Nadia Mohamed
AI-SEO & GEO
ChatGPT's Query Fan-Out Changed on 8 August: What site: Queries Mean for Your Domain
What is ChatGPT query fan-out?
Query fan-out is what an AI assistant does instead of running your question as a single search: it decomposes the prompt into several derived queries, retrieves against each, and synthesises one answer from the union. ChatGPT does this, and on 8 August 2026 the mix of queries it fans out to changed — the share of responses containing a domain-scoped
site:query rose from 0.37% to 16.8% of responses in a single day, according to Promptwatch’s measurement of live assistant interfaces.
Nearly everything written about query fan-out is about Google. The term entered the SEO vocabulary through Google’s AI Mode and a Google patent, and the pages that rank for it are explaining Google’s implementation. That is a different engine doing a different thing, and the distinction now matters, because ChatGPT’s fan-out behaviour visibly moved.
This article covers what was measured, what can reasonably be inferred from it, and — the part that takes up most of the page — what it does not license you to conclude. One publisher measured this. There is no second measurement. That constraint shapes every recommendation below.
TL;DR — Key takeaways
- On 8 August 2026, the proportion of ChatGPT responses containing at least one
site:-scoped fan-out query went from roughly 1 in 280 to roughly 1 in 6. - The average number of fan-out queries per response rose the same day, which is the basis for reading the change as additive — new queries added to the retrieval budget rather than substituted for existing ones.
- A
site:query resolves against an index. Pages that are not indexed cannot be surfaced by one, however good they are. - A domain-scoped question asks your site something. One excellent page does not answer it if the rest of the domain is silent on the topic.
- Every figure here comes from a single publisher, using its own dataset, with fewer than two weeks of post-change data. Treat the magnitude as provisional.
- Nothing about this requires an emergency change to your site. It sharpens an argument for indexation hygiene and topical coverage that was already sound.
What changed on 8 August
Promptwatch tracks which queries assistants issue behind their answers, sampling live interfaces rather than an API. On 10 August it published a measurement of one specific query shape: fan-out queries that use the site: operator to scope retrieval to a single domain.
The share had hovered between 0.3% and 0.5% for weeks beforehand, and dipped to 0.15% between 3 and 5 August. Then, on 8 August, it reached 16.8%.
In plainer terms: before 8 August, roughly one response in 280 involved ChatGPT asking a specific domain a scoped question. After, roughly one in six did.
The dip immediately before the jump is worth noting. A metric that sags slightly and then steps up is the shape a staged rollout leaves behind, and that is how Promptwatch reads it. It is a reading of a chart, not a confirmed deployment.
Fan-out, and why ChatGPT’s version is not Google’s
Fan-out is a retrieval strategy, not a ranking factor. Faced with “what’s the best headless CMS for a marketing site”, an assistant that runs that string verbatim gets one set of results shaped by one phrasing. An assistant that fans out generates several derived queries — comparisons, specific products, requirements, objections — retrieves against each, and composes from the union. The answer is better because the evidence base is wider.
The published SEO literature treats this as a Google AI Mode phenomenon, traced to a Google patent, and the advice that follows is about covering the sub-questions Google might generate. That advice is not wrong. It is about a different system. If you want the Google side of this, the comparison between AI Mode and AI Overviews covers how those two surfaces differ from each other, let alone from ChatGPT.
ChatGPT query fan-out is the same idea running inside a different pipeline, and a site:-scoped query is a different instrument again from a topical derived query. A topical fan-out asks the web a narrower question. A site: fan-out asks one domain a question.
The difference is not cosmetic. The first is a search for the best available answer; the second is an interrogation of a specific source. If ChatGPT is doing more of the second, it is spending more of its retrieval budget checking what particular domains have to say — which is a materially different thing from ranking pages against each other. How ChatGPT search works sets out the retrieval pipeline this change sits inside.
Additive, not a substitution
The most useful single number in the dataset is not the 16.8%. It is that the average fan-outs per response rose from about 1.08 to about 1.83 on the same day.
That is the evidence for the shape of the change. Promptwatch reads the change as additive rather than a substitution, and the reasoning is straightforward: if ChatGPT had swapped generic queries for site: queries, the count per response would have stayed flat. It rose by about 70%. So the site: queries were added on top.
That matters because the two readings imply opposite things. A substitution would mean domain-scoped retrieval crowding out open retrieval — fewer chances for an unfamiliar domain to be discovered. An addition means an extra retrieval pass, on top of whatever was already happening.
The inference is sound given the data, but it is an inference about a system nobody outside OpenAI can see. No announcement, changelog or documentation corroborates a rollout on that date. What exists is a behavioural measurement and a publisher’s reading of it.
What a site: query can and cannot surface
Here is the part that actually touches your site.
A site: query resolves against an index. It is not a crawl. It asks what a search index already holds for a domain, filtered to that domain. If a page is not in the index, a site: query cannot return it — not because the page is weak, but because the query has nothing to match against. Everything in crawled, currently not indexed is therefore upstream of this: pages sitting in that state are invisible to a domain-scoped retrieval pass in a way that no amount of on-page quality fixes.
A domain-scoped question is answered by the domain, not by a page. If an assistant scopes a question about pricing models to your domain, and your domain has one pricing page and nothing else on the subject, the retrieval returns one thin result. A competitor with a pricing page, two explainers, a comparison and a FAQ returns a body of material. This is the coverage argument that underpins topical authority in AI search, arriving through a different door: not “authority earns trust” but “a scoped query needs something to retrieve”.
The old failure modes still apply, now scoped to you specifically. A page blocked in robots.txt was already unreachable; a page that only renders client-side was already a gamble. What changes is the blast radius. Previously those pages lost a chance among many candidates across the web. Now, when a query is scoped to your domain, a gap is not a lost opportunity among competitors — it is your domain returning nothing for a question that was asked about you. What crawlers may and may not fetch is worth re-reading with that framing in mind, as is making a page legible once it has been retrieved.
And there is a limit to how much of this you control. You cannot make an assistant scope a query to your domain. You can only ensure that when it does, there is something there. If you want to be explicit about what an assistant should find, an llms.txt file is one way to state it — though whether any engine honours it is a separate question from whether you have published one.
What this does not tell you
This section is longer than it would be on most sites covering this story, and deliberately so.
Promptwatch says its data shows when, not why. The publisher states this plainly about its own charts. A step change on a date tells you something changed on that date. It does not tell you what was deployed, by whom, or whether it will persist.
There is no corroboration. Both the fan-out measurement and the companion Reddit citation measurement come from the same publisher and the same dataset. The story was picked up by the trade press on 18 August, but reporting on a dataset is not a second measurement of it. At time of writing, no independent source has measured ChatGPT site: fan-out rates. That is not a criticism of the publisher — it is unusual and valuable that anyone is measuring this at all. It is a statement about how much weight one dataset can carry.
The post-change window is short. Fewer than two weeks of data sit on the far side of 8 August. A step change that holds for two weeks may be permanent, may be an experiment, or may already have been rolled back by the time you read this. None of those is distinguishable yet.
Coincident dates are not causation. Reddit’s ChatGPT citation share first stepped down on the same day. It is easy to construct a mechanism — a domain-scoped retrieval pass would naturally favour the domain being asked about over a general forum — and that mechanism may well be right. It is also exactly the kind of story that sounds right and turns out to be two unrelated changes shipped in the same release train. The same scepticism belongs on citation share that decays over time generally: a number moving is not a number explained.
How to see this in your own data
You cannot measure ChatGPT’s fan-out rate yourself. You can measure the parts of this that touch your infrastructure.
Server logs are the only first-party record of assistant retrieval. Third-party citation trackers infer; your logs observe. If assistant crawlers are fetching more of your pages, or fetching different ones, the log is where it shows up first — with real timestamps, real status codes and real user agents. Log-file analysis is the method; the point here is that it is the only view that is genuinely yours.
Referral data catches the downstream half. A retrieval that produces a citation that produces a click lands in analytics like any other referrer, if you have configured it to be distinguishable. Tracking AI referral traffic covers the setup. Both halves are worth having: logs tell you what was fetched, referrals tell you what converted into a visit.
Separate what you control from what you don’t. Citation share is an outcome of somebody else’s ranking system and it moves for reasons you will never be told. Index coverage, render reliability, and whether your domain has substantive material on the topics you want to be asked about — those are yours. The 8 August change does not create new work. It raises the cost of work already left undone. If you want a structured pass over that, the GEO audit checklist is the version I run.
Set a review date rather than reacting. The honest response to a two-week-old signal from one source is to note it, instrument for it, and look again in a month. If the pattern holds and your own logs agree with it, that is when it becomes a reason to change a plan.
Frequently asked questions
Do I need to change anything on my site because of this?
No — not urgently, and not because of this specific finding. The change strengthens an existing argument for index coverage and topical depth rather than introducing a new requirement. If pages you care about are not indexed, fix that; you should have anyway. If your domain covers a topic in one thin page, deepen it; you should have anyway.
Is this the same as Google’s query fan-out?
No. Google’s fan-out, in AI Mode, is documented in a Google patent and discussed almost exclusively in the context of Google’s own systems. This measurement is of ChatGPT, a different engine with a different retrieval pipeline. The concept is shared; the implementation, the timing and the query shapes are not.
Does a site: query mean ChatGPT trusts my domain more?
There is no evidence for that reading. A scoped query indicates the assistant decided to ask a particular domain about something, which could reflect the prompt naming your brand, the domain already appearing in results, or mechanics not visible from outside. Nothing in the data speaks to trust.
Should I block site: queries from my domain?
There is no mechanism to block one specifically — a site: query runs against an index, not against your server, so there is nothing arriving at your infrastructure to refuse. Blocking assistant crawlers entirely is a separate decision with much broader consequences.
How reliable is the 16.8% figure?
It comes from one publisher measuring live assistant interfaces, published two days after the change it describes. The methodology is stated and the sample is large. It has not been independently reproduced, and the publisher itself frames its data as showing when things happened rather than why. Reliable enough to act on directionally; not reliable enough to build a forecast from.
Did this cause the Reddit citation drop?
The dates coincide, and a plausible mechanism exists. That is all. The publisher explicitly declines to attribute cause to its own charts, and so should anyone reporting them.
What would change my mind about how significant this is?
A second independent measurement showing the same step, or a month of sustained data at the new level, or a matching pattern in first-party server logs across several sites. Any one of those would move this from an interesting signal to a planning input.