ChatGPT picks the site before the page

Domain-scoped fan-out is when an assistant decides which sites to ask before it searches, then runs queries restricted to those domains, so being a site the model thinks to name becomes a precondition for being retrieved at all.

That is the shape of the change ChatGPT Search appears to have made this month, and Reddit is the most visible casualty of it.

#What the tracking data shows

Promptwatch, a GEO analytics firm that samples live assistant interfaces, reported that reddit.com held an average of 3.83% of ChatGPT Search citations from July 18 to August 7, then averaged 0.52% from August 14 to August 17, a relative drop of 86.4%. The measurement is Reddit's share of citations among responses that included at least one citation, drawn from millions of collected AI responses.

The same window looks nothing like that on Google. Over those weeks Google AI Overviews showed an 11.3% relative decline in Reddit citations and Google AI Mode showed 30.5%, a gradual slide rather than a cliff. Whatever happened was specific to one engine's retrieval behavior, not a web-wide shift in how much Reddit content exists or how good it is.

#The mechanism worth paying attention to

The interesting number is not the one about Reddit. On August 8, the share of ChatGPT fanout queries using the site: operator jumped from about 0.37% to 16.8% in a single day, roughly a 46x increase, while the average number of searches run per response nearly doubled from about 1.08 to 1.83.

Read that as a pipeline change. Query fan-out is the step where one user question becomes several web searches. Previously most of those searches were open: send the sub-query to the index, see what comes back, and let page-level relevance decide. A large minority of them are now domain-scoped: pick a site, then search inside it.

That inserts a gate above everything a GEO checklist normally addresses.

Stage	Open fan-out	Domain-scoped fan-out
What is selected first	A query	A domain
What decides inclusion	Page relevance in the index	Whether the model names your site
What an on-page rewrite affects	Retrieval and citation	Citation only, if the domain was picked
Failure mode	Your page ranks poorly	Your page is never eligible

Reddit is a clean illustration of the asymmetry. Reddit hosts an enormous amount of relevant discussion, and it ranked well in open retrieval for exactly that reason. But almost nobody, human or model, resolves a product question by deciding to search inside reddit.com specifically. Documentation sites, vendor sites, and reference works are the natural targets of a domain-scoped query. Community platforms are not.

Note what this does to the value of page-level optimization. If a meaningful share of the retrieval budget is spent on searches your domain was never selected for, then the ceiling on those searches is set by domain identity, not by anything in your markup or your prose.

#Why you should not over-read it

Promptwatch itself is careful here, and the caveats deserve more space than they usually get.

The firm said its data shows when the shift occurred, not why, described the size of the drop as provisional, and stated it could not yet rule out a data-collection issue on its own end. The timing also does not line up cleanly. The site: operator change landed on August 8 and coincided with a first, smaller decline into the mid-2% range. The sharper collapse below 1% came on August 14, six days later, and the operator change does not explain it. There are two drops and one proposed cause.

There is also precedent that cuts against confident attribution. In September 2025 Reddit's ChatGPT citation share fell from roughly 7% to roughly 1%, then rebounded to roughly 3%, based on Profound's analysis of over 4 billion AI citations and 300 million answer engine responses. That episode was traced to Google removing the num=100 search parameter, a change in how data could be collected rather than a change in what ChatGPT valued. A collapse of similar magnitude, with an unrelated cause, fully recovered eleven months ago.

Claims about who picked up the lost share are weaker still. The most-repeated figure, that documentation and help centers grew from 2% to 32% of citations, comes from one practitioner's model-version comparison and has not been independently verified. It is directionally consistent with the site: theory, which is exactly why it should be held loosely rather than quoted as a finding.

#What to actually do with this

Three things follow, none of them a content rewrite.

Separate platform shocks from your own performance. If your AI citation share moved in the second week of August, the first question is whether the source mix moved under you. A metric that collapses in four days with zero content change is not a content problem, and treating it as one produces a quarter of wasted work. Your baseline can shift for reasons that have nothing to do with your pages.

Treat domain selection as a distinct objective. Page-level GEO answers "is this the best passage for the query." Domain-scoped fan-out asks a prior question: "is this the site I would think to search for this topic." Those are optimized differently. The second is served by topical concentration, a clear and consistent entity, being the primary source rather than a commentator, and having a site an assistant can characterize in one sentence. It is closer to positioning than to markup.

Do not restructure a channel strategy on four days of one vendor's data. Community presence on Reddit and similar platforms was never justified solely by ChatGPT citation share, and the September 2025 round trip is a reasonable prior for what happens next. Watch it for a few more weeks before you move budget.

The durable lesson is structural rather than tactical. As assistants run more searches per answer and spend more of them on domains they chose in advance, the competition moves upstream. Being the best page stops mattering when the engine never asks your site.