The Assistant Sending You Buyers Is No Longer Just ChatGPT

Two things landed on 2 September 2026 that belong in the same conversation, though almost nobody covered them together. Anthropic put out a package of prebuilt commerce agents. Amazon switched on a feature that lets a shopper ask its assistant whether a message claiming to be from Amazon is real. One is about machines that buy. The other is about machines that vouch. Both push the assistant one step further into the part of the shopping journey a merchant used to own outright.

Underneath them sits a quieter number that matters more to a small shop than either announcement. The traffic an assistant sends you is growing fast, it buys more when it arrives, and the name on the referrer is not the one it was six months ago. The August 2026 Similarweb reading shifted the ranking again, which is part of why the assistants moved the fight to discovery rather than checkout.

Which AI assistants actually send retail traffic?

ChatGPT still sends the most, and it is losing share to everyone else. Digital Commerce 360 compared its Top 1000 retailers across the first and second quarters of 2026 and found the leaderboard reshuffling underneath a still growing total. The figures come from its quarterly breakdown of which AI sources refer traffic to online retailers, and they are counts of retailers, not shares of visits.

Read that table honestly. ChatGPT is the top AI referrer for roughly seven hundred of the largest thousand online retailers in the United States, and the challengers are counted in tens. A shop that optimises only for the leader is still optimising for the overwhelming majority of this traffic. But the trend line is the story: the number of retailers whose biggest AI referrer is something other than ChatGPT roughly tripled in one quarter. Scot Wingo, chief executive of ReFiBuy, told Digital Commerce 360 that if Claude's trajectory held for eight quarters it would be competing with Gemini and ChatGPT at the top. That is a conditional forecast, not a measurement, and it should be read as one.

The practical consequence is dull and useful. If you have been checking how your shop looks inside one assistant, you now need to check three. They do not read your pages the same way, they do not cite the same pages, and the fixes that make you legible to one tend to make you legible to all of them. We went through what actually moves those citations in a piece on what changes your citation rate inside AI answers, and nothing in this quarter's data changes that method.

Why does AI referred traffic convert better than the rest?

Because the assistant has already done the filtering that a search result leaves to the shopper. By the time somebody clicks through from a conversation, they have described what they want, been shown options, and picked one. Adobe's measurement, reported in Digital Commerce 360's write up of the July 2026 Adobe Analytics figures, puts numbers on how differently those visitors behave.

Adobe says the underlying sample covers more than a trillion visits to United States retail sites, and that July was the eleventh consecutive month in which AI referred traffic outperformed. It also gives the growth figure since October 2024 as 1,219%, which sounds enormous until you remember what the base was. A number that starts near zero produces spectacular percentages for a long time.

What did Anthropic actually ship on 2 September?

A blueprint, in its own words, rather than a product. Anthropic released reference implementations of two agents that a retailer's engineers can run and adapt. Digital Commerce 360's report on the Claude commerce features describes the pair as a shopper agent facing customers and a merchant agent facing the back office.

The shopper agent searches a catalogue, makes recommendations with context, remembers a customer's preferences, shows products visually inside the conversation, and moves a cart towards checkout. The merchant agent answers questions about sales and inventory, flags problems with a suggested action attached, and drafts campaign and pricing material. Accenture, Mastercard and Visa are named as partners taking it to their client and merchant networks. Shopify published a repository letting retailers on its platform connect an agent to catalogue data, and Priceline rebuilt its Penny assistant on Anthropic's models.

Anthropic's own performance claim is that retailers running shopping agents built on Claude see carts as much as 35% larger and shoppers 60% more likely to complete a purchase. PYMNTS, in its analysis of what the blueprint leaves out, correctly notes that these are vendor reported figures with no independent benchmark behind them. Treat them the way you would treat any conversion number in a launch post.

Does the blueprint touch my checkout?

No, and that is the most interesting decision in the whole release. There is no payment protocol in it, no catalogue backend, and no advertising layer. Checkout, payments and transaction processing stay inside the retailer's existing systems. PYMNTS calls it the shopping brain without the wallet, which is a fair description of the architecture.

The reason is not modesty. It is liability and consumer nerves. PYMNTS cites research finding that while 45% of consumers, and 54% of Gen Z, would accept an agent completing a purchase autonomously, 95% report concerns about agentic commerce generally. Keeping the card details outside the agent means the payment networks keep the trust relationship and the model vendor keeps out of the dispute chain. If you sell things, that is the sentence to remember: the reasoning moved, the money did not.

What that means for you legally has not changed either. When an agent places an order on your site, the questions about who agreed to your terms and who is on the hook for a mistake are the same ones we worked through in the article on what agentic commerce changes at your checkout. A blueprint that avoids payments does not answer them. It just declines to make them worse.

What does the code actually do, as opposed to the announcement?

It runs a simulation. The public commerce agents repository ships four runnable verticals covering retail, travel, telecom and entertainment, three runtime options, and a shared library handling config, memory, skills, grounding and events. It is Apache 2.0 licensed. Running the demo needs Python 3.11 or newer, Node 22 and an API key.

Two details in that repository are worth more to a merchant than the entire press cycle around it. First, every write passes a merchant approval gate: changes are staged rather than applied, so the agent proposes and a human commits. Second, the demo explicitly performs no real orders, no card charges and no live listing changes. Everything is simulated until somebody implements the interface that connects it to a real catalogue, real inventory and real pricing.

That is the honest shape of the thing. It is scaffolding for an engineering team, not a switch a small merchant flips. If you run your shop alone, the correct reading of this announcement is not that you should deploy a commerce agent next week. It is that the large retailers you compete with now have a cheaper path to shipping one, and their storefronts will get better at answering an assistant's questions. Yours needs to be answerable too, which is a much smaller job.

Is the merchant agent the half a solo shop should care about?

Probably, and not as software. Strip the merchant agent down to what it is described as doing and you get a short list of standing questions that somebody should be asking your data every week: how did sales move, what is running low, what is priced wrong, what should the next campaign say. Most merchants running alone answer none of those on a schedule. They answer them when something goes wrong.

You do not need an agent framework to fix that. You need the questions written down and a time in the week to ask them. What the reference implementation adds is that it asks them without being prompted and attaches a suggested action, and then stages that action behind a human approval rather than applying it. The staging is the part worth copying even if you never run a line of the code. An assistant that proposes a price change and waits is useful. One that applies it is a liability with a good user interface.

The same logic applies to the listing work. Handing a model your catalogue and letting it rewrite titles in bulk is how a shop ends up with three hundred confidently wrong product pages and no record of what changed. Propose, review, then commit. If you want the longer version of that argument, we set out which catalogue fields should never be handed to a model in the piece on moving a catalogue without letting AI touch the wrong fields.

Does this change what a scam message looks like to my customers?

It changes who your customer asks. Amazon added scam detection to Alexa for Shopping, letting a shopper paste a suspicious message and get an answer on whether it genuinely came from Amazon. TechCrunch's report on the scam detection feature says it compares sender information, content, timing and message metadata against the billions of messages Amazon has sent, and that around 360,000 customers a year already contact Amazon customer service to ask exactly this question.

That 360,000 figure is the one to sit with. It is a measure of how routinely people cannot tell a real transactional email from a fake one, at a company with more brand recognition than any shop reading this will ever have. Your order confirmations are being judged by people in that same state of doubt, and they have no verification service to ask about you.

The defensive work is old and boring and it still decides the outcome. Send from a domain you control, with authentication records that pass, in a consistent template, from a consistent address, with no link shorteners. We wrote the mechanics up in the piece on the rules that decide whether your email reaches the inbox. A message that authenticates cleanly is also a message an assistant can more easily confirm on your behalf, which is a benefit nobody was designing for two years ago.

What should a small shop actually do this month?

Very little, done properly. The temptation after a week like this is to conclude that you need an agent of your own. You almost certainly do not. What you need is for the agents other people already use to be able to read you accurately, and that is a checklist rather than a project.

Put price and availability where a machine can find them. Structured product data on the page, matching what a human sees. An assistant that cannot confirm a price will hedge, and a hedged recommendation loses to a confident one. If your stock number on the page disagrees with your stock number at checkout, you will convert an assistant's referral into a support ticket.

Write the returns policy in sentences. Not a table of exceptions, not a PDF. The question an assistant gets asked before a purchase is almost always about returns, delivery time or sizing, and it answers from whatever text it can read. If your policy is ambiguous to a careful human, it is ambiguous to the model, and an ambiguous answer is where invented policies come from. We took that failure apart in the article on assistants inventing a policy you never wrote.

Give order status a page that answers without a login. A tracking lookup keyed on an order number and an email, rate limited, is enough. This is the single highest volume question in most support inboxes and the easiest one to hand off.

Keep one canonical URL per product. Variants that each generate a distinct URL with near identical content give an assistant several candidate answers and no reason to prefer one. Pick one, and make the others point at it.

Look at your own referrer report before you believe any of this. Filter your analytics for the assistant hostnames and see what actually arrives. If the answer is eleven sessions last month, you have learned something worth more than any industry average: this channel is not yet your problem, and you can spend the weekend on the one that is.

None of that requires new software. A storefront that already renders clean product data, has readable policy pages and exposes an order lookup is legible to an assistant by construction, which is one of the reasons we build a shop from a description rather than from a template: the generated pages carry the structured fields whether or not the merchant knows what they are for.

What this week does not mean

It does not mean shoppers have moved to assistants. The Adobe figures describe a small, fast growing and unusually valuable slice, not a migration. It does not mean agents are buying autonomously at scale, because the release that got the most attention deliberately declined to touch payments. It does not mean the assistant leaderboard is settled, since the direction of travel in one quarter was away from the leader rather than towards it.

What it does mean is that the machine reading your product page is now a customer of a kind, and it has preferences. It prefers pages where the price is unambiguous. It prefers policies it can quote. It prefers shops whose stock data does not contradict itself. Those preferences happen to match what a hurried human wants at eleven at night on a phone, which is the least surprising outcome in all of this and the reason the work is worth doing regardless of how the quarter's referral counts move next.

The reasoning layer moved into the shopping journey this quarter. The money, the catalogue and the liability stayed exactly where they were.MaShop, reading the commerce agents repository, 4 September 2026