For AI Shopping Assistants, the Fight Is Discovery

Two numbers landed within a day of each other this week and they tell one story. Similarweb's August measurement, reported on 7 September 2026, put ChatGPT back at 55.5 percent of chatbot website traffic after three months of losing ground, while Gemini slipped from 27.8 to 25.6 percent and Claude sat at 9.3 percent. A year earlier ChatGPT held 73.3 percent. That is a market reshuffling fast enough that any shop betting on one assistant is betting badly.

The second number is quieter and it matters more. The attempt to move the actual purchase inside the chat window has been walked back, and the competition has settled on the step before it. Whoever decides which three shops get named when someone asks for a birthday gift under forty pounds is the one holding the valuable ground. That is a discovery problem, and it is the one you can do something about this month.

What changed between the launch and now?

The launch was ambitious and the retreat was quiet. In September 2025 Stripe and OpenAI released the Agentic Commerce Protocol, an open standard letting a shopper complete a purchase without leaving the chat, starting with United States Etsy sellers and a wave of hosted storefronts behind them.

The design was careful about merchant control. Orders arrived through the protocol's interface, and the shop kept the right to accept or decline each one, charge the payment method, calculate and remit sales tax, and handle fulfilment and returns exactly as before. A shared payment token carried the authorisation without handing over the card. On paper a merchant gave up nothing except the page the sale happened on.

Google shipped a parallel answer a fortnight earlier. Its Agent Payments Protocol arrived on 16 September 2025 with more than sixty organisations attached, among them Mastercard, Adyen, American Express, PayPal, Worldpay and Etsy. It answered a narrower question than the OpenAI standard: not how an agent buys, but how a merchant proves it was allowed to. The protocol signs three separate mandates, one recording what the shopper asked for, one freezing the exact basket and price the shopper approved, and one binding the payment method to that frozen basket. If a charge is later disputed, the signed record is the evidence.

Both standards were built for a future where the sale closes inside the assistant. That future got postponed. The in chat purchase step was withdrawn in March 2026 and shoppers were sent back to the merchant's own checkout, with the assistants concentrating on product discovery instead. The protocols survive, the plumbing survives, and the thing they were built to carry has moved.

Why did the purchase step go back to your own site?

Because shoppers did not want it and merchants did not want it either. Only about one third of consumers surveyed by Forrester said they would complete a payment inside an answer engine, citing privacy concerns, which leaves two thirds who would rather finish on a page they recognise.

The merchant side of that reluctance is more concrete. A sale completed in someone else's window costs you the customer journey data, the loyalty enrolment, the post purchase email and the returns conversation. Retail Dive's reporting on the risk describes the destination bluntly: retailers becoming something closer to a fulfilment company while the platform keeps discovery, decision and relationship. Eighty one percent of retail executives in that piece expect generative AI to weaken brand loyalty by 2027.

For a shop with three people and no loyalty programme, disintermediation sounds like a word for someone else's problem. It is not. The email address is how you sell the second thing. If a channel converts well and never gives you an address, you have rented a customer rather than earned one. We covered the mechanics of that handoff in more depth in our piece on what agentic commerce changes at checkout, including who has agreed to what when a machine places the order.

How many shoppers actually use these tools?

Enough to matter, and fewer than the headlines imply. A CapitalOne Research fact sheet cited by Practical Ecommerce said almost 60 percent of consumers have used AI to shop at some point, while NielsenIQ put the figure at 42 percent of American consumers in the past month, which is the more useful measure because it counts habit rather than curiosity.

Set that against volume. Research from OpenAI's economics team with Harvard found roughly 2 percent of ChatGPT queries are shopping related, which works out at something near 50 million shopping questions a day. Two percent sounds like a rounding error until you multiply it by the base.

Growth rates are where the reporting gets loudest and least useful. Retail Dive's figure of 758 percent year on year growth in AI driven United States ecommerce traffic is real and it is also a statement about a small starting number. Our own reading of the platform disclosures found agentic shopping still under one percent of Etsy's traffic, and both things are true at once. A channel can multiply eightfold and remain the smallest line on your dashboard. What it cannot do is stay that small if 42 percent of shoppers are already in the habit.

One caveat travels with every row. Similarweb measures websites, not apps. Phone usage through the native apps is invisible to it, and phone usage is where a lot of casual shopping questions get asked. Treat the table as a ranking, not a census.

The shelf analogy is doing real work here

Armando Roggio's framing in Practical Ecommerce is the most useful sentence written about this all year: the assistant conversation is shelf space. Brands have always competed to get on the shelf, to sit at eye level, and to be picked once they are there. Anthony Ferry, chief executive of the commerce technology firm Wayvia, told that publication the job of a brand has not changed, it has only added a new party to persuade, and this party is a language model rather than a category manager.

Ferry's other observation lands harder for a one person business. Advertising budgets used to split across television, radio and print. Then the internet arrived and took a slice. Now, in his count, there are thirty channels asking for a share of the same money. A solo merchant cannot work thirty channels. The question is not how to be everywhere, it is which two or three surfaces return anything, and everything else gets ignored on purpose.

It now includes educating LLMs to recommend the brand's products over competitors'.Anthony Ferry, CEO of Wayvia, in Practical Ecommerce

Did the purchase protocols die with the checkout?

No, and that distinction matters if anyone tries to sell you an integration. The standards are plumbing, and plumbing outlives the first thing it was built to carry. What was withdrawn was the consumer facing step where money changed hands inside the chat, not the machinery for passing a structured order to a merchant and getting an acceptance back.

Large retailers kept building on it, generally in the direction of a branded surface inside the assistant rather than a generic buy button. That model gives the retailer its own account linking, its own loyalty rules and its own payment path, which is close to what a merchant wanted in the first place. It also takes engineering nobody running a shop alone possesses, which is the honest reason it is not your problem this quarter.

The mandate structure in Google's protocol is the piece most likely to reach you eventually, because it is about liability rather than convenience. When an agent places an order and the shopper later says they never approved it, somebody has to hold the proof. A signed record of the exact basket at the moment of approval is a better answer than a chat transcript, and card networks will push in that direction long before small merchants ask for it.

What should a small shop do about it this month?

Start by finding out what the assistants already say about you, because most merchants have never checked. Open each of the three that matter, ask the question a customer would ask, and write down whether you appear, what price it quotes, and whether the description matches what you sell.

That exercise usually produces one of three findings. You are absent, in which case the problem is that your pages are thin or blocked. You are present and wrong, which is worse than absent because a confidently quoted stale price loses a sale and a bit of trust. Or you are present and accurate, in which case the job is to keep it that way through your next catalogue change.

The fixes that follow are unglamorous. Put the price, the stock status, the delivery window and the returns window in text on the product page, not inside an image and not behind a tab that loads on click. A model reading your page has none of a human's tolerance for finding things. We went through what actually moves the needle in what moves your citation rate in AI answers, and the short version is that specificity beats volume every time.

Then decide what you are doing about crawlers, deliberately rather than by default. Blocking every AI bot keeps you out of the answers entirely. Allowing everything means your catalogue trains models that may recommend a competitor. Most small shops want the read that produces citations and are indifferent to the rest, which is a position you can express in a robots file rather than a philosophy.

The measurement problem nobody solves for you

Traffic arriving from an assistant frequently carries no referrer, or carries one that your analytics files under direct. So the channel that is growing fastest is also the one your dashboard is worst at seeing, and the temptation is to conclude it does not exist.

Two cheap habits close most of the gap. Add a question to your checkout asking how the customer found you, with the assistants listed by name, and accept that a third of people will skip it. Then watch your branded search volume and your direct traffic together, because an assistant that recommends you without linking you produces a person who types your name into a search bar an hour later. Neither method is precise. Both beat a channel report that shows zero.

There is a third signal worth watching if you sell anything with a model number. Assistants tend to quote exact specifications when they have them, so a spike in long, oddly precise search queries hitting your site search box often means a machine repeated your own words back to a shopper who then came looking. That pattern shows up in site search logs before it shows up anywhere else, and reading a week of them costs nothing.

Our breakdown of which assistants are actually sending buyers goes through what shows up in server logs versus what shows up in analytics, and the two disagree more than most merchants expect.

Is any of this worth a solo merchant's time yet?

Yes, but only the cheap half of it. Auditing what three assistants say about your shop takes under an hour and occasionally catches a wrong price that was costing you orders. Rewriting your product pages so the facts sit in readable text is work you owed your human customers anyway, and it happens to be exactly what a model needs.

The expensive half can wait. Building against a purchase protocol, paying for visibility monitoring, or restructuring a catalogue around what an assistant might ask are all bets on a channel that is under one percent of most shops' traffic. Julie Bornstein, who founded the shopping platform Daydream, put the timing plainly in Modern Retail: perfecting these experiences will take longer than anyone thinks, and this year is a step along the way rather than the finish line. Andy Jassy, running the largest store on the internet, said in the same reporting that most AI shopping agents still fail to give a satisfactory customer experience because they lack personalisation and quote inaccurate prices.

When two people with that much to gain from the channel are describing it as unfinished, a merchant with limited hours can reasonably do the free work now and hold the budget.

What this means if you are building the shop itself

Everything above assumes you can edit your own product pages, control your own robots file, and see your own server logs. On a hosted platform where the template decides what the markup looks like, some of these fixes are simply not available to you, and that constraint is worth knowing before you pick where to build. It is one of the reasons we built an AI store builder that hands you the code, because a page you own is a page you can make legible to whatever reads it next.

The pattern underneath this news is older than the technology. A new intermediary appears, promises to own the transaction, discovers that the transaction is the hardest part to take, and settles for owning the introduction instead. Search did exactly this. Marketplaces did it. The assistants are doing it now, faster, and with better manners about giving the sale back.

Your defence is unchanged and slightly boring. Be findable, be accurate, and get the email address. The channel that introduces you will keep changing. The thing you are protecting will not.