ChatGPT now searches brand domains directly. A readiness audit for your own pages

KEY TAKEAWAYS

1. The shift is real in the consumer app. Any single percentage belongs to one tracker's surface and prompts.
2. In Peec's data 84% of scoped searches target brand-owned domains, and 91.9% name a domain, not a page.
3. Own pages help brands ChatGPT already names. Unnamed brands still need third-party coverage first.
4. Audit by running site: searches for pricing, specifications and integrations on your own domain.

On 8 August 2026 several tracking tools saw ChatGPT change how it searches. A large share of its background searches began to name one domain, written like a site: search, and most of those domains belong to brands. Coverage since then has been about Reddit losing citations. This article is about the other side: what a scoped search finds when it lands on your own site, and how to check before it does.

What changed on 8 August?

When ChatGPT answers with web search, it writes its own search queries in the background. These are called fan-out queries. Our piece on how AI engines choose which brands to name covers the general mechanics.

On 6 August 2026 OpenAI updated the GPT-5.6 Sol and Luna models in ChatGPT, replacing GPT-5.5 Instant. Its post says nothing about search behaviour. Two days later Promptwatch, which monitors the consumer app, recorded the share of fan-out queries using the site: operator rising from 0.37% to 16.8% in one day. Searches per response rose from 1.08 to 1.83. Promptwatch does not state its sample size.

OpenAI has not documented any of this. Simon Willison's reading is that the search tool probably takes a domain as a parameter, and that the model is not literally typing an operator. For a site owner the effect is the same: the search is confined to one site.

The behaviour is not entirely new. Search Engine Land published Resoneo's analysis in May 2026, which noted that an earlier thinking model already restricted searches to trusted domains. August moved the pattern into the default experience.

How big is the shift? It depends on who measured

No single number is right. Here is what each tracker reported.

Dejan's result looks like a contradiction. Dan Petrovic found 20 scoped queries among 14,460 OpenAI fan-outs in August, and none among 371 on 8 August itself. He does not say Promptwatch is wrong. He says the two tools sample different things: "Surface, prompt wording, and prompt mix are all candidate causes of the gap."

His own data shows how much prompt style matters. Long ad hoc prompts produced scoped searches 6.29% of the time. Short fixed monitoring prompts produced them 0.44% of the time.

Elmo's data points the same way. It tracks both the app and the API. In its numbers Reddit's citation share collapsed in the app in August while the API stayed low and flat, so the two surfaces were already behaving differently.

Our reading: the shift is real in the product buyers use. Every tracker that samples the app or logged-in sessions shows a large rise. But each percentage is a property of one surface, one model tier and one prompt set. Quote the range, roughly 17% to 76%, and do not plan around any one figure. This is the same caution we apply to AI visibility sampling.

Where do the scoped searches go?

Mostly to brands. In Peec's data, 84% of the domains targeted were brand-owned and 16% were not. Among the non-brand targets, Reddit made up 71%. Trustpilot was 54% of the review and comparison category.

The second finding matters more for planning. Peec tracked 500 prompts a day for 17 days after the model update, with no brands named in the prompts. In 91.9% of scoped searches ChatGPT named only a domain. About 9% included a path such as a pricing or product page. A core set of domains, 18.1%, came back almost every day. Pages did not: only 3.5% of URLs repeated across the dataset.

So ChatGPT settles on domains it trusts and then explores inside them. You cannot prepare one page and be done. The domain as a whole has to answer attribute-level questions. Peec's examples pair the domain with terms such as specifications, pricing, model names, official and review.

Lily Ray adds that scoped searches also go to government domains on legal and health topics, to established review platforms and to major retailers. She compares the behaviour to a spam filter. Qwairy's data points the same way for regulated topics: institutional and government sources gained while brand sites held steady.

What moved in citations?

Keep two things apart here. A retrieved page is one ChatGPT read. A cited page is one it credited in the answer. Many articles quote the two interchangeably.

Reddit fell, in ChatGPT. Promptwatch measured Reddit at 3.83% of ChatGPT citations from 18 July to 7 August, and 0.52% from 14 to 17 August. Elmo's separate dataset shows 4.67% falling to 0.38%. On Google the same Promptwatch data shows small, gradual moves: 2.37% to 2.10% in AI Overviews and 2.22% to 1.54% in AI Mode.

Own and official pages rose. Discovered Labs analysed 291,607 ChatGPT citations from B2B software prompts between October 2025 and August 2026. Documentation pages went from 36.9% to 55.1% of citations and listicles from 6% to 1%. Green Flag Digital re-tested 215 finance pages in September 2026 and found first-party brand content rising from 2% to 22% of top citations. Both cover long windows with several model changes, so neither isolates August.

Retrieval shifted the same way. Peec found listicles falling from 15.47% to 8.19% of retrieved pages, with product pages at 16.39%.

Two cautions. First, the scoped-search change is dated 8 August and Reddit's sharp fall is dated 14 August. Search Engine Journal pointed out the six-day gap, and nobody has shown that one caused the other. Second, Reddit is not gone. Ahrefs' list of most-cited ChatGPT sources, updated on 2 September, still puts Reddit first on its own metric. No tracker we read has published a figure for September or October, so whether the fall has held is not established.

Your own pages only help if ChatGPT already knows you

This is the part most coverage leaves out. A scoped search needs a domain to name, and ChatGPT chooses that domain before it searches.

Suganthan Mohanadasan captured network traffic from 57 logged-in conversations. ChatGPT retrieved 3,554 pages and cited 110 of them. Brands that ChatGPT wrote into its own queries reached the answer 68.9% of the time. Brands that only turned up in retrieved pages reached it 2.1% of the time. It is a small sample from one account, and he labels it directional.

The direction still matters. The August change raises the value of first-party pages for brands already on the shortlist. It does little for a brand the model does not name. That brand has a coverage problem, which is what the press coverage and brand entity pillars of digital authority measure, and what the placement marketplace exists to close.

Which situation are you in?

This table is our synthesis of the findings above.

A seven-step readiness audit

1. Run the shortlist test. Write 10 to 20 buyer prompts with no brand in them. Ask each in the ChatGPT app with the default model. Record whether your brand is named.
2. Write the attribute queries. For each prompt where you were named, list what a scoped search would ask your domain: pricing, specifications, model names, integrations, security, comparisons.
3. Run them as site: searches on Google and Bing. This is a proxy. It shows whether a search engine can surface a page of yours for that attribute.
4. Score each one. A page that states the fact in HTML text is a pass. A PDF, a script-only calculator, a gated page or no result is a fail. If nothing returns at all, check crawler access first.
5. Fix fails on stable URLs. Add the fact to a page that already exists and will keep its address. Since ChatGPT targets domains far more often than pages, coverage across the site beats one hero page.
6. Check the non-brand targets. Look at what Reddit, Trustpilot and the review sites in your category say, since ChatGPT still searches them on purpose.
7. Repeat monthly. Log two numbers: how often you are named, and your pass rate. Seer's study argues that citation counts alone now miss what is happening earlier in retrieval.

One thing not to do: build a page for every fan-out query. Lily Ray makes this point, and the Peec data supports it, because the queries rarely repeat.

A worked example with invented numbers

This company is hypothetical. It sells invoicing software and tests 12 buyer prompts.

• ChatGPT names the brand in 5 of 12 prompts, or 42%.
• For those 5 prompts it writes 4 attribute queries each, 20 in total.
• 9 of the 20 return a page with the fact in HTML text. The pass rate is 45%.
• The 11 fails have three causes. Pricing sits in a script-driven calculator (4 queries). The integrations list is a PDF (3). There is no security page (4).

The company publishes a static pricing table, an HTML integrations page and a security page. A re-test passes 18 of 20, or 90%.

The other 7 prompts, where the brand was not named, do not move. They are a third-party problem. Page fixes will not change them.

The pass rate is a readiness measure the company controls. It does not predict a citation rate.

What is still unknown

OpenAI has not explained the change. Semrush reports that the company told Yahoo Tech it does not set a fixed level of visibility for any site. The trackers disagree on basics: Peec says ChatGPT now reads about twice as many sources per prompt, while Seer counted 43% fewer. And every figure here comes from August or early September.

Treat the direction as settled and the decimals as provisional. If you want to see how your brand is named across engines, join the waitlist.

Questions

What is the ChatGPT site: operator change?

From 8 August 2026, tracking tools recorded ChatGPT limiting many of its background searches to a single domain, written like a site: search. Promptwatch measured a rise from 0.37% to 16.8% of fan-out queries in one day. OpenAI has not documented the behaviour, and its 6 August model update post does not mention search.

Did ChatGPT stop citing Reddit?

No. Promptwatch measured Reddit's share of ChatGPT citations falling from 3.83% to 0.52% in mid-August, and Elmo's data shows a similar fall. But Peec's data has Reddit as 71% of the non-brand domains ChatGPT still searches on purpose, and Ahrefs still lists Reddit as a top cited source.

Does the change apply to the OpenAI API?

Not uniformly. Dejan found site: in only 0.14% of its OpenAI API fan-outs in August, while Nectiv measured 64% on the API with a different model and prompt set. Check the consumer app with the default model, because that is what buyers use.

Should I create a page for every fan-out query?

No. Lily Ray warns against it. The queries are long, vary by prompt and change between runs, and a page per query risks Google's scaled content abuse policy. Add the missing facts to stable pages that already exist.

Sources

1. OpenAI, Improving GPT-5.6 Sol in ChatGPT openai.com
2. OpenAI Deployment Safety Hub, GPT-5.6 August Updates deploymentsafety.openai.com
3. Promptwatch, Why Did ChatGPT Stop Citing Reddit? promptwatch.com
4. Promptwatch, How ChatGPT Searches the Web in 2026 promptwatch.com
5. Peec AI, ChatGPT site-operator fan-outs docs.peec.ai
6. Peec AI, GPT-5.6 retrieval changes docs.peec.ai
7. Peec AI, ChatGPT fan-out domains vs URLs docs.peec.ai
8. Peec AI, GPT-5.6 product pages vs listicles docs.peec.ai
9. DEJAN, Our OpenAI fanout data shows no rise in site: operator use dejan.ai
10. Seer Interactive, AI Citations Aren't Your Top KPI, ChatGPT-5.6 Luna Runs More Targeted Searches seerinteractive.com
11. Nectiv, What We Learned From Analyzing 28K+ ChatGPT And Gemini Fan-Out Queries nectivdigital.com
12. Elmo, ChatGPT UI Reddit Citations Fell to API Levels elmohq.com
13. Qwairy, ChatGPT's Reddit Citations Collapsed 95% qwairy.co
14. Suganthan Mohanadasan, ChatGPT research suganthan.com
15. Discovered Labs, How ChatGPT retrieves in 2026 discoveredlabs.com
16. Green Flag Digital, ChatGPT Citation Study greenflagdigital.com
17. Ahrefs, The 50 Most-Cited Websites in ChatGPT ahrefs.com
18. Simon Willison's Weblog, ChatGPT Search now uses the site: operator at scale simonwillison.net
19. Search Engine Journal, Why Reddit's ChatGPT Citation Drop Isn't Fully Explained searchenginejournal.com
20. Search Engine Journal, What We Can Learn from Evolving ChatGPT Fan-Out Queries searchenginejournal.com
21. Lily Ray, What We Can Learn from Evolving ChatGPT Fan-Out Queries lilyraynyc.substack.com
22. Search Engine Land, Inside ChatGPT Search, web.run fan-out queries searchengineland.com
23. Semrush, Reddit's citations in ChatGPT fall semrush.com

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →