August 31, 2026

Retailers Put Numbers on the AI Channel, Google Puts a Checkout in the Answer

Period of August 18 to August 31, 2026. Two weeks rather than one, and they were unusually dense. Q2 earnings season put hard figures on AI-assisted selling for the first time at scale. Google added a booking checkout to AI Mode that completes in Google Pay while leaving the partner as merchant of record. OpenAI switched on advertising across 31 European markets, exactly on the schedule it signalled a fortnight ago. Cloudflare shipped a robots.txt control that finally separates “do not train on me” from “do not find me”. And the trade body responsible for ad measurement admitted it has no way to credit an AI-influenced sale, and put a date on the fix.

Last period we covered ChatGPT advertising arriving in Europe, the finding that 96 percent of advertisers never appear as a source in the answer above their own ad, and the first server-log tooling for agent traffic. The through-line was cost and visibility: what the AI channel charges, and whether you can see anything happening inside it.

These two weeks answered the first half of that question and sharpened the second. The channel now has published performance numbers from four large retailers, a live paid slot in Europe, and a Google surface that completes a transaction. What it still does not have is attribution. That gap now has a date on it, and the date is after peak season.

Here’s what happened.

Q2 earnings: the AI channel gets its first real numbers

Three large US retailers reported within about a day of each other, and all three volunteered AI commerce figures without being pushed.

Walmart said on its 20 August call that customers using its Sparky assistant spend 40 percent more per order than customers who do not, with the Sparky user base up 70 percent year over year and global ecommerce up 23 percent in the quarter. CEO John Furner gave the example of a shopper asking Sparky for a week of high-protein meals and getting recipes and meal kits back as an addable cart, with ingredients they had already bought filtered out.

Lowe’s said on its 19 August call that its Mylow assistant, which also powers the associate-facing Companion app, has fielded more than 25 million questions from customers and associates since launch, and that online customers who use it are three times as likely to convert as customers who do not. That was CEO Marvin Ellison’s phrasing. Lowe’s online sales grew 15.7 percent year over year, the second consecutive quarter above 15 percent.

Target reported on 19 August that digital traffic from external AI platforms is growing at more than 3.5 times the industry rate, and that AI-powered wish list recommendations drove wish list creations up more than 50 percent with items added more than doubling. CEO Michael Fiddelke credited the company’s agentic commerce partnerships and then, to his credit, said the AI-driven volume remains “small in total today”.

Two figures from outside the period round out the picture, and both are worth dating precisely. Albertsons put average order value up 10 percent where conversational search is used, and up 26 percent where the assistant matches recipes and ingredients to dietary preferences, figures given by Jill Pavlovich, its SVP of digital shopping experiences, to the Wall Street Journal on 17 August rather than on an earnings call. And Andy Jassy told Amazon’s 30 July call that more than 350 million shoppers used Alexa for Shopping, the assistant that absorbed Rufus in May, over the past twelve months, and that US customers using it spend 40 percent more per order.

Why it matters for merchants: These are the numbers you will be shown in every vendor deck for the next six months, so it is worth being precise about what they are and are not.

Every figure here is a comparison between users and non-users of an assistant. None of them is a controlled experiment. People who open a shopping assistant and ask it a question are, on average, further along and more committed than people who do not, so a large part of any 40 percent gap is selection rather than causation. Read them as evidence that the channel is real and that engaged shoppers behave differently inside it. Do not read them as a promise that bolting an assistant onto your store lifts basket size 40 percent.

What survives that discount is still substantial. Notice, though, that all of these except one measure behaviour on the retailer’s own surface: Sparky, Mylow and Albertsons’ conversational search are assistants the retailer built and controls. Target’s number is different in kind. It measures inbound traffic from external AI platforms, which is the number closest to what most merchants actually care about, and it is the only one of the set that reflects whether anyone else’s model can find you. That traffic is not won with an on-site chatbot. It is won by being legible to the model doing the recommending, which is the same argument we have been making since most catalogues turned out not to be ready.

Google adds a checkout to the answer, and hands the merchant relationship back

On 27 August, Google added three new travel capabilities to AI Mode in Search, one of them transactional: flight price tracking across more than 300 partner airlines and travel sites in over 180 countries and territories, award pricing in points and miles for an initial set of airline and hotel loyalty programmes, and a hotel checkout that completes through Google Pay. The booking flow runs behind a “Continue on Google” option: the user picks a room type, reviews the cancellation policy, and pays without leaving the surface. Ten partners are named for the rollout, Booking.com, Choice Hotels, Expedia, Hilton, Hotels.com, IHG, Marriott, Priceline, Trip.com and Wyndham, and the hotel or booking platform acts as merchant of record and handles customer service. It is United States only, English only, with partners activating over the following weeks. No European availability was announced.

Why it matters for merchants: The merchant-of-record detail is the whole story and almost every write-up buried it.

Google built a surface that takes the payment and then does not take the customer. The partner still owns the booking, the refund, the complaint and the loyalty relationship. That is the opposite of the marketplace bargain, and it is the same design choice that separates UCP from the platform-mediated model: Google is positioning itself as the surface, not the seller. For anyone who watched Google ship a cart in June, this is the same architecture extended from retail into travel, which tells you the pattern is intentional rather than category-specific.

The uncomfortable half is what happens to everyone outside the list of ten. When the answer completes a booking with a partner, the non-partner is not a slightly worse option in the same list. It is in a different interaction entirely, one that requires the user to leave. That gap is not closed by better product data, and it is the clearest case yet of being outside a partner programme carrying a mechanical rather than a ranking cost. Watch whether the partner list stays curated or opens up, because that decision sets the shape of every category Google extends this to next.

Note also that this landed on travel, where inventory, cancellation terms and price are already highly structured and machine-readable, and where a small number of aggregators hold most of the supply. That is not a coincidence. The categories that get an in-answer checkout first will be the ones whose data was already clean enough to transact on.

Europe’s paid slot went live on 24 August, on schedule

OpenAI announced on 18 August that ChatGPT advertising would reach 31 European countries the following week, and it went live on Monday 24 August. The 31 markets are the 27 EU member states plus Iceland, Liechtenstein, Norway and Switzerland. Ads appear only on the Free and Go plans. Personalisation is not active in the EEA and Switzerland at launch: selection runs on the context of the current conversation plus general signals such as language and broad location, without past chats, memory or interaction history. Buying runs through OpenAI’s ads solutions team, agency partners and technology partners, with beta self-serve Ads Manager access following on 31 August, and OpenAI shipped a pixel and a Conversions API alongside it.

The line worth extracting is not the go-live date. It is that OpenAI states, in its own materials, that advertising does not influence the answers ChatGPT generates.

Why it matters for merchants: That is the platform confirming what independent measurement found a fortnight ago, when SE Ranking found only 3.63 percent of advertisers appearing as a source in the answer above their own ad. Two separate inventories, now attested by both the study and the vendor. If your internal debate is still “should we buy ChatGPT ads or fix the feed”, the answer is that those buy different things and neither substitutes for the other.

The other consequence is that with personalisation off there is no behavioural profile doing the targeting work, so selection leans on how well your offer matches the topic of the live conversation. Every merchant in the EEA starts from the same absence of profile data, which makes this the flattest competitive footing the channel will ever have. If you want to see what a model currently reads of your products, the Shopify Feed Previewer shows the input side of that match.

Practical caveat, now with an expiry date: being shown ads and being able to buy them were separate rollouts for exactly one week. On 31 August OpenAI opened beta self-serve Ads Manager access to eligible advertisers in all 31 European markets, alongside India and 11 Middle East and North Africa markets. A budget can now be deployed without an agency or an OpenAI partner, subject to entity verification.

Your discounts do not exist if the agent cannot read them

PYMNTS argued on 24 August that loyalty benefits and promotional offers are becoming irrelevant unless an AI system can detect and apply them at the moment of recommendation. The supporting evidence is circumstantial but consistent. Target’s non-merchandise revenue, which includes its loyalty and membership economics, rose more than 20 percent in Q2 alongside the AI traffic figure. Synchrony announced an enterprise collaboration with OpenAI on 17 August to put promotional financing offers from its marketplace inside a ChatGPT conversation. Sephora’s ChatGPT app, which launched back in March, connects Beauty Insider status so that member benefits such as samples and free shipping surface in the advice itself.

Avery Miller, Visa’s VP of global loyalty for value-added services, and Kipp Johnson, Braze’s senior director of AI solutions consulting, told PYMNTS that personalisation has become table stakes in loyalty, and that the differentiator now is whether that value can be surfaced by an AI system in real time.

Why it matters for merchants: This is the same failure mode as the returns-policy finding from last period, applied to the other half of your commercial terms.

Most promotional value in ecommerce is expressed in places a model cannot parse: a banner image, a code applied at the cart, a members-only price revealed after login, a “10% off your first order” popup. An agent comparing three products on your behalf sees your list price and someone else’s list price. If your competitor’s discount is in a structured field and yours is behind a login wall or baked into a JPEG, you lose that comparison while being the cheaper option. That is a genuinely infuriating way to lose a sale and it is happening now.

The fix is unglamorous and mostly free. Put price reductions in Offer markup with priceValidUntil rather than only in creative. Expose member pricing where you can, or at minimum publish the member price as a visible tier rather than a post-login surprise. Ensure your product feed carries the sale price field and that it is actually populated, because feed sale price is the single field most commonly left blank in the catalogues we look at. And if you run financing, note what Synchrony is doing: the terms are becoming a discoverable attribute of the product, not a checkout-page disclosure.

There is a strategic point underneath the tactical one. Loyalty programmes were built on the assumption that the customer knows they are a member and factors it in. An agent shopping on their behalf does not know unless you tell it, in a field. Every locked benefit is now a benefit that does not enter the comparison.

The industry admits it cannot measure any of this, and sets a date

On 24 August, the IAB confirmed it is drafting a framework for measuring and crediting AI’s role in conversions, with a target release of 12 November 2026. Caroline Giegerich, the IAB’s VP of AI, is leading the drafting with a working group spanning tech companies, publishers, agencies, measurement vendors and brands. The framework splits AI’s influence into two layers, when an AI presents something to a user and when it assists the decision, and tries to attach credit to each. The problem it addresses is blunt: UTM parameters and referral data do not survive a journey where an agent reads several product pages, compares options and hands the user a conclusion. Separately, IAB Tech Lab continues work on Agentic Advertising Management Protocols covering how autonomous systems discover, plan and buy media.

Why it matters for merchants: Note the date and count backwards. A framework published on 12 November is a framework that arrives after Black Friday planning is locked and roughly at the start of peak trading. Nobody is measuring this season’s AI-influenced revenue with an industry standard. You will be measuring it with whatever you build yourselves, or not at all.

That has a budgeting consequence right now. Channels that cannot be measured lose budget arguments to channels that can, regardless of which one actually performs. If you go into Q4 planning with paid search reporting a clean ROAS and AI-influenced revenue reporting nothing, you will underfund the second one and the numbers will appear to justify it. The defence is to instrument something imperfect before the season rather than waiting for something rigorous after it.

The cheap version, which most merchants can do in a week:

Pull your server logs and count declared agent user agents separately, as we set out last period.
Filter your existing analytics by referrer for the AI surfaces, chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com, and keep it as a monthly series. That is your own version of the Target number.
Add a coarse “how did you hear about us” option naming AI assistants at checkout. Unfashionable, and surprisingly effective when nothing else works.
Record a monthly baseline of whether your top products appear in answers to your top commercial queries.

None of that is attribution. All of it is better than a zero that you know is wrong.

Cloudflare separates “do not train on me” from “do not find me”

On 21 August, Cloudflare launched Bot Preference Sync, which generates a site’s robots.txt directly from the AI bot policy already set in the dashboard, on every plan from Free to Enterprise. The generated block is prepended to any existing robots.txt between BEGIN and END markers, and the user-agent list refreshes periodically from Cloudflare’s own bot database. The substantive change is in what “Disallow” now means: it writes a no-training preference in a form that still lets cooperating mixed-purpose crawlers reach the content for search indexing. Operators running crawlers that do both search and training must meet four disclosure conditions to qualify, including URL-level reporting on which pages were used for training, or they stay blocked outright. New domains that declare advertising as a monetisation model get Training set to Disallow by default at onboarding; other new domains get no blocks.

Why it matters for merchants: Last period we warned that a single blanket bot policy almost certainly blocks live customers, feeds scrapers, or both. This is the first mainstream control that keeps that distinction in sync with what your edge is actually enforcing, rather than leaving robots.txt to drift away from your firewall rules.

Two things to do. First, if you are on Cloudflare, go and look at what your AI bot policy currently says, because this feature makes that setting authoritative over your robots.txt and a lot of merchants have never opened that panel.

Second, check the default. A new domain that declares ad monetisation now starts with training disallowed, which is a reasonable default for a publisher and a potentially expensive one for a retailer that also runs a content operation, because the four disclosure conditions decide whether a mixed-purpose crawler is treated as a search crawler or shut out entirely.

The broader point is that the industry is converging on the distinction we keep coming back to. A training scraper, an index crawler and an on-demand fetcher answering a live customer question are three different visitors, and treating the bot gate as a single on-off switch keeps being the expensive mistake. For context on scale, Cloudflare Radar’s June 2026 reading put automated requests at 57.5 percent of HTML traffic against 42.5 percent human, which CEO Matthew Prince flagged on 3 June as the first time machines had passed people, arriving well ahead of his own end-of-2027 prediction. That is a June number, not a new one, but it is the backdrop against which a blanket block is a bigger decision than it looks.

Two shorter items worth knowing

Amazon asked the full Ninth Circuit to rehear the Perplexity ruling. On 18 August, Amazon petitioned for rehearing en banc of the 4 August panel decision that vacated its injunction against Comet, arguing the ruling “degraded website owners’ ability to set the terms on which powerful, fast-evolving, and potentially destructive AI agents may enter their secure systems”. Its central counter-argument is that control, not initiation, is what matters: “Users may or may not choose to access Amazon, but the assistant will access Amazon regardless, acting under Perplexity’s control.” Nothing has changed operationally, the agent is still permitted, but the holding that the agent is the user is not settled law and should not be planned around as if it were.

The agent moved to the buying side of advertising. Announced on 21 August and documented on 24 August, X’s Ads MCP server exposes 23 tools to AI agents, ten of which can write to live ad accounts, with campaigns arriving paused and requiring a separate activation call. That follows Meta enabling write access in April and extending it in July, and TikTok announcing a full-lifecycle server in May, against Google’s read-only implementation. This is adjacent to commerce rather than inside it, but it is the same protocol layer arriving on the media-buying side. The practical question for a merchant is one to put to your agency this month: are any agents currently holding write credentials to your ad accounts, and who reviews what they change before activation. Campaigns arriving paused is a safeguard that only works if somebody is looking at the queue.

The European angle: live channel, no checkout, no timeline

Europe’s position changed materially in these two weeks, and in one direction only.

The paid slot is real. On 24 August, ads went live in all 27 EU member states plus Iceland, Liechtenstein, Norway and Switzerland. There is no longer any version of an EU AI channel plan that treats paid placement as a future problem.

The thing to watch next is the personalisation opt-in. OpenAI has built the European launch as contextual by default with personalisation available only on explicit consent, which is a GDPR-shaped architecture rather than a permanent product decision. When that opt-in prompt starts converting at scale, the flat footing described above erodes and behavioural signal begins doing what it already does in the US. Nobody has published opt-in rates and it would be a guess to predict them, so the useful posture is to treat the current window as finite and to note that the catalogue work holds its value either way, whereas an advantage built on everyone lacking profile data does not.

The completion layer is not real here. Google’s hotel checkout is US-only and English-only with no announced European timeline, and the announcement said nothing about Digital Markets Act considerations that plainly apply to a Google surface taking payment for third-party inventory. Do not read the silence as a delay you can rely on, and do not read it as an imminent launch either. Nobody outside Google knows. Two things would tell you the position has changed: the partner list opening beyond the initial ten, or any DMA compliance statement attached to the checkout flow. Until one of those appears, European merchants have an unearned interval in which the same product and price data that feeds the answer today will feed the checkout when it arrives.

The measurement problem is worse in Europe, not better. The IAB framework is a US trade body’s output arriving on 12 November, and the pixel-and-conversions-API instrumentation OpenAI shipped alongside the European ad launch sits inside a consent regime that limits what you can join up. If your Q4 plan depends on proving AI-influenced revenue to a finance team, build that proof now with server logs and self-reported attribution, because nothing standardised is arriving before the season.

What moved this period

Development	What happened	Why a merchant cares
Q2 earnings numbers	Walmart 40% higher order value, Lowe’s 3x conversion	Real figures, but user vs non-user, not causal proof
Target AI-platform traffic	Growing 3.5x faster than the industry rate	The one number measuring external models finding you
Google AI Mode hotel checkout	Completes in Google Pay, partner stays merchant of record	The answer takes payment without taking the customer
ChatGPT ads live in Europe	31 markets on 24 August, unpersonalised at launch	Paid slot live EU-wide, semantic match does targeting
Loyalty and promo visibility	Offers invisible to agents unless machine-readable	You can lose a comparison while being the cheaper option
IAB attribution framework	Confirmed 24 August, ships 12 November	No standard before peak season, so instrument yourself
Cloudflare Bot Preference Sync	robots.txt generated from dashboard AI bot policy	Block training without blocking the answer you want in
Amazon v. Perplexity	Rehearing en banc petitioned 18 August	The agent-is-the-user holding is not settled law yet

What merchants should do this period

1. Put your promotional and loyalty terms in structured fields before Q4 creative is locked. An agent comparing your product to a competitor’s reads list price unless the discount is machine-readable. Populate the sale price field in your product feed, use Offer markup with priceValidUntil for time-limited reductions, and stop expressing your best pricing exclusively in banner creative or behind a login. If you have a members-only price, decide now whether the loyalty gate is worth being excluded from the comparison entirely.

2. Instrument AI-influenced revenue this month, imperfectly, rather than waiting for 12 November. The IAB framework arrives after peak season. Three things you can do in a week: count declared agent user agents in your server logs as a weekly series, add AI assistants as a named option in a post-purchase “how did you find us” question, and record a monthly baseline of whether your top ten products appear in answers to your top ten commercial queries. Unmeasured channels lose budget arguments, and that is the actual risk here.

3. If you are on Cloudflare, open the AI bot policy panel and check what it now says. Bot Preference Sync makes that dashboard setting authoritative over your robots.txt, and the Disallow option no longer has to cost you search and answer visibility. Verify which of the three visitor types you are currently blocking. If your domain declares advertising monetisation, training is disallowed by default, which may or may not be what you want.

4. Treat the earnings numbers as evidence of a channel, not as a forecast for your store, and build your own version of Target’s. Walmart’s 40 percent and Lowe’s 3x are user versus non-user comparisons with obvious selection effects, and both measure a retailer’s own assistant. Target’s is the one worth benchmarking, because inbound traffic from external AI platforms reflects whether other people’s models can find you. To produce it: segment sessions by referrer for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com, add the declared agent user-agent counts from your logs, and track both monthly. It will be an undercount, because agents relaying through the user’s own browser are indistinguishable from the customer. Track the trend rather than the absolute number.

5. European merchants: split the plan into a live half and a pending half. The paid slot is live in your market as of 24 August, and since 31 August you can buy it directly through Ads Manager rather than only through an agency, so budget for it. The completion layer is not live and has no announced date, so the correct posture there is readiness rather than spend: clean prices, stock, variants, cancellation and returns terms in fields, because that is the same data an in-answer checkout will require whenever it lands.