How to Make Your Entire Site Relevant for ChatGPT's Domain-Scoped Queries

What changed in how ChatGPT searches the web?

ChatGPT started choosing a domain before it chooses a page. Search Engine Journal's review of tracking data reported that on August 8, 2026, the share of ChatGPT Search fan-out queries using the site: operator, the query type that restricts a search to one domain, jumped from about 0.37% to 16.8% in a single day, a roughly 46-fold increase (Search Engine Journal, 2026-08-19). The average number of fan-out queries per response nearly doubled the same day, from about 1.08 to 1.83.

That is a mechanical shift, not a ranking tweak. Instead of running one open search and pulling snippets from whatever comes back, ChatGPT is now issuing a first round of queries to decide which domains deserve a visit, then a second round scoped to those domains. Klaas Foppen, co-founder of Promptwatch, described the behavior plainly: ChatGPT 'is no longer just searching the open web and seeing what comes back; it is deliberately going to specific sites to pull information from them' (Search Engine Journal, 2026-08-19). Domain selection is now a gate that sits above page-level optimization, and a page can be well built and still never get a look if its domain does not clear that gate.

Fan-out is the term for the round of background searches ChatGPT runs before it writes an answer, often several per turn, each aimed at a different angle of the question. A domain-scoped fan-out query is one of those background searches restricted to a single site with the site: operator. Before August 8, that restriction was rare, under half a percent of fan-out queries. After August 8, it is closer to one in six. The model is not searching less. It is deciding, earlier and more often, which handful of domains are worth entering at all, and treating everything outside that shortlist as noise.

Who does this affect, and how do we know it is real?

Reddit is the clearest casualty so far. Its share of ChatGPT Search citations averaged 3.83% from July 18 to August 7, 2026, then fell to an average of 0.52% from August 14 to August 17, an 86.4% relative decline (Search Engine Journal, 2026-08-19; Forbes, 2026-08-20). Over the same stretch, Google's AI Overviews and AI Mode showed only gradual declines in Reddit's presence, down 11.3% and 30.5% respectively, which points to a ChatGPT-specific mechanism rather than a broad, engine-wide drop in Reddit's authority.

Any publisher whose visibility depends on being pulled into an AI answer is exposed here, not just forums. Sites built around a single strong page sitting inside a domain with no clear topical identity are the most likely to lose fan-out visits, because domain-scoped selection rewards a site ChatGPT can label quickly. That is the same test behind Citation Engineering: coverage and clarity at the domain level, not a single page optimized in isolation.

B2B software teams and local service businesses feel this differently. A SaaS company competing for 'best [category] tool' prompts needs its domain to read as the category authority, not a marketing site with one buried comparison page. A local or multi-location business competing for '[service] in [city]' prompts needs the reverse problem solved: enough consistent, entity-tagged location pages that the domain reads as a real, current source for that service, not a thin directory listing. Both groups lose the same way if the domain itself gives ChatGPT nothing clear to point to.

What does the before-and-after actually look like?

The table below lines up the two windows analysts have published data for. Read it as a gate closing on August 8 and a citation reshuffle showing up roughly a week later, not as one clean cause and effect. Four numbers matter most: the operator's own jump, the response-level query count that jump produced, Reddit's citation share on ChatGPT specifically, and Reddit's citation share on Google's two AI surfaces over the identical window, included so the ChatGPT-specific pattern is visible rather than assumed.

Why is there a six-day gap between the change and the collapse?

This is the part worth being honest about. ChatGPT's site: operator behavior changed on August 8, 2026, but Reddit's sharpest citation decline landed on August 14, six days later, and Promptwatch has not explained that second drop (Search Engine Journal, 2026-08-19). Promptwatch also told Search Engine Journal it cannot rule out a data-collection issue on its own end. Promptwatch's tracking data, as reported by Search Engine Journal, shows when the shift happened. It does not show, conclusively, why Reddit specifically lost share.

There is also a precedent for false alarms. Reddit saw a similar citation collapse in September 2025 that turned out to be caused by Google removing a search parameter, unrelated to OpenAI or to Reddit itself, and citations eventually rebounded (Search Engine Journal, 2026-08-19). None of that erases the underlying mechanism, domain-scoped fan-out is real and measured, but it argues against rebuilding a whole content strategy off two and a half weeks of data. Treat the operator jump as confirmed and the specific site-by-site fallout as still forming, two different confidence levels living inside the same news cycle.

How should you respond right now?

Two things can be true at once: the site: operator jump is confirmed and dated, and Reddit's exact citation drop still has an unexplained gap in it. Respond to the confirmed part and watch the uncertain part.

Treat your whole domain as the unit ChatGPT is evaluating before it ever opens a page. A domain that reads as one clear topic, with a consistent entity behind it (the same brand, author, or organization named the same way across pages) and dense coverage of the questions in that topic, is easier for a model to select with confidence. A domain that mixes unrelated subjects across a handful of thin pages is not. This is the practical core of answer engine optimization: winning the domain-level gate before the page-level one.

Separate a platform-wide shift from your own performance problem before you touch anything. If your citations dipped in the same window as Reddit's, check whether the drop lines up with August 8 or August 14 specifically, and whether competitors in your category moved too. A dip that started earlier or that no competitor shares is your problem, not ChatGPT's.

Do not move budget on this alone. Give it a few more weeks of data before restructuring a content program around one operator's fan-out behavior. Publish for topical concentration and a clear entity regardless of how this specific pattern resolves; that work compounds under any future version of domain-scoped search.

1. Audit your domain's topical concentration: one clear subject beats five scattered ones.
2. Name your entity the same way (brand, author, organization) on every page a crawler might land on.
3. Compare your citation trend against the August 8 and August 14, 2026 dates specifically, not just 'recently'.
4. Check whether competitors in your category moved with you before assuming ChatGPT singled you out.
5. Hold budget for two to three more weeks of data before restructuring strategy around this pattern alone.

Does this replace traditional SEO and page optimization?

No, it adds a step in front of it. Page-level work, clear headings, direct answers, real sourcing, structured data, still decides whether ChatGPT uses a page once it opens a domain. Domain-scoped fan-out just decides whether the model bothers to open the domain in the first place. A single excellent page on a scattered, unfocused site is now competing at a disadvantage against a mediocre page on a domain that reads as a coherent authority, because the second domain is more likely to survive the shortlist stage.

The practical implication for a publishing program is sequencing. Coverage breadth across a topic, the kind that comes from consistently answering the full range of questions a buyer or reader asks in a category, builds the domain-level signal that gets a site onto the shortlist. Page-level craft is what happens once it is there. Programs that publish at real scale, ten or more articles a day rather than one a week, close both gaps at once: they build topical density fast enough to read as a coherent domain, and each individual piece still has to meet the bar on its own.

What should you watch next?

Watch whether the site: operator share holds near 16.8% of fan-out queries or drifts back down, and whether Reddit's citation share stabilizes, keeps falling, or rebounds the way it did after the September 2025 episode. Citation share, the percentage of relevant AI answers in a category that cite a given domain, is the metric that will show whether domain-scoped fan-out is a permanent gate or a temporary tuning pass. Either way, the sites that win it will be the ones a model can identify as the single clear authority on a topic before it ever opens a page.

None of this changes what has always driven citation, it raises the entry cost. A model that scopes its search to a handful of domains before reading a single page is making its shortlist decision on thinner evidence: a domain's apparent focus, its consistency, and how often it already turns up as an answer to the surrounding questions in a category. Coverage across the full question universe for a topic, not one flagship page, is what earns a spot on that shortlist. Track fan-out behavior for a few more weeks, but build for the shortlist regardless of how this particular pattern resolves.

Sources

Source: Search Engine Journal

The site: operator's share of ChatGPT fan-out queries jumped from 0.37% to 16.8% on August 8, 2026, and Reddit's citation share fell from 3.83% to 0.52%. Search Engine Journal, 2026-08-19

Source: Forbes

Reddit's presence in ChatGPT Search citations collapsed within days following OpenAI's search change, while Google's Overviews and AI Mode declined only gradually. Forbes, 2026-08-20