Something changed in how ChatGPT researches an answer, it changed on a single day, and the operator implication is this: for a meaningful share of the background searches the model runs before it recommends a product, it now decides which website is authoritative before it decides anything about your product.
That is a source-selection step sitting upstream of every asset you control. If the model scopes a search to amazon.com, your detail page is in the running. If it scopes to a review site, a retailer, or a manufacturer's own domain, your listing was never in the conversation and the thing deciding your product's fate is a page you have no edit access to.
Nothing about your September P&L moves because of this. I want to say that in the first hundred and fifty words rather than pretend otherwise. But this is the clearest look anyone has gotten at how the AI discovery layer actually assembles a recommendation, and it arrived without an announcement describing it.
What happened
Promptwatch, which monitors AI search interfaces at scale, published data showing that ChatGPT's fanout queries using the site: operator jumped from 0.37% to 16.8% of all fanout queries — roughly a 46x increase in share inside a single day, on August 8, 2026. The share had sat between 0.3% and 0.5% for weeks beforehand, dipped to 0.15% on August 3–5, then jumped and has held around 16–17% since. In the same window, the average number of fanout queries per response went from about 1.08 to about 1.83.
Simon Willison wrote it up on August 20. The timing lines up with OpenAI's vague August 6 note that it was updating GPT-5.6 Sol in Chat "to be more reliable with facts and provide more focused answers." That sentence does not describe this behaviour at all, which is the second-most interesting thing about the story.
Two caveats, held honestly, because I am about to build an argument on somebody else's panel. Promptwatch sells in the GEO space, so it has a commercial interest in AI visibility being a subject people pay attention to. And as Willison notes, these figures only reflect the prompts for which their automated tracking is enabled. Treat the exact percentages as directional. The shape — a step change on a specific date, sustained since — is not really in dispute, and you can go read the underlying data yourself.
A fanout query, if you have not run into the term: it is the background search the model fires to gather material before writing you an answer. You never see them. They are the retrieval layer under the conversation.
Why most brand owners will read this wrong
The dumb take is that this is search engine news, and search engine news belongs to whoever owns the website. Amazon sellers filed the whole GEO conversation under "not my surface" eighteen months ago and have not reopened it, largely because the advice on offer has been terrible.
The other dumb take is the opposite reflex — that ChatGPT is about to eat marketplace discovery and you need an agentic readiness project by October. Adoption is still a rounding error against Amazon search volume and I am not going to manufacture urgency about it.
The real signal is narrower and more useful than either. The model used to search the open web and see what came back. It now frequently decides where to look first. That is not a ranking change, it is a change in the composition of the pool your product is being compared inside. And the pool is what determines whether "which stainless water bottle should I buy" resolves to a set of Amazon listings or to somebody's roundup post from 2024.
The corollary landed ten days later. On August 18, Promptwatch reported that ChatGPT appeared to have sharply reduced the likelihood of Reddit being used in those searches. Willison could not confirm whether the system prompt was changed to discourage it, and Search Engine Journal ran a piece arguing the citation drop is not fully explained. So: something moved in which sources get consulted, nobody outside OpenAI can tell you exactly what, and it moved on a Tuesday.
I have written some version of "the thing moved underneath you and nobody sent a memo" four or five times this year. What is new here is that the artifact does not belong to the company that made the change. There is no changelog entry, no pricing note, no retirement date. The only record that this happened is a tracking panel run by a third party who happened to be watching. You can now be affected by a retrieval decision that leaves no trace in any system you own.
What actually changes for someone running $200K a month
The comparison set gets assembled somewhere you cannot see, and it is now assembled deliberately. Before, being in the answer was partly a function of what surfaced from a general web search. Now, for one in six of those background searches, inclusion depends on a domain being chosen. There is no version of your listing copy that influences that step. It happens before your page is read.
Fanouts per response nearly doubled. More searches per answer means more chances to be pulled in and more chances for a competitor to be. A wider consideration set is being assembled per question than was being assembled three weeks ago. That cuts both directions and it is the part I would actually watch over the next quarter.
Domain-scoped retrieval rewards being the best page on a domain, not the best page on the web. If the model scopes to amazon.com, your competition inside that scope is roughly the same set you fight in the search grid — twenty results, most of whom you can name. That is a fight you already know how to have, and every asset that wins it is one you already own. If the model scopes to a third-party editorial site, you are competing on whether you are mentioned, which is a PR problem, not a listing problem, and it is the first time a share of your discovery is decided by content you cannot edit.
None of it shows up in your reporting. An AI-referred session that lands on your detail page is a session. It frequently has no search term behind it, because no search happened on Amazon. It never attaches to a keyword, never clicks an ad, and blends into an average with no line item. Unmeasured does not mean small. It means unassigned.
And the assets that transmit to this reader are the boring ones. A model reading a page is reading text: title, bullets, structured attributes, A+ copy, alt text, price, review summary. A page whose entire argument lives inside a 2000px infographic is making that argument to a reader who has to work considerably harder to receive it. The exact same work makes you legible to Amazon's own AI layer. Two readers, one job. That does not happen often, and it is why I keep pushing the unglamorous version over anything with "AI visibility" printed on the invoice.
What I'd do this week
1. Read your top revenue detail page with the pictures off. Reader mode, or paste it into a document and strip the images. Five minutes. That text is what every non-visual reader gets — Amazon's layer, a third-party agent, a screen reader, Google. Most brands have never once looked at it, and what they find is empty attribute fields and an A+ section whose whole case was carried by a graphic.
2. Fill the fields that have no alarm. Category attributes to completion. Alt text on A+ modules, which takes seconds per module. Item Highlights as a readable phrase rather than a comma salad. Every one of these is correct regardless of how any of this plays out, which is precisely why it is the move and not a bet.
3. Find out which third-party pages already rank in your category. Not to game them — to know whether the domains a model would plausibly scope to have you in them at all. Search your main category term the way a shopper describes a problem rather than a product, and write down the non-marketplace domains that come back. That list is your exposure, and most brands have never written it down.
4. Stop treating "are we in ChatGPT" as a yes/no. Ask the question the way a buyer would, in your own category, three or four different ways, and note whether the answer is built out of marketplace listings or out of editorial. Those are two completely different competitive situations and the answer differs by query inside the same category. Ten minutes, no tooling, and it is more than almost anyone in your category knows.
5. Do not start anything new before January. Nine weeks from peak. Nothing in this story justifies putting a new class of tool anywhere near a live catalog write path, and the things most likely to cost you real money between now and Christmas are an unreviewed catalog change, a Q4 budget carried forward as a dollar figure into peak CPCs, or inventory sitting in a receiving queue.
What I'd ignore
The retainer that is about to be built on this. By mid-September somebody will be selling an "AI source authority audit." It is attribute completeness, readable A+ copy, and a list of domains, which I have now published free twice. Make anyone pitching it describe the deliverable in one sentence.
The Reddit-is-finished discourse. One vendor's panel showed a reduced likelihood in a specific type of query over a specific window, and the person best placed to check could not confirm why. That is an observation, not a verdict on a platform, and it has no bearing on what you put in a bullet point.
Anyone quoting you a number for how much of your traffic is AI-referred. Nobody has that number for your account. Amazon does not report it as a category. If someone hands you a percentage, they made it up, and we have had enough of that this year.
The urge to build a separate machine-facing asset. Target and a handful of others spent this month building agent-facing front doors, and every one of them owns the surface an agent reads. You have one detail page, rendered by Amazon, and whatever a model extracts from it is what it gets. That is not a strategic problem you can solve. It is a housekeeping problem you can finish this week.
The line I keep landing on
For most of a decade the honest answer to "who reads your listing" was a shopper. Then it was a shopper and Amazon's model. Then it was a shopper, Amazon's model, and whatever third-party agent a court has now said is allowed to open the page.
What changed on August 8 is that one of those readers started choosing where to look before it looked. Nothing you can put in a title influences that decision. The only thing you control is whether, once it has chosen, the page makes its case with the pictures turned off.
Which is a thing I would have told you to fix anyway.