Where AI actually touches the money in a Shopify store

This report was built the slow way. Every LinkedIn and X post in the corpus was read and scored by a model, one post at a time — 5,328 of them — and 1,000 kept as relevant and substantive. Every number in those 1,000 posts was extracted with its source, then 711 raw claims were consolidated into 329 distinct facts: the same figure repeated by forty accounts counts once. Each fact carries the organisation it originated with and a reliability grade, and the grading is the same for everyone, so a vendor's launch claim is rated low even when the vendor is a company we like.

What that process keeps producing is a gap between what happened and what was said about what happened. Shopify disclosed a 13x figure for one quarter from a small base and a 3x figure the next; the 13x is the one still circulating. Walmart's assistant premium is a comparison between two self-selected groups; it is quoted as a lift. Adobe published 42% in March and 60% in July; the number that spread fastest, 54%, was published by nobody.

So the chapters are arranged by where the money moves — discovery, agents, the Shopify stack, on-site experience, funnel, ads, retention, support, content — and each one states what is live, what has shipped but is unproven, and what is still a demo. Chapter 11 lists the numbers this report refuses to use, and gives you a way to produce your own instead.

Read the strength column first. Three stages have first-party or research-grade numbers behind them; three have nothing but vendor claims. That asymmetry is the most useful thing on this page — it tells you where to move budget and where to run your own test before believing anyone.

Two things fall out of the table. First, AI's grip is tightest at the two ends — discovery and retention — and weakest exactly in the middle, on the pages merchants spend most of their time editing. Second, the stages with the best evidence are the ones where a platform publishes numbers because it is selling you the channel; the stages with no evidence are the ones where the tooling is cheapest to buy. Those two facts are not unrelated.

Goldman Sachs tracked the share of traffic that generative AI sends to five of the largest ecommerce platforms in the world — Alibaba, Amazon, JD, Coupang and MercadoLibre. Across 2024 and 2025 the lines sit at or below 1%. From March 2026 they bend sharply upward. They are still at or below 1%. That is the honest opening position for this report: the growth rates are real, the base is small, and a large share of the numbers circulating this year are one of those two facts wearing the other's clothes.

Read the chart for its shape, not its level. The steepening from March 2026 is the thing worth planning around; the level says you have time to plan. One caveat travels with it: the post attributes the data to Goldman Sachs, but the image itself carries no source line, so treat it as named-but-secondhand.

What is actually measured

The most defensible market-level read comes from Similarweb and Statista, who put AI in 11.4% of US shopping journeys in 2026, up from 4.5% in 2024, measured on desktop. The same dataset carries the number that most retellings drop: 89% of shopping journeys that involve AI also involve search. Only 11% of them involve AI without search. AI is being added to the journey, not substituted into it, and that distinction decides whether your search budget is under threat or merely under-instrumented.

Similarweb's 23% conversion figure on the same card is a journey-level rate, not a session rate. It cannot be laid next to Adobe's or Shopify's session-level numbers, and Chapter 11 shows what happens when people do it anyway.

Stated preference points the same way. Semrush asked consumers to pick a single source they would keep before buying: 19.4% chose an AI chatbot, behind reviews at 37.1% and search engines at 36.7%. A fifth of shoppers naming an AI assistant as the one source they would keep is a real result. It is also a hypothetical question, and it is third place.

Generational data narrows it further. PYMNTS Intelligence found 41% of US millennials using ChatGPT for product discovery in July 2026 — behind Google at 57%, but already ahead of Amazon at 37%. The purchases themselves still happen largely in stores. Discovery has moved faster than checkout, which is why a chapter on discovery (02) has thirty-nine high-reliability facts and a chapter on checkout (06) has one.

The adoption number depends entirely on the question

Every headline share in the corpus is technically defensible and none of them are comparable. Here is the same phenomenon measured six ways:

The top and bottom of that table differ by more than three times, and the gap is almost entirely in the wording. Deloitte's 56% clears a very low bar — once, ever. Accenture's 57% is not about shopping at all. BCG's roughly one third is the most useful of the six because the base and the window are both stated, and because the same BCG report gives the number that matters more: 13% of consumers say they buy whatever the AI tool recommends. Presence in the journey is about a third of shoppers. Handing over the decision is 13%.

Three times the disagreement, one word of explanation

US agentic commerce by 2030: forecasts differ by more than 3x

The houses define the market differently: spend agents complete, versus spend agents merely orchestrate.

Named but secondhandBain & Company; McKinsey; one unattributed news projection (low reliability). Both houses publish ranges; these are the ends of each range.

Show data table

McKinsey — high	$1,000B
McKinsey — low	$900B
Bain — high	$500B
Unattributed projection	$385B
Bain — low	$300B

McKinsey's US figure and Bain's US figure describe the same country, the same year and the same technology, and differ by a factor of more than three. The explanation is one word: orchestrate. McKinsey's $900 billion to $1 trillion counts retail revenue an agent had a hand in directing. Bain's $300–500 billion counts a narrower agentic market. McKinsey's global equivalent, $3 trillion to $5 trillion of consumer commerce by 2030, is a scenario range, and the word "up to" falls off it in most of the posts that repeat it.

Share forecasts fragment the same way, and the base is always the trick. Gartner's 20% by 2030 is 20% of all transactions. McKinsey and Rye's 15–25% is of ecommerce transactions. EMARKETER's 15.8% is of US ecommerce sales, under a base case, up from 3.2% in 2026. Three shares that look like the same claim, three different denominators, all landing on the same year.

EMARKETER's curve is the most operationally useful forecast in the pack, for a reason that has nothing to do with its end point: it starts at 3.2% in 2026, so it tells you the size of the thing you are being asked to budget against today. The report's own analyst adds the detail that reframes the whole category: the majority of those AI-driven sales are expected to come through retailer-native assistants such as Alexa for Shopping and Walmart's Sparky, and fewer than half through platforms like ChatGPT and Gemini. If that holds, the surface merchants control matters more than the surface they are panicking about. Chapter 05 takes that surface apart.

Ambition is cheap; production is not

Consumer companies: agentic AI ambition against operational readiness

Share of surveyed companies. The gap between the first bar and the rest is the story.

First-party / researchDeloitte, State of AI in the Consumer Industry 2026.

Show data table

Plan to deploy agentic AI within 2 years	73%
Report at least moderate adoption today	24%
40%+ of AI experiments in production	23%
Mature governance for autonomous agents	20%

Deloitte's survey of consumer companies is the cleanest picture of the gap in this chapter's ledger. 73% plan to deploy agentic AI within two years. 24% report even moderate adoption today. 23% have moved at least 40% of their AI experiments into production. 20% have mature governance for autonomous agents. And 82% have not redesigned a single job around AI capability — which is the number that predicts the other four, because an agent that nobody's role changed to accommodate is an agent nobody is accountable for.

65% of shoppers plan to use AI for their holiday shopping this year. 8% of retailers say they feel ready for it. I read that gap twice.

Sarah Willersdorf, Kinship AI — LinkedIn, 2 Sep 2026

Narvar's survey behind that quote is small on the retailer side: 100 executives against 1,348 shoppers. Treat the 8% as directional. The direction is corroborated by where the money is pointed inside marketing teams.

Epsilon's split is the quiet finding of this chapter. 71% of marketers aim AI at productivity; 9% aim it at revenue growth. Intuit's QuickBooks survey of small businesses, January 2026, lands in the same place from the other side: 78% report productivity gains, 43% report revenue increases. Firms are buying AI to do the current job faster, then reporting the results as if they had bought growth.

Who is actually shipping

Four operators in the ledger are past the pilot stage, and none of them got there through a shopping assistant. Walmart's US ecommerce grew 24% in Q2 FY27 while US comparable sales grew 2.6%. The gap is real and widely quoted, and Walmart does not attribute it to AI, whatever the listicles do. Alibaba's merchant-facing ecommerce AI agents took more than 60,000 paid users in five months, which is the only clean paid-adoption number in the chapter. Shopee's in-house 245-billion-parameter commerce model, Compass, went from 3 billion to 340 billion API tokens a month in eight months. Tesco consolidated 250 separate AI initiatives into one group strategy backed by a 6,000-person technology team — the least glamorous item on the list and probably the most transferable.

Against that, the headline multiples are decelerating exactly as a small base predicts. Shopify reported AI-referred orders up about 13x year over year in Q1 2026 and about 3x in Q2 2026. Both are official; both are correct for their own quarter; neither is a slowdown in absolute volume. Chapters 02 and 04 take those numbers apart properly.

89% of merchants say they are actively preparing for agentic commerce. Agents are involved in about 3% of transactions today. That gap is being filled with a bet.

Nikki Baird, retail analyst — LinkedIn, 8 Sep 2026

Neither of Baird's two numbers can be traced to a named source in this corpus, so do not quote them as measurements. Quote the sentence after them. The preparation is real, the transaction share is small, and the space between is being filled with capital allocated on forecasts that disagree by a factor of three.

How much traffic is actually at stake

Every retelling of AI discovery opens with a growth rate, and the growth rates are genuinely large. They are also shrinking each quarter, because the base is no longer near zero. Adobe's successive readings of one metric — AI-referred visits to US retail sites, year on year — run up to 4,700% (July 2025), 693% (holiday 2025), 393% (Q1 2026), 125% (April–June 2026), 62% (July 2026). Quoting the 2025 number in 2026 is the most common mistake in this corpus.

The absolute shares are the sobering half. Bernstein's reading of Similarweb data puts generative-AI referrals at 0.8–1.4% of total traffic at major US retailers in August 2026 — while the same analysis shows them at 22–39% of referral traffic, a denominator that excludes direct, organic and paid. That gap is why Bessemer's widely-shared "15–20% of referral traffic from AI chat" is compatible with a store where AI barely registers in the totals.

First-party disclosures land in the same range. John Lewis says product searches arriving through AI agents went from 0.3% to 2.5% of its product searches in a year. Target's CEO told the Q2 2026 earnings call that traffic from external AI platforms grows 3.5x faster than the industry average, but is still small, with no absolute figure given. At the low end, Ethercycle's study of ten Shopify stores found AI referrals drove $75K of $76.5M in revenue — 0.10% — over January–June 2026, on last-click attribution that undercounts influence on a purchase AI did not close.

What those visitors are worth

The case for caring about 2% of searches is that they are not an average 2%. Shopify's Q1 2026 commerce data found sessions landing on a product page from AI platforms converted about 50% better than the same kind of session from organic search — average gap 56%, holding in 23 of 25 categories — with 14% higher average order value. About half of AI-referred sessions start on a product page at all, against roughly 20% for organic search. That last number is the mechanism behind the other two: the assistant does the browsing and hands over a shopper who has already chosen.

Adobe measures something different and lands in the same direction: AI-referred retail traffic converting 42% better than non-AI traffic in March 2026, 60% better in July 2026, revenue per visit 53% higher. Read the baselines, not the headline. Shopify compares AI with organic search on product-page-landing sessions only; Adobe compares AI with every other channel a retail site tracks, at site level. In both, the AI group is made of people who chose to shop through an assistant. That is a selection effect, not a measured lift — nobody in this corpus randomised anything.

Similarweb's finding is the most useful for planning precisely because it is not a channel comparison. Journeys combining AI with search convert at 23% against 12.5% for AI alone — and 89% of AI-using journeys include search anyway. AI appeared in 11.4% of shopping journeys in 2026, up from 4.5% in 2024, and it usually appears mid-journey: 76% of those sessions, against only 23% at the start.

That is the practical shape of the channel. AI is rarely the first touch and rarely the last. It is the research step that used to be a comparison blog, a Reddit thread and four open tabs, compressed into one answer — which is why it arrives with intent attached, and why last-click attribution will systematically understate it.

Which assistant sends them

ChatGPT is still the channel for practical purposes, and it is slipping at the edges. In the ReFiBuy / Digital Commerce 360 AI1000, top-1,000 US retailers whose largest AI referral source is ChatGPT fell from 844 to 722 between Q1 and Q2 2026, while Gemini went 16 → 32, Perplexity 8 → 21, Claude 1 → 15. These are counts of retailers, not shares of traffic.

The same index makes a more durable point about who wins. Only 10 of the 100 largest online retailers made the AI1000's top 100, and the new number one, Nixon, ranks 722nd by online sales. 71% of Shopify's AI-attributed orders in 2025 came from long-tail, specialised products. Size on the old shelf does not transfer.

10 July 2026: retrieval moved to the feed

Then the mechanism changed under everyone. Profound analysed roughly 1.75 million ChatGPT shopping prompts and found the share of product recommendations retrieved from integrated merchant feeds went from 8.26% to 61.54% in a single day, with the release of ChatGPT 5.6. Concentration followed: the top ten merchants' share of recommendations rose from 22.5% to 41.8%, distinct merchants shown fell from 13,524 to 10,607, and in the days around 10 July, 450 Profound-tracked brands lost at least a third of their ChatGPT Shopping visibility while 67 gained a third.

Profound tracks its own customer base, and a second reading of the same event matters. Novi ran 1,500 queries across 13 product categories and concluded that feeds decide rendering, not selection: which products the model picks still comes from how it understands them across the web, while the feed supplies the verified price, variant IDs, imagery and stock that let a pick be drawn as a shopping card.

Both readings point the same way for a merchant: if your product is not in a feed ChatGPT ingests, your best case is a text mention beside competitors who get a card with a price on it. Shopify and Etsy catalogs are already integrated with no application needed, and Profound puts Shopify at about 35% of all feed-integrated retrievals — the specific reason a small Shopify merchant is not automatically locked out of a shelf that just got more concentrated.

The shelf got shorter, and more expensive

Google's side changed shape rather than plumbing. Productrise tracked 2 million listings over 23 days in August 2026: AI Mode shows 3.9 products per answer against 27.8 in the classic shopping carousel, only 1.28% of listings overlapped daily between the two surfaces, and on 49.6% of shared products the top seller differed. The lead offer in AI Mode averaged 21.6% more expensive than classic search for the same product. This is one small firm's study, of the lead listing only, and deserves that caveat every time it is quoted — but the direction matches Ahrefs finding only 38% of AI Overview citations now come from the organic top ten.

That is the honest read of a four-product shelf selected on undisclosed criteria. A brand that only ever won on price has lost its mechanism; a brand with reviews, complete attributes and third-party coverage has a shot it never had in a 28-product carousel sorted by price.

What a merchant can actually control

Very little of the ranking, and most of the inputs. Google's Merchant Center AI performance report, in beta and expanding through 2026, names the AI shopping intents a merchant appears for and the product attributes shoppers ask about that the feed does not contain.

Feed work is the only lever here with a before-and-after measurement attached, and the measurement is mixed: in a four-feed audit, adding missing product images lifted organic clicks by 18 to 23 percentage points, fixing vague colour values did nothing measurable, and category fixes were too few to judge. Treat "complete the feed" as a portfolio of bets, not one bet.

It also helps to see what an assistant actually receives. Pulled from ChatGPT's retrieval stream, a product page arrives as raw meta tags plus a flattened markdown render — no layout, no alt text, no links. Design decisions do not survive the trip. Title text, attribute values and structured data do.

The conclusion is smaller and more concrete than the discourse suggests. AI discovery is not replacing search: search still drives about a third of Shopify storefront sessions and grew 1.3x over two years, and almost every AI journey passes through it anyway. What AI has taken is the comparison step — the place where a shopper decides which product to want.

Three trades get filed under "agentic commerce". An AI recommends and a human buys. An AI assembles the cart and a human taps pay. An AI holds a credential and buys alone. The first is a real channel, the second is shipping, the third is a rounding error — and the distance between them is this chapter.

Agent-mediated influence is scaling faster than agent-mediated checkout.

Molly Schonthal, agentic commerce consultant — LinkedIn, 18 Aug 2026

What actually transacts today

The useful question is not which protocol wins but which surface can take an order this quarter. Sorted that way the field thins fast.

The checkout that died

OpenAI and Stripe launched Instant Checkout in September 2025 and ended it on 4 March 2026. Shopify's count, secondhand, was roughly twelve merchants actually using it; an unattributed post says about thirty integrated with sales near zero. Both can hold — "integrated" and "in use" are different tests. It reportedly took about 4% from merchants. One week before the shutdown, OpenAI and Amazon announced a $50B partnership; reading cause into that sequence is inference, not reporting.

That screen is what a working agentic purchase looked like — a Glossier fragrance bought in the chat, Glossier as seller of record. It worked, and it still lost to a redirect. The technical problem was solved a year ago; the conversion problem was not. ChatGPT now discovers and hands the shopper to you, which one fashion operator called affiliate marketing with a chat interface on top.

The protocol map, minus the war

UCP, co-built by Google and Shopify, covers discovery, cart, checkout, order status and identity linking across four transports; its Tech Council added Amazon, Meta, Microsoft, Salesforce and Stripe in April 2026. ACP, from OpenAI and Stripe, is now mostly a discovery format with a handoff. AP2 handles rules-based agent payments. Muse speaks several at once.

The diagram is worth ten minutes because it shows where your obligations sit: capabilities you must expose, transports you may choose. It says nothing about adoption, and adoption is the weak leg.

UCP has 51 live merchants. Growth is stagnating. The infrastructure is real. The adoption isn't.

Philip Qiu, co-founder, Henry Labs — LinkedIn, 5 May 2026

That count is unattributed and four months stale — a snapshot, not a running total. The harder evidence against the war framing is what the machine-to-machine rails move.

Bernstein's read of Artemis data puts real agent-to-agent volume on x402 at a $5.1M peak in November 2025 and under $1M a month through 2026, MPP near zero. The same agency posts that count ten competing agentic protocols are counting them over a flow smaller than one mid-size store's year.

Fees, and who holds the bag

Where a company names its fee, it is modest. Square charges its standard ~2.9% + $0.30 on restaurant orders placed inside ChatGPT and Claude, no added marketplace commission, against the 15–30% delivery marketplaces take. Instacart inside Gemini shows the shopper a 4% service fee at handoff.

Watch what that Gemini screen does: it builds the basket in Google, then bounces to "Checkout on Instacart". The disclosed fee is the price of the handoff, not of the checkout. Agent assembles, owner of the payment relationship closes — that is the pattern earning money in 2026.

Liability did not move with the interface. ACP states the merchant remains merchant of record and keeps refunds, chargebacks and compliance; UCP takes the same position. What changed is that the buyer is now a probabilistic system, and most order records were never built to record what the human authorised the agent to do.

When a customer disputes what their agent bought, who produces the evidence?

Richard Emanuel, Claude and n8n partner — LinkedIn, 9 Sep 2026

The trust ceiling is real; the surveys disagree about where

This corpus holds at least a dozen 2026 numbers for "would you let an agent buy", from 7% to 85%. They are not measuring the same act. Seven of them, and what each actually asked:

Consumers use AI to research, not yet to pay

Two surveys, kept in their own pairs. The Visa pair asks the same people about the same act with and without a named payment brand.

First-party / researchIpsos "Shopping with AI" (AI-aware consumers); Visa Trust Index for agentic commerce (2,065 US consumers, 2026).

Show data table

Visa · accept when Visa is named as handler	61%
Ipsos · use AI for product research	27%
Visa · trust GenAI to pay (unbranded)	23%
Ipsos · let AI buy autonomously	9%

Read the pairs, never the spread. The Visa pair is the most instructive line in the chapter: same people, same act, 23% to 61% on the strength of a named payment brand. Bain's 3x higher trust in a retailer's own agent than a third-party bot points the same way. Trust attaches to whoever takes the money.

Meta's App Store listing for Muse shows the compromise everyone landed on: the agent proposes, the human taps Allow, then Link pays. Stripe says Muse checks out instantly at more than a million Link businesses — every one still behind a human tap. That is the honest state of the art, and why 3% of transactions and 89% preparing are both true.

Connect now, wait on the rest

Microsoft's Merchant Center states the real entry fee plainly: publish a return policy and support contacts and you are eligible for AI shopping surfaces. The work is hygiene, not integration.

1. Now: the free default rails

   UCP through Shopify Catalog costs nothing and is on by default; Square sellers are opted into ChatGPT and Claude ordering automatically. The work is catalog completeness, live inventory and price, a return policy, a support contact.

2. Now: the dispute record

   Log agent identity, approval event and stated constraint on every agentic order. One practitioner's read of the timing: Visa's VAMP excessive-chargeback threshold dropped from 2.2% to 1.5% on 1 April 2026 for US, Canada, EU and APAC merchants. Your room for a new dispute class shrank this year.

3. Wait: integrations sold on fee arbitrage

   Anything priced against the unattributed 7.2%-versus-zero framing or promising 40% more volume for running both. If your platform already speaks UCP, a second direct integration buys little at 3% of transactions.

4. Wait: agent-to-agent payment rails

   Under $1M a month across all of x402 in 2026. Revisit when a named merchant publishes a volume number.

Keep the vendor numbers in the vendor column

Anthropic's materials say retailers running Claude shopping agents saw carts up to 35% larger and shoppers 60% more likely to complete. Twenty posts here repeat them; most drop the "up to", several swap "completion" for "conversion". The underlying disclosure is one early partner self-reporting 30–35%, no sample size, no stated comparison group. A separate +40% conversion case in the same materials is credited to a startup called Vambe.

They claim 40% increase in conversions, and then when you find the source, it's some no-name startup called "Vambe" whose website currently reads "404: NOT_FOUND Code: DEPLOYMENT_NOT_FOUND."

Kelly Goetsch, President, Pipe17 — LinkedIn, 3 Sep 2026

That is not an argument against the blueprint, which is useful engineering, free and forkable. The discipline cuts both ways: unattributed posts claim Amazon's Buy for Me covers 100M products across 400,000 merchants, with 170+ merchants organising legal action and chargebacks up 28%. None of it is sourced. Both sets of numbers are selling something.

Nobody installed this

In March 2026 Shopify began turning Agentic Storefronts on for eligible stores. Aleyda Solis noted at the time that the channel was active by default, with permissions sitting in Settings → Sales channels — an opt-out screen, not an install flow (X, 25 Mar 2026). Six months on, the same screen lists ChatGPT, Microsoft Copilot, Google and Shop, and since 8 September, Meta.

The plumbing under that screen is what makes the default stick. Shopify Catalog syndicates structured product data to AI channels, and Shopify's help documentation states plainly that blocking crawlers in robots.txt does not stop it — different pipes. Every store also serves /agents.md, /llms.txt and /llms-full.txt without anyone creating them. Read the documentation below as the list of what a merchant cannot switch off with the tools they already know.

That default has a measurable consequence off-platform. Profound, studying ChatGPT Shopping retrievals from July to September 2026, put Shopify integrations at roughly 35% of feed-integrated product retrievals. Shopify's own matching claim — AI searches drawing on structured Catalog data convert at 2x the rate of AI searches relying on scraped or outdated web data, Q1–Q2 2026 — has never had its comparison set published (post).

Every new channel repeats the pattern. September's Meta launch is the clean example: products go to Meta by default via Catalog, so the merchant's decision is an opt-out they have to notice first.

The 13x belongs to Q1 2026

Shopify said AI-driven traffic to merchant stores was up more than 8x and AI-driven orders up nearly 13x year over year in Q1 2026. Three months later, on the Q2 2026 earnings call of 5 August, the same two lines came in at about 3x each. Together they look like a collapse. They are not.

A year-over-year multiple divides this quarter by the same quarter a year earlier. Q1 2025 was close to nothing, so almost any volume in Q1 2026 produced a large number. By Q2 2025 the base had grown, so Q2 2026 was divided by a bigger denominator. A multiple falling from 13x to 3x is fully consistent with absolute volume still rising — it is what the first year of a new channel looks like once the denominator catches up.

The practical version: a September slide reading "AI orders are up 13x", with no "Q1 2026" attached, is quoting a stale quarter as current. Three further versions circulate — 15x for full-year 2025 (secondhand, no link), 7x traffic and 11x orders from an older disclosure with no period attached, and an unattributed 385% for Sidekick sitting awkwardly beside Shopify's disclosed 4x. Only the figures with a quarter attached survive checking.

One real dashboard

Shopify's disclosures are multiples and rates. None says what the channel is worth on one store, and the Agentic Storefronts graphic that circulated after the June launch is no help: its store name and its $103,813 read as sample data in a launch image. The only real dashboard in this corpus is a client store Kurt Elster posted in May 2026 — ChatGPT $5,673.68 over 30 days, Copilot $301.79, next to $52,483.10 through the Shop channel. Read the channel lines, not the tile's headline total, which does not reconcile with them; the "0.27% of revenue" in Elster's caption cannot be checked from the screenshot.

At this size the channel differences matter more than the totals. Google, Copilot and Meta close checkout inside the AI chat, where browser pixels, bundles and subscriptions do not apply; ChatGPT redirects the shopper to the merchant's own Shopify checkout (Stack Architect, X, 24 Aug 2026). A store whose margin rests on post-purchase upsells is not "live on AI" in the same sense across all five. Nor was the rollout clean: in March, Ben Kennedy found every store he tested returning a "Product not available" error after launch was announced.

Sidekick's most common job is meta tags

Sidekick usage was up 4x year over year in Q1 2026 per Shopify's COO and the Spring '26 Edition, and daily active merchants grew 3.6x in Q2. At DotDev 2026 Shopify showed the request mix over 30 days, and its shape says more than the growth rate does.

SEO and meta tags lead at 1.1M requests, ahead of fulfillment (716k) and inventory (646k). Merchants are not asking an assistant for strategy; they are handing it the copy work they were already doing badly. The structural change is app intents: Sidekick now reaches into AfterShip, Consentmo, Klaviyo, Judge.me and Loop, asking permission before each read, and since an August changelog it opens the app's own page full-screen when it invokes one — moving app discovery out of the App Store and into the assistant (post). The supply side answered: App Store review went from 40 days to 4 (Atlee Clark, VP Partnerships, Shopify), and May 2026 alone brought over 2,000 new apps, 27% mentioning AI (App Store Pulse). What none of it captures is the failure mode operators complain about.

The connector is really a permissions screen

In May 2026 Shopify shipped connector apps for ChatGPT and Claude so merchants could run the store from a chat window they already had open. Harley Finkelstein's framing was a merchant survey finding 83% already use ChatGPT (LinkedIn, 4 May 2026) — method undisclosed, so read it as direction, not measurement. What deserves study is not the demo but what the connector exposes.

Nineteen interactive tools, four read-only, with GraphQL mutations and shop switching held behind a "needs approval" gate. That list is the real governance surface of the whole stack: it decides what an agent may change on a live store without a human present. Evaluate that screen and the app-intent approvals; the marketing page can wait.

The one AI surface a merchant fully controls

The previous three chapters are about places somebody else sets the ranking. Your own domain is not one of them. You choose what search returns, you see the whole session, the customer record stays with you, and nothing about the experience depends on a protocol being ratified. EMARKETER's forecast puts the money there too: retailer-native assistants — Rufus, Sparky and their kind — are expected to drive 54.1% of US AI-driven retail ecommerce sales in 2026, more than every general chatbot combined. Shoppers report the same preference: Evercore's agentic commerce survey found 55% use a retailer's built-in assistant against 48% who use a general AI tool, and 66% were satisfied against 3% dissatisfied.

In practice the surface is less exotic than the launch videos. Constructor's Ask Cleo sits on the Rugs Direct search results page, answers "which of these is actually washable", and hands back product cards with an offer to narrow by shape — a retrieval layer with a mouth on it, anchored to a catalog the merchant already maintains.

The numbers everyone is quoting this quarter

Four disclosures set the tone of the whole conversation. Walmart's CEO John Furner told the Q2 FY27 earnings call in August 2026 that Sparky users spend roughly 40% more per order than customers who do not use it, and that Sparky users were up 70% year over year. Amazon's Andy Jassy said in the Q2 2026 call that US customers using its shopping assistant spend over 40% more than non-users, and that 350 million shoppers used it in the past year. Lowe's says shoppers who engage Mylow convert at about three times the rate of those who do not, across 25 million questions answered. Williams-Sonoma says its Olive assistant converts at 3x the normal rate, with revenue through it up 620% and engagement up 700% since January 2026.

Below the enterprise tier the same claim repeats with bigger multiples and thinner sourcing. Constructor reports that Rugs Direct and Lightopia shoppers who engage Ask Cleo add to cart at 8–9x the rate of other shoppers. Nosto claims a +92% conversion increase in A/B tests of an LLM assistant wired to its intent model; Loops AI claims +83% conversion and +77% add-to-cart.

Retailer AI assistants: how much more users spend than non-users

These compare shoppers who chose the assistant with shoppers who did not. Self-selection, not a measured lift.

Named but secondhandWalmart and Amazon Q2 2026 earnings calls; Albertsons via newsletter (low reliability).

Show data table

Walmart Sparky (per order, Q2 FY27)	40%
Amazon AI assistant (US, Q2 2026)	40%
Albertsons (AOV, secondhand)	26%

Read that chart as a chart of one methodology, not of three results. Every bar is the same subtraction: the average of people who used the assistant minus the average of people who did not. The Albertsons bar is the weakest — secondhand, its metric undefined, and described in at least one account as experience-versus-experience rather than user-versus-non-user.

What the gap actually measures

A shopper who opens an assistant to plan a week of dinners was already the bigger basket: further down the funnel, more likely logged in, doing a stock-up trip rather than grabbing one item. The assistant did not create her; it attracted her. The 40% is the selection effect plus whatever the tool contributed, and nothing in the disclosures separates the two.

Both can be true. Neither is a lift. Those groups are not the same shoppers. Whoever opens an AI assistant to plan a week of dinners was already the bigger basket. The number describes who uses the tool, not what the tool caused.

Manish Sharma, omnichannel commerce executive — LinkedIn, 9 Sep 2026

Walmart's own number has a second lesson in it. The 40% is an increase on the 35% premium the company gave in Q1 FY27, three months earlier. Retellings that quote a single figure lose the trend, and several restate the Q1 number as "baskets 35% larger", which is a different metric from spend per order. When a claim survives three retellings, check which quarter and which denominator it started in.

Walmart Sparky: the spend premium its users carry, quarter to quarter

Two disclosures three months apart. Retellings that quote only one number drop the trend.

First-party / researchWalmart investor remarks and Q2 FY27 earnings call. Users self-select.

Show data table

Q1 FY27	35%
Q2 FY27	40%

The clearest illustration of the trap is a vendor slide. Loops AI reports A101 Ekstra visitors who use its assistant converting at 37% against 21% for standard visits. Printing both numbers is more honesty than most vendors offer, and it lets you check the arithmetic: 21% to 37% is a 76% increase, not the +83% quoted elsewhere in the same campaign, while a companion card calls assistant sessions 2x. Three numbers, one dataset.

Survey evidence gets closer to causation without reaching it. Evercore found 57% of Amazon assistant users bought a product they had not been considering — a genuine discovery effect, self-reported. But in the same survey only 36.4% said the assistant raised their Amazon spending, while 44.5% said it was flat. Discovery moved. Total spend mostly did not. That is a shift in which product gets bought, not a shift in how much gets spent, and for a brand those are opposite problems.

What shoppers say decides where they shop

One finding cuts against the premise of the category. Digital Commerce 360's 2026 holiday report asked shoppers what would influence where they shop this season. Recommendations and personalization came last, at 9.9%. Price and discounts led at 81.8%, fast or free shipping at 54.6%; even "website experience" edged personalization out, at 10.3%. Stated preference is not behaviour, and nobody notices that a collection page was reordered — but it caps what this work can be sold as internally. A store with a thin margin and a slow courier does not have a personalization problem. Put the letters AI in the question and the answers warm up — a vendor survey with eTail Insights has 65% of shoppers planning to use AI for part of their holiday shopping, 78% if it were more personalized — but only the first question is attached to a purchase (post, post).

Maybe good personalisation isn't something customers are supposed to notice at all. It just makes those answers easier to find.

Liliia Maliutina, ecommerce consultant — LinkedIn, 2 Sep 2026

Peer-reviewed work goes further. A study of 299 fashion ecommerce customers in the Journal of Retailing and Consumer Services found no direct link between recommendation quality and post-purchase satisfaction; perceived value carried the relationship, and dissatisfaction — not recommendation quality — was the key predictor of returns. In a category where returns eat the margin, that moves the case downstream: the number to watch is the return rate, not the click.

The largest controlled effect anywhere in this chapter is not a recommendation at all. A randomized field experiment with 35,625 users, reported by Wang, Zhang, Gao and Tan in September 2026, varied one thing in an AI-generated summary of customer reviews: how strongly it framed them as customer consensus. Product clicks rose 24%; time spent reading fell 10%. It is an SSRN working paper, not yet peer reviewed, so hold the magnitude loosely — but it is the design nobody else here ran, and the lever it moved was wording, not ranking. Any store already showing an AI review summary is running that experiment on its customers without a control arm (post).

Personalization that is not a chatbot

The results at merchant scale come from ranking and recommendation work, not conversation, and they are smaller and more believable. DFS added AI recommendations and photo-led visual search and reports +10% online conversion, +8% AOV and a 3% drop in bounce. THG's foundation finder for LOOKFANTASTIC reports a 5% revenue uplift, with 96% of buyers through it purchasing a product they had never bought before. Anphonic reports personalized upsells lifting Vocca's AOV 14.49%. Dr. Squatch put recommendations on its order-tracking page and reports a 31.89% click rate and $32,978 in Q1 2026 revenue from them — a small number, and one of the few in this chapter with a currency symbol and a period attached.

Every figure in that paragraph is vendor- or agency-published, with no controls disclosed. Treat them as a plausible range for what good merchandising work returns — single digits to low double digits — rather than as benchmarks.

The least-discussed version of this is silent. Cooee's category merchandising for Shopify scores each SKU on more than a hundred signals and reorders the collection page: featured and new items up, out-of-stock down. No chat, no persona, nothing for the shopper to opt into — which also means no user-versus-non-user number to quote, because everyone gets it. That is a feature, not a shortcoming, and it is why this kind of change is the easier one to test properly.

One operator has published the test. Matthew Bertulli of Pela says he spent a couple of all-nighters with Claude Code building a small version of TikTok's For You page over a collection page: it watched how a visitor scrolled and reordered what came next. One split test, the static page left as the control, profit per session up 28%. His own caveat: "One test isn't a benchmark. It might do nothing on your store." Unaudited — and still one of the few numbers here with a control group behind it, and the only one whose denominator was chosen before the test ran (post).

The biggest production version of the quiet kind also has no chat window. RecGPT — Alibaba Group and Renmin University of China, in a paper accepted by ACM Transactions on Information Systems — runs on Taobao's "Guess What You Like" homepage feed. It throws away plain clicks as noise and keeps purchases, favourites, add-to-carts, detailed views and searches; an LLM compresses that history into an interest profile matched to a taxonomy of 169 interests, refreshed every two weeks. Every LLM stage runs offline in batch; the online path is a retrieval model answering in roughly 25 ms. No revenue lift is published with it, and the architecture is the transferable part: the model is not in the request path, and the input is deliberate behaviour, not clicks (post).

The floor most stores have not poured

Algolia and The Retail Hive's barometer found 52% of retailers still rely on out-of-the-box third-party search tools, and just 8% feel ready for AI agents to navigate their sites on a shopper's behalf. The models are not the constraint; the catalog, the attributes and the query understanding underneath them are.

Nearly half say they'd shop more with grocers that use AI well. But AI won't save you if "birthday cake" returns birthday candles. It'll just be wrong faster.

John A. Stewart, Algolia — X, 8 Sep 2026

The same research found 44% of grocery shoppers have no go-to store, which is the commercial reason to care: the shopper is not loyal, and a bad result page is a live churn event. On the platform side the tooling is arriving fast — Shopify's in-house 0.8-billion-parameter model scaled buyer-profile generation from 2 million to 72 million profiles a day — but for most merchants this arrives through the Shopify App Store, where a search for "personalization" returns 4,008 apps and five of the ten fastest-climbing new apps in one week of September 2026 were upsell tools. Supply is not the constraint either.

How to get a number you can act on

1. Fix retrieval before you add a mouth

   An assistant reads the same index your search box does. If "birthday cake" returns candles today, the assistant will say the wrong thing more fluently. Attributes, synonyms and stock accuracy come first.

2. Hold out, do not compare

   Withhold the feature from a random slice of traffic and compare slices, not users. Every spectacular number in this chapter is a user-versus-non-user comparison; every modest one came closer to a controlled test.

3. Pick the denominator before you run it

   Conversion per session, per user, or per visitor who reached a product page are three different numbers, and Walmart's 35% drifted into "35% larger baskets" in the retelling precisely because nobody carried the denominator.

4. Measure substitution, not just addition

   Evercore's 57% who bought something unconsidered against 44.5% whose spend stayed flat is the whole risk in one line: the assistant may be moving which of your SKUs sells rather than how much sells. Carry the return rate beside it: the 299-customer study found dissatisfaction, not recommendation quality, is what predicts a parcel coming back.

What ten facts look like

Every other chapter in this report argues about which measured number to trust. This one cannot. The whole ledger still fits in a table, and none of the three new rows is first-party.

Note the one high-reliability row. Walmart cut prices 11,000-plus times in the quarter against 7,200 the quarter before, disclosed on its own earnings call: a real number about promotional pressure that says nothing about AI. Two medium rows are surveys asking people what they would do; the three new ones are single-source. The four low rows are a pricing vendor, two agency cases, unattributed benchmarks, and a startup's first case study.

Both survey numbers are stated intent, collected by companies that sell into the problem they describe. Neither is a conversion rate. They point at one thing: the site is where the sale closes, even when the recommendation happened elsewhere.

Production got cheap. Evidence did not.

The published work is almost entirely about making pages faster. ScaleBot shipped a landing-page builder that writes an on-brand lander from scratch or, in its author's words, "rip[s] a competitor page in 1-click", carrying over fonts, logo and colours.

Read that screenshot for what it proves. The build step is now minutes. It says nothing about whether the page sold anything: no session count, no conversion rate, no control. Every AI page-building post in this corpus stops there.

The audit side is the same story, better instrumented. An n8n workflow fetches a landing page, converts it to Markdown and has Gemini 2.5 Flash return prioritised CRO fixes.

The run shown took 29 seconds and 2,381 tokens. That is the asymmetry of 2026: the cost of a CRO opinion is known to four significant figures, its accuracy nowhere. A bundle app handles the same trade more carefully — each natural-language prompt lands as a logged set of structure and settings changes, so a merchant can see what the model altered.

The logging is the part worth copying. If an AI edits your page, the change set has to stay readable afterwards, or you cannot attribute a result to it even when you run a test.

One store ran the test

The exception is not a page builder. Matthew Bertulli, co-founder of Pela, spent two nights with Claude Code building what he calls a dumb little version of TikTok's For You page for the brand's bestsellers collection: it watched what a visitor scrolled past, stopped on and opened, then reordered what came next. Split-tested against the static page, profit per session went up 28%. The caveats are his own — one test, one store, no session count or window published, and the man reporting the number owns the company. What earns it the space is not the 28% but the design: the only claim here with a control group and a metric fixed before the test, on a surface nobody was optimising. The order of products on a collection page has not changed on most stores since the theme went in.

The other control group belongs to a study nobody named: a randomised field experiment on 35,625 users varied how strongly an AI summary of customer reviews presented them as showing consensus, and product clicks rose 24% (Agnes Kaczmarek, X, 14 Sep 2026). Best design in the chapter, weakest attribution. Clicks are not orders, and the dial that moved was how confidently a machine summarises other people's opinions.

A/B testing with a model in the loop

Shopify's Winter '26 Editions put two relevant tools in the admin: Rollouts, native A/B testing with scheduled theme changes, and SimGym, which sends simulated AI shoppers through a storefront before real customers see it. Both are shipped, not proven. No merchant in this corpus has published a result from either, and a simulation is a prior, not an outcome.

Be careful with simulated shoppers. One post summarises research finding that LLMs predict purchase intent at around 90% accuracy when asked to roleplay a customer, and "fail completely" when asked to rate a product one to five, because models default to safe distributions (CrazyShyyt, X, 8 Sep 2026). The method decides the answer, and buying the tool means inheriting whichever method it chose. Amazon's 70–90% is the strongest version of the claim and still only triage: it predicts which arm wins, not by how much, and one call in five is wrong. Point it at the bottleneck below — which of fifty variants gets the month's traffic — not at the test itself.

Checkout: the hard numbers still point down

The closest thing to a measured AI conversion result belongs to a channel rather than a page, and it is negative: after testing roughly 200,000 items, Walmart found purchases completed inside ChatGPT converted about three times worse than click-outs to its own site (Gagan Ghotra, X, 20 Mar 2026). Chapter 03 owns that story. The point here is that it took a retailer of Walmart's size to produce it.

September added a second, from the other end of the pipe. HUMAN Security's traffic data puts 87% of AI-agent visits to retail sites on product pages and barely 2% at a checkout (Sreekant Vijayakumar, LinkedIn, 8 Sep 2026). Beside the Shopify figure Chapter 02 owns — about half of AI-referred sessions land straight on a product page, against roughly 20% for organic — the shape holds: AI traffic arrives deep in the funnel and does not close. Two populations with opposite intent land on the same PDP, read through one conversion rate. Split PDP conversion by referral source before buying anything to fix it (Bharti Prasad, LinkedIn, 25 Aug 2026).

Against that, one vendor case study: Ava's AI phone agents called shoppers who had abandoned checkout at pet&home.

Measure it yourself

One split test is not a benchmark, so the only honest number is still your own. Here is the cheapest version that produces evidence, inside a fortnight.

Two separate things happened to ads in 2026 and operators keep discussing them as one. Paid placements began appearing inside AI answers — the surface merchants were just told to earn their way into. Separately, making an ad got cheap enough that volume stopped being a budget question. The first has revenue behind it and almost no merchant-side evidence; the second has evidence, most of it written by people selling the tool.

The inventory arrived before the evidence

In late August 2026 OpenAI said ChatGPT Ads had passed a $1 billion annualised revenue run rate in under 200 days. Read that as written: a run rate annualises a recent period, it is not money recognised for a year. eMarketer's forecast has all chatbots together — ChatGPT and Google AI Mode included — generating under $1B in 2026 ad revenue, against OpenAI's own projection of $2.5B. Both can be true; only one describes invoices.

The revenue line matters less to a merchant than the shape of the placement. In Google AI Mode the sponsored listing sits inside the answer, in the model's own visual language, beside the citation panel a brand spent the quarter trying to enter. There is a Sponsored label and no ad block to scroll past: the unit competes for the same attention as an organic recommendation because it is rendered as one. Not every surface renders it that way. Through August 2026 ChatGPT had one placement, directly below the generated response, labelled sponsored and laid out like a search result — adjacent, not woven in.

Ads are filling AI answers: sponsored placements in ChatGPT hotel results

Share of answers carrying a sponsored placement. Travel queries; retail ad load was not measured.

Named but secondhandComscore, spring 2026. Context: OpenAI reports a $1B annualised ads run rate in under 200 days.

Show data table

	Answers with a sponsored placement
Mar 2026	6%
Apr 2026	14%
May 2026	24%

Comscore's hotel series is the only public measurement of ad load inside AI answers. It is travel, not retail, and three points is a short series — take the trajectory as the finding, not the level. A surface that launched ad-free was carrying a sponsored placement in roughly one answer in four within a quarter.

Buying it is now a Tuesday afternoon

Access moved faster than proof. In February 2026 ChatGPT Ads meant a $60 CPM and a $200,000 minimum; by May it was self-serve CPC bidding with no minimum. On 31 August 2026 the Ads Manager opened for direct buying in 31 European markets, and in India on 4 September. Amazon Ads now has a pilot letting selected US advertisers buy ChatGPT inventory through Amazon DSP — a retail-media team gets in without ever talking to OpenAI.

The console is familiar on purpose — campaign, ad group, ad; conversions, feeds, audiences. A team running Google Shopping can operate this second feed-driven auction in an afternoon, which is the trap: a mature console implies a mature channel.

It's unproven at scale. OpenAI's only public proof point is one anonymous e-commerce advertiser at 3x ROAS over 28 days. Vendor claim, no methodology.

Ingvar Kraatz, Co-Founder / COO at Bidnamic — LinkedIn, 4 Sep 2026

Bidnamic's second warning costs more. By OpenAI's own documentation the conversion pixel initialises consent to true unless told otherwise, and hashes customer data detected in site forms. Installed without a consent platform behind it, that is GDPR/PECR exposure on a test budget.

What it returns, honestly

Start with what buyers are told to expect. The most specific public set was published in August 2026 by paid specialist Nina Welt: CPM around $25–45 in scaled US segments, down from the early $60; a CPC benchmark near $1.72 against OpenAI's advice to open bids at $3–5; CTR around 0.7% blended, some campaigns at 1–2.5%; conversion rates from 0.2% to 5.8%. Collected expectations — Vendor or unsourced — not a measurement: no account count, no period, no method. The structural read is the durable part — at that CTR the unit performs like display while the user context reads like search. Price it as search and you switch it off inside a month.

The merchant-side evidence is two small tests pointing opposite ways. Windmill Strategy ran ChatGPT Ads for 30 days against industrial and manufacturing buyers: CPCs lower than expected, engagement solid, zero conversions. Planit, across three anonymised brands, reported organic ChatGPT traffic up 42% while ads were live (+10%, +41%, +164% by brand) — before-and-after, no control group, measured while AI referral traffic was growing everywhere. Both still beat the vendor 3x, because both name their method.

Amazon's are the strongest numbers here and still need care. They came on the Q2 2026 call, past the $19.8B advertising headline. The 48% describes the shopper as much as the placement: whoever asks an assistant a question and clicks the answer is already deep in a decision. It proves intent and volume are real on that surface, not that the ad created the purchase. It is also the placement a seller controls least. Sponsored Products Prompts left beta on 25 March 2026 as a billable CPC unit, only automatic and broad-match campaigns qualify, and a seller can opt out of specific prompts but cannot write one, bid on one or give it a budget; exposure is read after the fact in the Prompts report. No holdout, no line item to pause — so the merchant cannot turn Amazon's 48% into a number of their own.

The ad load lands on the visibility you were just told to chase

On mobile the sponsored merchant offers sit directly beneath the organic recommendation — same answer, same scroll, paid row under the earned one. Every point of ad load is answer real estate no feed, schema or citation strategy wins back. After AI Overviews, Smarter Ecommerce measured median Google Shopping impressions falling from about 1.85M to about 1.4M while median CTR rose from about 1.20% to about 1.55%: fewer impressions, better qualified. Vendor analysis over ~175B impressions, confounded by everything else Google shipped that year.

Amazon is charging my company money to answer customer questions and then directing our customers to inferior competitor products who have paid Amazon. We never signed up for Rufus ads.

Molson Hart, Founder and CEO of Viahart — X, 21 Apr 2026

Creative: the cost collapsed, the hit rate did not

Adoption is measured here, not claimed. IAB put 62% of ad buyers using generative AI for video creative in July 2026, up from 51%. XR Extreme Reach found 26% of marketers using digital replicas and 22% synthetic AI talent. ADWEEK reported research putting 88% of advertisers on AI in creative production; that study is unnamed.

The economics of testing did not move. Motion's analysis of 578,750 creatives and $1.29B of spend found roughly 5% of creatives are winners absorbing 55% of spend, and drew the useful conclusion: low hit rates are a statistical feature, not a quality signal. Cheaper production multiplies the denominator without improving the ratio. It moves the constraint from making an ad to judging one.

Volume does work somewhere. A Brazilian TikTok Shop account runs what the poster reads as one AI-generated presenter across dozens of products, top videos at 1.3M–4.1M views; the AI claim is his, unverified. That is media arbitrage with no brand to damage, and it now carries disclosure duties: New York's GBL §396-b took effect in June 2026, and Amazon began requiring sellers to tag AI-generated people in product media the month after.

Two findings cut across the volume play from opposite sides. IAB research reported in September 2026 found 56% of recent AI-shopping users preferred recommendations carrying creator perspectives, and 65% felt more confident when credible creator reviews informed the answer — secondhand, and stated preference rather than observed behaviour. Against it, a small blind study by Kristie Marsela, matching AI iterations against the source photographs of one D2C brand, put detection accuracy at 51.08%, with the AI images scoring higher on perceived quality; sample size not stated. The scarce input is the perspective, not the pixels.

The platform is editing the thing you are testing

The sharpest change in creative testing this year was not a generator. In July 2026 Meta flipped ten Advantage+ Creative modifications to default-on — backgrounds, image templates, music, 3D animation, text overlays among them — with no opt-in and no notification, as reported by agency founder Kody Nordquist after finding them running on client accounts. He reports it; Meta published no number for what the edits do. The consequence is arithmetic: if the platform can add music, swap a background or layer text after upload, the ad that won is not the file you uploaded, and what you concluded about hook or format belongs to a variant you never saw.

A batch of 30 videos that each differ in 6 ways isn't a test. It's a lottery with a spreadsheet attached. You get a winner and no idea why it won.

Michael Godzic — X, 1 Aug 2026

The same runs on the buying side. Google's AI Max widens matching and rewrites assets inside an existing Search campaign; agency Mabo ran it across 40 live client accounts for seven weeks and reported 9% of spend generating 14% of conversion value, account ROAS up 43% over the window — its own clients, no control group, everything else moving too. Google sells the measurement as well: Data Manager, now wired into Analytics and DV360, arrives claiming advertisers who connect offline and app data see an average 26% rise in incremental ROAS Vendor or unsourced. Google's own number, selection built in — advertisers able to connect offline sales are the ones who have offline sales worth connecting.

What survives this is testing discipline, not a bigger generator. One vendor framework at least counts the right thing: Creative Testing Velocity measures distinct ideas tested per unit of spend rather than renders or cost per video. The idea is free: if you cannot say how many separate ideas last month's spend bought, the answer is smaller than the file count.

The surface is worth the attention

Retention is the one chapter in this report where the base rate is well documented before AI shows up. Klaviyo's 2026 email benchmarks, drawn from more than 183,000 brands, put automated flows at 41% of email revenue off 5.3% of sends — 18 times the revenue per recipient of a broadcast campaign. The ratio matters more than the 41%. Flows win because they fire on something the shopper just did, which is why they are the first place to spend a personalization budget: the context is already there, and the volume is small enough that better messages are cheap to make.

Klaviyo: automated flows earn 41% of email revenue from 5.3% of sends

Flows are triggered by behaviour, so they are the surface where AI personalization pays first.

Share of sendsShare of email revenue

First-party / researchKlaviyo 2026 Email Benchmarks, 183,000+ brands.

Show data table

	Share of sends	Share of email revenue
Automated flows	5.3%	41%

Read the chart as an allocation argument, not a performance one: a brand that puts almost all of its sending effort into campaigns is spending most of its attention on the low-yield side of the ledger. Every AI retention pitch in this corpus is, underneath, a claim that the same machinery can now be pointed at the high-yield side without adding headcount.

Headless, and who got there first

At K:BOS in September 2026, Klaviyo opened the CRM to agents: 260+ MCP tools and 490+ APIs against live profile data, so a marketer or an agent can read, write and send without opening Klaviyo. In a same-week interview the CEO gave the figure as 400+ APIs — a rounded lower bound rather than a contradiction, but the kind of drift worth pinning to the announcement and its date.

An agent can read, write, and go live on its own. No need to open Klaviyo at all.

Christopher Di Censo, Klaviyo — LinkedIn, 10 Sep 2026

That image illustrates the workflow, not a screen recording — treat it as a diagram. What is verifiable is the tool surface and its date, and the surface is no longer a differentiator: Salesforce shipped Headless 360 in April 2026 and added 60+ more MCP tools on 19 August. Driving your ESP from a chat window is table stakes, so the question for a vendor is not whether they have an MCP server but what their agent may do without a human.

Who approves the send

The StoreCRM beta above is a real capture, and the design choice worth copying is the last step. A merchant describes a branching post-signup scenario in plain language — wait one day or two depending on how much the customer spent — and the system produces a draft, saved inactive, for a person to check before anything sends — the cheapest available guard against an agent with write access to a hundred thousand people.

The org chart above is a proposal, not a deployment: one operator's seven-agent retention department, built to see what was possible. The annotation is the finding — humans keep strategy, creative direction and approval of every send, and the agents are grounded in warehouse data so they cannot invent metrics. Nobody here is running an unattended lifecycle programme, and nobody claims to be.

What a holdout changes

The most useful sentence of the K:BOS week was not about tools.

Personalization is getting a new layer, in preview, to coordinate audiences, content, product recommendations, timing and channels. With holdout testing to measure whether it makes a difference.

Einar Thor, Pipar\TBWA — LinkedIn, 10 Sep 2026

The layer is narrower than its name. Part of it is Klaviyo's Marketing Analytics renamed; what is new is a separate predictive model per decision — who to send to, what to show, when, on which channel — written back onto the customer profile, so a marketer can read what the system believes about a person. Predicted repurchase timing becomes a flow trigger. Some of it, the new segment layers included, was not live on announcement day (Jordan O'Connor's K:BOS breakdown, 12 Sep 2026).

A holdout is a randomly withheld slice of the audience that never gets the treatment, so the comparison runs against people who would otherwise have been targeted the same way. It is the difference between attributed and incremental revenue, and the measurement almost every AI marketing claim in this report is missing. Across eleven chapters this is the only vendor building the correct instrument into the feature instead of leaving the customer to construct it — and it is a preview feature, not a proven result. Nothing has been published about what its holdouts show.

Two operator results in the corpus used that design. At Fellow, a 30-day 50/50 test of AI-personalised messages inside the brand's existing Klaviyo flows reported 70% more browse revenue and 9x incremental ROI. On another account, a 30-day trial of an AI inbox-placement and subject-line tool reported $19,700 incremental.

The lift is measured against a holdout group, so this isn't platform attribution. It's a real incremental number.

Justin Guimera, retention agency — LinkedIn, 16 Mar 2026

Both are published by the party selling the tool: the design is right and the sample is not, because no one publishes the 30-day test that came back flat. Take them as existence proofs — user-level AI personalization can clear a holdout on some accounts — not as planning numbers.

The loop above is the pattern worth stealing from the vendor pitches: generate variants, judge them blind on revenue per send rather than on taste, then put the winner in front of a 5–10% holdout before it becomes the default. The agent count is decoration; the holdout at the end makes the rest of it legible.

What gets claimed without one

A fourth scraping round was aimed at exactly this chapter and came back with 15 qualifying posts, fewer than any other area of the funnel. None was measured against a control. The imbalance is the finding: Klaviyo shipped agents, an MCP surface and a predictive layer inside one quarter, and almost everyone reporting what that did to revenue is selling it.

The widest number in the corpus shows the shape. Win-back emails across ecommerce are put at about $0.11 per send; win-back flows where AI continuously updates products and messaging are put at $5 to $10 per email. That is a 45- to 90-fold gap with no period, no account count and no baseline for the treated group, and it comes from the co-founder of the company selling the optimisation, relayed on LinkedIn (Pietro Saccomani, 4 May 2026). Keep it as the ceiling of what a vendor will say out loud, and for nothing else.

The cart-recovery claims cluster the same way: +73% abandoned-cart revenue in 30 days at an unnamed medical supply brand (CustomersAI, 3 Jun 2026), 20% cart recovery at one wine merchant on WhatsApp, $4,157.70 in fourteen days from a phone agent. In each, a new channel arrived at the same moment the AI did, and nobody separated the two.

Two audits in this sweep put the binding constraint further upstream. One mid-size store — Shopify Plus, Klaviyo, live automation flows — still had only 36.7% of its database active and sent everyone the same message at the same time; the author's ordering is segmentation first, predictive marketing last (Olimpia Valentina Chiriacescu, 6 Jul 2026). An agency that audits Klaviyo accounts puts the typical one at roughly 30% of what the platform already does: three to five flows, a weekly campaign, agentic optimisation never switched on. Both are single-operator reports, and both point where the predicted-CLV gap points.

Production speed is where AI in retention is least disputed and least interesting. Audit outputs like the one above are routine now: 90 days of flow data in, a ranked list of gaps out. The dollar figures on that screen are a model's estimate of opportunity from an unknown account, not booked revenue, and they circulate as if they were. Claims that a flow now takes 5–10 minutes instead of 2–3 hours are probably true, and they are a cost story: one practitioner reports agencies re-quoting retention retainers from $4,100 to under $3,000 a month. If AI only makes the same programme cheaper to run, the saving lands with the client.

Speed is also where the sharpest dissent sits. An email agency spent a year testing the category — copy generators, design assistants, segmentation models, send-time optimisation — and published a sentence no vendor deck contains.

AI is incredibly good at making you faster. It's terrible at making you interesting.

eCom Email Marketer — LinkedIn, 15 Apr 2026

Ten subject line variations in seconds, they report, all ten sounding like one template. A practitioner's read, not a measurement — but from someone with no tool to sell, and it matches the only controlled comparison the round turned up. Digital Applied paired 100,000 cold emails, half AI-written and half human, October 2025 to April 2026: the variable that moved inbox placement was not authorship. Three days between sends landed 93%, one day landed 71%, copy unchanged (relayed by William Doom, 9 Sep 2026). Cold outbound is not permission email, so leave the digits there. The shape carries: the writing got faster and the schedule around it mattered more.

What none of this fixes

One consultant estimate in the corpus is worth keeping precisely because it is unproven: roughly 7 in 10 first-time ecommerce buyers do not return within a year, and Zalando's order frequency has stayed near 5 orders per customer per year for six years, with SHEIN near 4. That is a practitioner's framing, not research, and the framing survives even if the digits do not: none of the AI work above gives a customer a reason for a second purchase. It makes the second ask faster, cheaper, better timed and better worded. That is worth money. It is not the same problem.

The ROI number, and what it measures

Gartner's analysis of 432 customer-service AI use cases found 25% with positive ROI and 42% unclear. Companies run nearly five use cases on average, commit 13% of the service budget to AI, and 42% cannot say what value those systems create. Yet 56% of service leaders expect their incentives tied to AI outcomes this year — pay attached to a number four in ten cannot produce.

The rest of the 432 was not reported, so it is not drawn. Most customer-service AI is unscored, not proven bad.

The gap often starts before deployment: the goal is simply "add ai" · use cases are chosen by volume · containment becomes the headline metric · measurement ends at deflection.

Alexander Abstreiter, AI vendor — LinkedIn, 4 Sep 2026

He sells this software and says so in the same post. The diagnosis still holds.

"Handled" is not "resolved"

ASOS's playbook is the largest production claim in the corpus: about 50% of inbound care requests handled by AI agents, plus 90% company-wide Copilot adoption, reached in phases rather than one launch. It also leaves the load-bearing word undefined. A ticket an agent touched, one it deflected and one the customer thought was solved are three different denominators.

Below it sit vendor cases. Shopify's agentic use-case page reports Klaviyo's Customer Agent resolving 75% of LifeStraw's inquiries without a human, agent-recommended sales up 111% in 90 days — no pre-AI baseline, no scope of inquiry types. Gorgias ambassador Max Wallace reports 20% of conversations automated in the first weeks at Tommy John, disclosing the relationship in his first line. Then the round numbers passed around as common knowledge: ~30% of tickets are "where is my order", AI deflects ~80% of those, every vendor claims ~60% resolution.

The benchmark above is the corrective. Across 107 real commerce workflows the best of 13 models completed 61.7%, on a leaderboard Accio runs itself, with domain task counts summing to 117 against a stated 107. Directional, not exact, and still: top agents fail about four operational tasks in ten. A resolution rate above 70% is a claim about a ticket mix, not capability.

The anecdotes circle the same hole. An agency reports a store's support email halving, 100 to 50 a day. An operator watched cost per interaction drop from $6 to $0.50, then read the reviews, which were furious. A vendor says its "dumber" agent, which hands off below 70% confidence, drew zero complaints in 14 days against three for the smarter one. All tiny, all self-reported, all about what most rollouts never instrument: the customer's outcome after deflection.

Per-resolution pricing changes the arithmetic

Gorgias bills per billable ticket — $10 a month for 50, up to $900 for 5,000 — and the AI Agent adds $0.90 to $1.00 per resolution on top, so a chat the AI closes is billed twice. One missing number now decides it: your fully loaded cost per resolved case today. A thousand AI resolutions a month is about $1,000 in resolution fees plus ticket fees, against the staffed hours removed. The meter is spreading: in September 2026 Salesforce closed its acquisition of Fin, ex-Intercom, 30,000+ customers.

The operational pattern: the agent drafts, a human approves

The clearest working deployments here are not chatbots. They are agents that stop at an approval gate.

Both captures show one loop in Gorgias's Slack agent, Cortex: it audits roughly a hundred human-handled tickets, finds a retention step the AI had been skipping, drafts the Skill update with evidence and a risk tier, then waits. In the second it reviews pending rules and help-centre drafts and recommends which to enable, changing nothing itself. Polar's Operator sells the same contract, "it acts when you approve", and its $3k/day of Meta spend on a broken URL is a self-report of what such a loop finds.

Above is the small-merchant version, posted free by a store operator: rules route the ticket, AI drafts the reply, code verifies the promises in it, a second AI reviews the day's answers, and any refund needs a person. His 0.45% dispute rate over 9,000+ tickets is in the post, not the screenshot. Copy the architecture, not the numbers.

Approval gates scale only while the action is rare and expensive. Albert Heijn's AI produces over a billion automated demand forecasts a day, 17,000 products across 1,200 stores, 50 days out, and nobody approves a billion forecasts. There a human approves the purchase order, the exception, the policy. Amazon is cited secondhand as reporting 35% fewer stockouts from AI inventory tools; the podcast citing it asks the right question, which is when a human should approve the order.

Returns, inventory and fraud: the cleanest numbers, and who published them

Shopify's agentic use-case page carries the tidiest outcomes here: Signifyd's fraud agents taking Cymbiotika to 98% order approval with 93% fewer chargebacks, Prediko saving Kuppa Joy 10 hours a week with 90% fewer order errors. Vendor-selected cases on the platform's marketing page, no baseline stated. SAP's Salling Group figure, up to 66% less time from supplier to shelf, is labelled a projection by SAP itself — never quote it as a result.

Returns suit the approval pattern best and are measured least. AfterShip Returns became one of the first Sidekick app extensions, so a merchant can ask "show me returns that need my approval" inside the Shopify admin. SHIPPED, not proven: no time-saved or resolution figure exists for it, and Dr. Squatch's 25% drop in WISMO from predicted delivery dates is unattributed.

The queue is changing from the other side too: above, a consumer agent opens Amazon's support chat for its owner and extracts a $70 refund a supervisor had promised. One tap for the customer, unlimited polite retries for your team; policy that reads clearly to a person and vaguely to a machine will be tested by machines first.

What separates the deployments that pay

Say the evidence is thin, because it is

Across the whole corpus, 87 of 329 de-duplicated numbers trace to a named first-party disclosure or a research firm with a stated method. In this chapter's original ledger that count is zero. Imagery, product copy and try-on are where operators spent real money in 2026 and where the industry published the least measurement. Read everything below against that.

We tested that rather than repeat it. A fourth scraping round, screened to the same rubric, was pointed at exactly these areas. Try-on, searched for on purpose, yielded four more qualifying posts: a vendor announcing Zara's entry, a listicle about L'Oréal, a note on fit-memory startups, a vendor's own usability observations. None controlled — the thinnest return anywhere in the corpus.

One controlled experiment — and it is about words, not pictures

The sweep did turn up one thing this chapter did not have: a randomized field experiment. Across 35,625 users reading AI-generated summaries of customer reviews, researchers varied one element — how strongly the summary asserted that customers agreed with each other. Consensus framing raised product clicks 24% and cut reading time 10%. The paper is Wang, Zhang, Gao & Tan, "When AI Speaks for the Crowd", September 2026, an SSRN working paper, not peer reviewed, and the post relaying it is its only record in this corpus (post). Treat the effect size as provisional; treat the design as the point. Every other number in this chapter compares self-selected users with everyone else — this one holds the shopper constant and changes only the machine's sentence, and the arm that made people read less sold more. That summary is product content somebody else writes about you, in wording you do not set.

Cost per asset fell. Cost per approved asset did not.

The production story is the most credible thing here, because the people who did the work tell it. A designer at Every Man Jack produced 2,800 new product images in a couple of months, white-background main shots through to generative lifestyle imagery, syndicated through the company's PIM into Amazon, Walmart.com, Target.com and Kroger.com. A separate practitioner reports 5,000 AI image edits for about $200 — and a week of human QA on top. That second half is the number to plan around: generation is nearly free, verification is a salaried person's week.

The failure mode is not a bad picture. It is drift across a set.

By pose seven, something shifts. The jawline softens. The proportions stretch. A PDP built from ten different women isn't a catalog. It's a casting error nobody caught.

Walid M Samuel, visual production at Pixofix — LinkedIn, 9 Sep 2026

Every Man Jack's VP of marketing says the same from the CPG side: models still botch small captions on packaging, so designers stayed in the loop rather than being replaced by one upload.

Try-on: the loudest numbers are the weakest

The commercial case is easy to state. Online fashion returns run 20-40%, wrong size the leading cause; US retail returns total about $816 billion (NRF) and cost 20-30% of revenue. Only the $816B and the cost share are sourced; the rest circulates unattributed.

What try-on returns in exchange is where the evidence gives out. The capability is real and now buildable by one person in a weekend, as the prototype above shows. The results attributed to it are not. A widely shared post cites a 2026 study of 1.2 million fashion shoppers at 50% higher purchase conversion and 3x more cart additions among try-on users, and a second of 577 Shopify stores at 3.1x add-to-cart, 90% of them on mobile. Neither publisher is named anywhere in the corpus, and both compare people who chose the feature with people who ignored it — a shopper who uploads a photo of herself to see a dress on her body has already decided she wants the dress.

What 2026 changed is the scale of deployment, not the quality of the proof. Zara entered virtual try-on, the clearest mainstreaming signal in the corpus. L'Oréal's are the largest numbers anyone quotes — 1 billion virtual try-ons, 20 million skin diagnostics, 3x higher conversion — reaching us through a marketing listicle on X, no period, no baseline, no L'Oréal document behind them (post). Irisphera, posting off Zara's move, claims 45,000+ products processed, 5-15 second renders, up to 3.5x conversion — and "20,000+ successful purchases with zero returns" (post). Zero returns on twenty thousand fashion purchases is not a measurement. When a vendor's headline number is impossible, read the ones beside it as advertising too.

The part with a checkable payoff

One place this work pays off without trusting anyone's study: retrieval. Chapter 02 documents the shift — on 10 July 2026 the share of ChatGPT Shopping recommendations sourced from merchant feeds jumped from 8.26% to 61.54% in a single day (Profound, ~1.75M prompts), and Shopify reports AI searches using its Catalog convert at twice the rate of those relying on scraped data. Retrieval runs on structured attributes, not art direction.

Adobe's Content Visibility Checker puts average language-model visibility on retail product pages at 66% (post): a third of what you wrote, shot and priced never reaches the software doing the shortlisting. Syndigo's State of Product Experience 2026 survey (8,736 consumers, six countries) has the shopper side — 83% leave for a competitor when the information they need is missing, 26% are more likely to buy what an AI recommends (post). Coverage and attitudes, not lift: they size the leak, they do not price the fix.

That reframes what "good imagery" means. Google's own guidance is unusually concrete for this chapter: crisp, multi-angle assets, up to 6% more clicks, and a 500x500 px floor — up from the legacy 100x100 — enforced by disapproval from 31 January 2027, with 1,500x1,500 recommended and secondary angles supplied through additional_image_link. Alt text is that job in text. Catalogue audits are it again, with the same evidence problem. A brand strategist put a few live product pages through an AI auditing tool and got back 511 inconsistencies: a missing video on a Caudalie page, a blurry image and mismatched pricing at London Victorian Ring Co, wrong descriptions on LG pages — drift no merchandising team reads its way through at catalogue scale. The recovery figures she quotes (Caudalie €700k recovered, LVR +7% conversion, LG $900k captured) come from the tool's vendor, in a post offering readers free access to it. ELEMIS reports an Akeneo agent turning one launch's three-hour product-data investigation into 30 minutes, a human approving before anything went live; the source is an Akeneo employee. Run the audit. Treat the numbers beside it as marketing.

It has forced us to really critique our imagery — and evolve our focus from aesthetics to retrievable, qualifying information. Better not just for agents, but human shoppers too.

Nicklaus Hasselberg, VP Marketing & Omnichannel, Every Man Jack — LinkedIn, 18 Aug 2026

Disclosure became a legal surface in 2026

Two rules landed while the measurement question was still open. A New York law (GBL §396-b) took effect in June 2026; Amazon began requiring sellers to tag AI-generated people in product photos and video the following month, and TikTok and Meta label realistic AI content automatically. In the EU, the AI Act requires visible disclosure of AI-generated or manipulated imagery, media published before 2 August grandfathered; MANGO, Adidas and Boozt were among the first storefronts seen complying. Tolerance is its own variable: one performance marketer wrote that AI models on an apparel site "completely threw me off from considering a brand".

Nine numbers, nine different comparisons

"AI traffic converts better" moved more budget in 2026 than any other claim, and nothing in this report has less agreement about what it means. Nine versions circulated, from +31% to +100%, and no two measure the same thing.

Adobe's +42% (March 2026) and +60% (July 2026) compare AI-referred traffic with all non-AI traffic — paid search, email and affiliates included, a mixed bag holding some of the worst traffic a retailer buys. Shopify's ~50% (Q1 2026) compares with organic search only, and only for sessions landing on a product page: a much harder baseline, which is why the smaller number is the more impressive one. Visibility Labs' 31% is ChatGPT against non-branded organic. Ethercycle's +46% is ten stores; Netalico's ~2x is nine.

One publisher measuring one thing does not give a stable number either. Three Adobe readings of the same comparison run -38% (March 2025), +42% (March 2026), +54% (May 2026). The sign flipped inside a year. A plan built on the March 2025 reading would have written AI referrals off as below-average traffic. Three points is not a trend, and the path between them is unknown.

What this report refuses to use

Seven widely-shared figures did not survive checking. They are named rather than quietly dropped, because all seven are still circulating.

Correlation wearing a lift's clothes

The second failure mode is quieter, because the sources are impeccable. Walmart told investors Sparky users spend about 40% more per order than non-users (post); Amazon reported over 40% for its assistant; Lowe's says Mylow users convert at roughly 3x non-users (post). Real disclosures, named executives, and not one of them a lift.

A shopper who opens an assistant mid-basket is already further into a purchase than one who does not, and an earnings call cannot separate cause from consequence. The contrast in the figure is the useful part: Albertsons' +26% and +10% compare one experience against another — a comparison that can come out wrong. A case study of users versus non-users shows you an audience, not a result.

Small bases move fast, then stop

Shopify's two 2026 disclosures sit badly beside each other: a Q1 blog reporting over 8x AI sessions and about 13x AI orders, then a Q2 call reporting roughly 3x for both. Both are official and nothing broke in between — Q1 was measured off a near-zero base a year earlier. Multiples decay while absolute volume rises. The absolute share is the figure nobody quotes.

On a store we work with: ChatGPT is 0.27% of revenue, +42% MoM. It's a small slice with a steep curve. The right approach here is "monitor monthly" not "overhaul everything."

Kurt Elster, Shopify consultant — X, 12 May 2026

The plumbing moved under the dashboard

Rank stopped predicting citation. Across 863,000 searches and four million AI Overview links, Ahrefs found the share of citations coming from a page already in the top ten fell from 76% to 38% (post). A team still reporting "we hold position 3" is reporting a proxy that lost half its predictive power.

The labels moved too. In mid-June 2026, ChatGPT sessions on this site stopped arriving as referral and (not set) and began arriving under a new ai-assistant medium; every saved report filtered on referral showed AI traffic collapsing on a Tuesday. Nothing collapsed. Then there is who is at the other end: Cloudflare puts bots ahead of humans at 57.5% of website traffic, and Lunio's 2026 report puts invalid ad traffic at 24.2% on TikTok and 7.6% on Google — discount that second one, because Lunio sells click-fraud protection.

Trust is a ladder, not a number

Trust figures look contradictory until you notice the surveys ask about different rungs. Researching with an AI, letting it build a cart and letting it spend money are three decisions, and the drop-off between them is steep.

Visa's Trust Index puts the bottom rung at 23% of 2,065 US consumers willing to trust GenAI with a payment — and 61% when Visa is named as the party handling it. Trust transfers to an institution; it is not extended to the model. That is a design instruction: let the assistant research, and put a name the shopper already banks with on the payment step.

The costs that arrive without a dashboard

Agentic checkout moves a liability before it moves a number: under the Agentic Commerce Protocol the merchant stays merchant of record and keeps refunds, chargebacks and compliance. Meanwhile the tolerance narrowed. On 1 April 2026 Visa's VAMP excessive-dispute threshold dropped from 2.2% to 1.5% for US, Canada, EU and APAC merchants, with fees around $8 per disputed transaction — secondhand from a conference write-up, so check the Visa bulletin before it enters a model.

When a customer disputes what their agent bought, who produces the evidence?

Richard Emanuel, on an Anthropic agentic-commerce session — LinkedIn, 9 Sep 2026

Produce your own number

The way out is not a better source; it is a measurement you ran yourself. Here is what the correction usually looks like: dashboard lifts of 68–112% in free-listing clicks shrank to +18 to +80 percentage points once seasonality was removed and an untreated control market used. Two-thirds of the headline was the calendar — assume that error inside any uncontrolled before-and-after, including your own.

The pipeline as it is assembled today

Nobody starts at the generator. Every serious version starts in a competitor ad library.

it starts in the meta ads library, pulling the winning ads in the niche before writing a single line

Machina, AI ad operator — X, 9 Sep 2026

MAX gives the model product, audience, offer and example ads, plus browser access, then tells it to open the Meta ad library and search five named competitors. Mike Futia's scrapes competitor pages, watches every video with a model and returns a brief with ten concepts in the brand's voice; his TikTok variant turns winning videos into shot-ready briefs. Step one is research, and it is the step most brands skip.

After the brief the chain is short. Script and shot list come out of the same conversation; a storyboard image is generated first as the anchor — one reference image driving every beat and camera angle — which stops the shot-by-shot drift that makes AI video obvious. Then it is a model call: Seedance inside a Claude skill, Higgsfield over MCP for research, scripting, images and animation in one conversation, Gemini Omni for multi-shot ads with a consistent creator across scenes, Kling over MCP for talking head, product shots, narration and music in one pass. On their own timing a variant takes minutes, a new concept a working day including research.

Two details separate daily operators from demos. Prompting: Stephen Bishop runs three fixed clauses on every generation, skin texture first, and argues people switch models when they should fix the prompt. Editability: Higgsfield's Motion Designer returns an After Effects project where camera, lighting and typography stay editable — an asset you revise rather than regenerate and hope.

The step nobody screenshots is the one Shann Holmberg draws: agents only produce output as good as what they were given. Performance data, customer language, brand rules and competitor context in one structured place, loaded before anything is written. Skip it and you get generic output, then blame the model.

Where the human still sits

Three places: choosing the angle, approving the cut, owning the account the variants go into. The clearest evidence that generation is not the hard part is a company reversing. Icon raised $9.2M for an AI admaker, then pivoted to running the human UGC workflow — creators, shipping, scripts, coaching, editing — at six filmed ads for $999, keeping AI for everything around the shoot.

A company that originally built an AI admaker ended up putting humans back into the part of the process that actually produces the creative.

Divyanshi Sharma — LinkedIn, 25 Aug 2026

Same shape in imagery, better sourced in chapter 10: 5,000 AI image edits for about $200, behind a week of human QA.

What is winning attention right now

Operator opinion, flagged as such. Song ads are the format of the moment, and the operator running them is their sharpest critic.

They retain attention. They get cheap clicks. They can absolutely crush CPA. But cheap acquisition and great acquisition are two different things.

Franky Shaw, ecommerce operator — X, 12 Sep 2026

Claymation is second, on an arbitrage argument: it reads as expensive, almost nobody runs it, and an eight-step workflow replaces the studio. Three months later it is taught in French and packaged into Chinese cross-border tooling — the half-life of a creative arbitrage. AI personas are third: one TikTok Shop account runs a single synthetic presenter across dozens of products, top videos at 1M–4M views by the poster's count. Long-form autopilot video is fourth. Static remix is fifth and least discussed, though most merchants could run it tomorrow: 45+ on-brand statics from one product and one angle, or Oliver Kenyon's tighter version, six carousel images where each has a defined job. That one sits on the PDP, not the feed, and it is the cheapest funnel test here.

The funnel under the format is moving too. Meta now answers shoppers inside the ad unit as suggested questions, then carries the conversation into Messenger, WhatsApp or the brand's site. If that holds, the ad's job shifts from carrying the pitch to starting a conversation something else finishes — which makes the hook, not the close, the part worth generating fifty of.

The economics operators report

What it replaces, in their accounting: creator fees of $150–500 a video, or $300–2,000 at 30–50 videos a month; retainers of $5,000–20,000 a month for 15–20 videos; a tool stack one post prices at $10,000+ a month. None are audited; their consistency tells you the market price, not the return. One operator did better than a claim: he ran the same listing task three ways and scored the outputs himself. Copy the method, not his verdict.

The durable half: judging creative, not producing it

Take the video out and what remains is a research stack that is aging better. Teardowns that score every creative an account ever ran and say why. Systems that scrape Instagram and TikTok and rank clips by sales intent. Vendors turning the Meta ad library into a chat over three million ads. Workspaces giving a brand's agents one context library instead of five silos, like Shadow. Skill libraries too: a marketing pack for coding agents credited with more than 48,000 GitHub stars, and a product-photography workflow one person spent 15 hours and $478 to nail, then published free.

The most interesting build is a scoring loop: ten runs against a written eval, prompt rewritten, retested, winner kept. A hook writer scored 32/50, then 47/50. His eval, his scale, no outside check. The shape is right though: if ~5% of creatives carry 55% of spend, the scarce skill is a rubric that predicts which 5%.

Rubrics also industrialise in a way generators do not. Here a small distilled model scores 84.6 against its much larger teacher's 83.0 on buyer profiles, with an eight times shorter prompt and daily output rising from 2 million to 72 million: once judgment is written down, it gets cheap enough to run on everything. The attribution to Shopify comes from the post, not the slide — read it as a pattern, not a disclosure.

Where it breaks

Ratio, not volume. Making 500 ads instead of 50 multiplies the denominator. If you cannot tag variants and read which carry the spend, more creative is a larger bill.

Slop. A practitioner ranking ecommerce ad formats puts AI UGC in F tier, below repurposed TikToks, the same week other operators call it the whole business. Both are opinions; the disagreement is the finding.

Brand risk compounds. A synthetic presenter across dozens of unrelated products is arbitrage with nothing to damage; a brand whose customers meet the same invented face next quarter is not. A vendor selling "the world's first ad that nobody knows is an ad" is describing a disclosure problem as a feature.

Disclosure is already law. The EU AI Act treats AI-generated imagery as a deepfake and wants it disclosed visibly to EU shoppers; by mid-August 2026 Linda Bustos could find three ecommerce sites doing it — MANGO, adidas, Boozt. Put the label in the asset template now; retrofitting 500 files is the expensive version.

What a merchant should copy

1. Build the input layer first

   One place holding performance data, customer language, brand rules and five competitors' live ads.

2. Buy the judging before the generating

   Tag every variant at upload so you can answer which 5% carry the spend. Without that, stop here.

3. Start with statics, not video

   A six-image PDP carousel where each image has a job is a same-week test with a readable outcome.

4. Run one fixed task three ways

   Same brief, same inputs, three tools, your own scoring sheet — the only cost comparison that means anything for your catalogue.

5. Keep a person on the angle and the cut

   The model is good at twenty variants of a chosen angle. Choosing which angle deserves twenty is still the job.

From the click: landing pages and PDPs

The cheapest thing to build right now is a page audit. Yumna Aziz published the whole pipeline in n8n on 19 August 2026: a form takes the URL, an HTTP node fetches the HTML, it is converted to Markdown, and Gemini 2.5 Flash returns prioritised CRO fixes with line-by-line rewrites — one run at 29 seconds and 2,381 tokens, she reports Vendor or unsourced. Mike Futia built the same idea as a Claude Code plugin covering technical SEO, product schema, Core Web Vitals and AI-search readiness, scored 0–100, and reports it replaces a $200/month Ahrefs subscription Vendor or unsourced.

None of them tell you which fix to ship. They read copy, not behaviour, and the ranking they return is an opinion with a number on it. Point an agent at marketing work and the answers come back "correct, but not much use" — which is why the popular remedy is a skills pack (marketingskills, ~50 marketing skills, 48,000+ GitHub stars) rather than a bigger model. Operators keep buying context, not capability. The page still decides the sale: Contentsquare found 97% of shoppers would reconsider an AI-recommended purchase if the brand's site falls short Named but secondhand — a survey about a hypothetical, so a direction, not a rate.

The best-documented PDP workflow here is Connor Gillivan's week inside Glara, an agent reading a live Shopify catalog rather than an uploaded export, where nothing publishes without approval and any change rolls back.

Glara had already drafted the fix, and it wasn't a rewrite. It was one line: "when to use." Mist onto face, pat in, don't rub. I read it, changed a word, approved it. That was the whole job.

Connor Gillivan — LinkedIn, 8 Sep 2026

On-site: assistants, finders and the handoff

Product finders are the older, better-measured half of this. THG's Foundation Finder for LOOKFANTASTIC reports 5% revenue uplift Named but secondhand — behind it, 3,400 foundation samples read under spectrophotometers against 6,100 shade options. That is what a working finder costs.

The rest of the category looks like the image above: a selfie in, styling advice and a generated look out. It is a vendor demo, not a deployment, and no conversion number has been published for it. Screenshots run well ahead of evidence here.

This one is a real capture, and it is the moment that matters: the assistant asks which colour before adding the fleece, then returns a two-item, $290 bag. Once an assistant can write to the cart, the question stops being answer quality and becomes permissions and escalation. Max Wallace, who discloses he is a Gorgias brand ambassador, reports 20% of conversations automated weeks after launching its AI Agent and Shopping Assistant Vendor or unsourced. Which model sits behind that permission is a procurement question. COLIBRIX ONE benchmarked 20-plus ecommerce agents over 2.4 million runs: the top-ranked model costs 14 times more and answers 8 times slower than the runner-up for three points of quality, and when planted text told the agent to run an unauthorised command, only one of the twenty refused Vendor or unsourced. The ranking is theirs; what it measures is the point. On a surface that can write to a cart, latency and injection resistance are the specification. Josh Gonsalves built the small version in Notion Custom Agents: the shopping assistant and the support agent are one machine seen from opposite ends of the funnel.

Lifecycle: the week that changed

Retention is furthest along because the tools opened first. Jordan O'Connor spent four days testing the Klaviyo × Claude integration on a client doing over $10M a year, and his framing is what to keep: the model is no longer analysing the account, it is building inside it. Tyler Phillips published the copyable version on 21 April 2026: load the brand guide, product doc and CDN image URLs first, connect Klaviyo over MCP, then ask for a three-email abandoned-checkout flow for a named segment and let it build the templates and wire the flow IDs. A month of calendar in under an hour, one variable tested per campaign. Kumud Deepali Rudraraju's Omnisend post is an ad; the part worth taking is that its MCP connection is read-only, so an audit runs without write access to the list. O'Connor's later write-up of K:BOS, 12 September 2026, records the same move as shipped product: 260-plus Klaviyo MCP tools, so an agent can set a flow live without opening the app, and predictive timing pulled inside the flow — a trigger that fires when a customer is likely to buy again rather than on a fixed delay. Timing is the part a merchant cannot draft by hand.

What changed in the week is throughput, not judgement. Chase Dimond's agency shipped 40 client flow emails in one week without adding a retention strategist — "the queue just stopped being the ceiling." Aim it at the dead flow, not the calendar: average revenue per send on a post-purchase win-back is $0.11, against $5–17 for brands doing it properly Vendor or unsourced.

The demo above answers "why is retention low?" by pulling ninety days of cancellations out of Shopify, Recharge, Klaviyo and Gorgias in about four minutes. The figures on screen are demo data — judge the shape, not the result. A general-purpose model cannot do it, because — as one email operator puts it — it "doesn't know your welcome flow is working and your win-back is quietly dead". Shann Holmberg's email walkthrough starts from the account's real open rate and revenue per send. Copy that order of operations.

After the sale is still the funnel

Support is a conversion surface that only ever shows up in the retention line. Dr. Squatch is reported to have cut WISMO contacts by 25%, at 94% delivery-date accuracy, by predicting delivery problems from carrier data instead of answering questions about them (Malik Usman, 7 Sep 2026) Vendor or unsourced. The same brand worked the other side of that page: recommendations placed on the order-tracking screen itself, clicked by 31.89% of customers and credited with $32,978 and 766 orders in Q1 2026 (Jalpesh Patel, 10 Sep 2026) Named but secondhand. A denominator and a period, which is more than most figures here carry — but no holdout, so it sizes the surface, not the lift.

Deflection is not the end of it: the answer itself is a place to recommend. Zak Cassady-Dorion's agency has Klaviyo's Customer Agent settle an address correction and then recommend a product off that order, and reports roughly a 10% lift on abandonment flows after swapping the discount code for a real reply to "what stopped you?" Vendor or unsourced. Shopify's write-up of LifeStraw puts a figure on both halves — 75% of inquiries resolved without a human, agent-recommended sales up 111% over 90 days Vendor or unsourced. The resolution rate is the number every vendor publishes and the wrong one to optimise alone. Between 25 April 2025 and 31 January 2026, e-commerce was the largest of 31 categories on India's National Consumer Helpline, at 47,743 refund grievances — behind a growing share of them, a bot with no visible route to a person Named but secondhand. A ticket closed fast and a problem solved are different events.

On somebody else's surface it gets worse, and the screenshot above is the warning: a seller reports being billed when shoppers click the assistant's suggested questions on his own listings — 1,815 interactions, $2,941, roughly 70% with no sale attached. On your own site you set the price of an assistant conversation. On a marketplace, somebody else does.

Merchandising, pricing and the catalog

Shopify's AI Toolkit put a write-capable agent in the terminal. Mike Futia's version is one prompt that reads the catalog, rewrites descriptions and pushes them live; Edward Deng's list of seven jobs adds re-ordering collections by what converts and post-purchase sequences written from order history. One item there is dangerous: inventory-aware urgency copy, rewritten automatically, is a claim about stock that has to be true. The remedy operators use is to shrink what the agent may touch: in Maxim Lazovsky's supplier pipeline the model normalises titles, extracts weights and drafts descriptions into a Shopify draft, but SKU, barcode, price, inventory and variant IDs stay protected, and an anomalous change is held for review automatically. ELEMIS runs the same split through Akeneo — the agent proposes a fix with its expected impact, a person approves — and reports one launch investigation dropping from three hours to thirty minutes Vendor or unsourced.

This is the checkable payoff, and it points back to chapter 02's feed shift. Google Merchant Center is the first Google product to expose query data for AI Overviews and AI Mode, split by journey phase and query type; the merchant above holds 1.0% share of voice where a competitor holds 8.3%. That is a number you can act on without trusting anyone's case study. Marketplaces have the cheap version: Ruben Alikhanyan's audit is five buyer questions and fifteen minutes — he typed them into Rufus for a $4M brand's hero SKU and it was cited zero times. Pricing has no such instrument: the claim that AI Mode shows the same products at 21.6% higher prices is summarised secondhand from a study whose publisher is unnamed — a prompt to check your own feed, not a number to repeat.

The seller's posture

Most pitches this quarter look like the card above: an agency case study, three precise percentages, one named brand, no control group, no denominator. Not dishonest, not evidence. What a deck never contains is a retirement. Starbucks spent nine months counting backroom shelves with iPad Pros and computer vision, then ended the programme in an employee newsletter Named but secondhand: reflective surfaces double-counted, staff were disciplined for the model's misses, and the inventory system underneath was 1990s AS/400 that could not hand it clean data. None of that was a model problem. Three buckets.

1. Adopt now

   Anything that drafts for a human to approve and can be rolled back: page and listing audits, flow and segment drafting over MCP, catalog rewrites with the money fields locked, WISMO prediction, and recommendations on a page you already own, starting with the order-tracking screen. Worst case is a wasted hour. Simon Jackson's library of 50 store agents rests on the right premise — "most of them are checklists, not judgement" — and he reports a first run flagging 17 dead SKUs and $2,300 of discount leakage Vendor or unsourced. Start there.

2. Pilot with a holdout

   Anything that claims revenue. Retention is the one surface where holdout testing is already normal; copy the design before you copy the numbers. Dr. Squatch's tracking-page result is the shape to repeat this way: a denominator, no control, one withheld cell would settle it. A 30-day 50/50 split costs nothing but patience, and it is the only thing that turns someone's case study into your own.

3. Ignore until someone publishes a controlled number

   On-site assistant lift, AI dynamic-pricing margin claims, and every user-versus-non-user comparison in this report — two more landed this quarter, Loops at A101 Ekstra, +83% conversion, and Constructor's Ask Cleo, 8–9x add-to-cart, each comparing shoppers who chose the agent with shoppers who did not. Buy an assistant for deflection if that maths works; never on the strength of a gap between people who chose the tool and people who did not.

Timeline

Only events with a date a merchant can check. Quarterly figures are filed under the day they were disclosed, not the quarter they describe.

Glossary

Where two sources in this report use a word differently, the disagreement is noted.

Agentic commerce

Shopping where software acts on the buyer's behalf: it searches, compares, assembles a cart, and in the strongest version pays. The forecasts disagree wildly because they draw the line in different places — McKinsey counts spend agents orchestrate (up to $1T in the US by 2030), Bain counts spend agents complete ($300–500B). When you read a number, ask which of the two it counts.

AI referral traffic

Visits arriving from an AI surface — ChatGPT, AI Mode, Gemini, Perplexity, Copilot. Measured inconsistently: as a share of referral traffic (Bessemer's 15–20%) or of all traffic (Bernstein's 0.8–1.4%). Same phenomenon, denominators 15x apart.

Feed-integrated retrieval

An AI assistant answering a shopping question from structured merchant product feeds rather than by reading web pages. After 10 July 2026 this is how most ChatGPT Shopping recommendations are sourced.

GEO / AEO / AI SEO

Optimising to be cited or recommended inside AI answers. The tactics that survive checking in this corpus are unglamorous: complete structured product data, entity clarity, third-party mentions. Rank alone predicts citation far less than it did — AI Overview citations from the top ten fell from 76% to 38% in a year.

UCP · ACP · AP2 · MCP

The plumbing. UCP (Universal Commerce Protocol, Google-led) and ACP (Agentic Commerce Protocol, OpenAI/Stripe) let an agent read a catalog and place an order; AP2 handles agent payment authorisation; MCP is the general connector standard that lets a model use a tool — the same standard Shopify and Klaviyo expose their admins through.

Agentic Storefront

Shopify's surface for agent traffic: catalog syndication, agent-readable policies, order attribution back to a channel. Enabled by default for eligible stores; blocking crawlers in robots.txt does not stop Catalog syndication.

AI UGC

Ad creative in the visual language of a customer testimonial, generated rather than filmed. Cheap enough to change the production question from "can we make one" to "can we judge fifty" — see Chapter 12.

Holdout

A slice of the audience deliberately left untreated so the effect of a change can be measured. Almost nothing in this corpus has one, which is why so many numbers here are comparisons rather than lifts.

Self-selection

The reason "assistant users spend 40% more" is not "the assistant adds 40%". People who open an assistant were already further along. Chapter 11 is mostly about this.

Collection

Three sweeps of LinkedIn and X between 14 March and 14 September 2026, using 58 search queries covering every commercial surface: discovery and AI search, agentic checkout, personalization, recommendations, funnel and CRO, email and SMS, ads and creative, pricing, support, product content, try-on, merchandising, inventory, and Shopify's own AI features. That returned 6,840 raw posts, 5,328 of them unique after removing reposts and near-identical text.

Screening

No keyword filter decided what stayed. Every post was read and judged on four questions: is it substantively about AI applied to selling online; does it carry a concrete tactic, case or number; is it crypto, hiring, event promotion or an empty pitch; and which commercial surface does it belong to. Crypto was excluded outright — 177 posts, many of them using agentic-commerce vocabulary to sell a token. Off-topic accounted for 1,369, thin self-promotion 442, generic hype 339, hiring 100. What remained was ranked and the top 1,000 kept.

As a check on the screening, 21 verdicts were pulled at random and re-read by hand against the post text; all 21 held. That is a small sample. It catches gross failure, not drift.

The evidence ledger

Every number in the 1,000 posts was extracted with its claimed source, period, population and baseline: 711 raw claims. Those were then consolidated into 329 facts, because the same figure repeated by forty accounts is one piece of evidence, not forty. Each fact records the organisation it originated with, the posts that carried it, and a reliability grade applied uniformly:

• First-party / research a named company disclosure (earnings call, official data release, documentation) or a named research firm or survey with a stated method and sample.
• Named but secondhand the source is named but the post is a retelling, or it is an agency/merchant case with concrete specifics but no independent check.
• Vendor or unsourced a vendor's own launch or marketing claim, an anecdote, or a number with no attribution at all.

Where facts disagreed, both were kept and linked, with the reason for the difference stated: a different quarter, a different baseline, a different population, or a misquote. Ninety-nine facts carry such a link.

Figures

Every image attached to a kept post was opened and classified — 1,099 of them — then the 382 plausible ones were re-examined against five tests: does it carry real information, will it read when printed, is the number in it sourced and consistent with the post, is it a genuine capture rather than a mockup, and is it a duplicate of a stronger image. 161 survived. Screenshots that turned out to be AI-generated illustrations carrying invented numbers were rejected, as were charts whose labels contradicted their own plotted values.

The second pass: the agentic-commerce corpus (Chapters 12–13a)

Chapters 12, 13 and 13a were built later and from a separate sweep, run on 14 September 2026, because the first corpus covered agentic commerce as one surface among fifteen and the subject had outgrown that share. Three pulls of LinkedIn and X returned 11,908 raw items, 9,105 unique after de-duplication. Two buckets were defined: posts whose own text contains the phrase "agentic commerce", and posts that do not but report a named platform's move in AI shopping, agent checkout or merchant AI tooling. Only posts carrying at least one image were considered, which narrowed the field to 3,780 candidates.

Those candidates were read one at a time — 2,728 of them — and graded on two axes: is the post or its author promoting crypto, and is the post substantive about agentic commerce or merely brushing against it. 249 were dropped as crypto and 587 as incidental (hiring ads, event promos, hashtag stuffing, stock chatter, and a large family of templated "I built a store with Claude in an hour" posts). The remaining pool was ranked and cut to 1,500: 500 phrase posts and 250 platform-move posts from each network.

All 1,840 attached images were opened and classified individually; 1,032 carried enough legible information to be usable and 31 are printed in these three chapters. Quotes in these chapters were machine-checked against the source post text: 26 that turned out to be paraphrase rather than verbatim were sent back and rewritten, and one claim that carried a figure absent from its cited post was removed outright.

Two differences from the first pass are worth stating plainly. Crypto was excluded far more aggressively here, because on X roughly a third of everything using the phrase was promoting a token — a stricter rule also cost the corpus a handful of legitimate posts whose authors work at crypto-native payment companies. And the monthly post counts in this second corpus cannot be read as market volume: the X sweep took the top posts per month and the LinkedIn sweep ranked by relevance, so the shape reflects the sampling, not the discourse.

What this method cannot do

• It measures discourse, not the market. A corpus of posts over-represents whoever posts: vendors, agencies, consultants and people with something to sell. Quiet operators are absent.
• It is English-first. Non-English posts were kept when relevant, but the queries were English, so the view of China, Japan, Korea and Southeast Asia is thin.
• It cannot verify a first-party claim. When Walmart says its assistant users spend 40% more, this report records who said it, when, and what it compares. It cannot audit the number.
• Six months is short. Several of the strongest findings — the July feed shift, the ads moving into answers — are weeks old at publication and may not hold.

Tooling

Collection ran through Apify actors for LinkedIn and X. All screening, number extraction, image classification and writing was done by Claude (Opus 5) under the rules above, with the outputs checked against the source posts. Both language editions were written independently from the same ledger rather than translated.

The ledger

Every first-party and research-grade fact used in this report, with its source, period, how many posts carried it, and a link to one of them.