ChatGPT Search’s underlying retrieval just changed: ChatGPT is now doing large-scale, targeted site scraping that upends GEO strategy

What actually changed in ChatGPT Search

Promptwatch is a GEO platform that tracks visibility in generative search. It runs client-defined questions repeatedly across real interfaces like ChatGPT, Claude, and Gemini, recording the answers, citations, and backend search queries it can observe.

Its public 16-day dataset shows a striking break in the trend. Computing unweighted daily averages separately for the 13 days before August 8 and the 3 days after:

• Queries containing site: as a share of all fanout queries jumped from 0.368% to 16.78%, roughly 45.5 times the previous level;
• Search responses that included at least one site: query rose from 0.345% to 12.1%;
• The average fanout per search response increased from 1.06 to 1.837, up about 73%.

The denominator here is the easiest thing to misread. The "roughly one in six" figure refers to the expanded backend search queries, not one in six of all ChatGPT answers. The share of search responses that actually contain a domain-restricted query is closer to one in eight.

The first raw chart shows the shape of the change: it hovered near zero for weeks, then spiked almost vertically to 16–17% on August 8. A step change like this, completed in a single day, doesn't look like the gradual result of many sites optimizing. It looks like a coordinated adjustment on ChatGPT's side, either in search strategy or in how queries are displayed.

The second raw chart shows that the increase on that same day wasn't just about the site: string. The total number of search queries triggered per response also rose from about 1.08 to 1.83. That suggests domain-targeted search is layered on top of the existing search process, not replacing ordinary search outright.

How the new search flow might work

To understand this change, you first need to understand query fanout. When a user asks a single question, ChatGPT Search doesn't necessarily pass the raw sentence to a search engine. It may rewrite that question into several more specific queries, hunt for different pieces of evidence, and then synthesize the results into an answer.

The appearance of site: queries suggests this flow may now include an intermediate step:

1. Run general queries across the open web to find candidate answers and candidate sources;
2. Judge which domains are most likely to hold primary information;
3. Generate domain-restricted queries for the selected sites, looking for documents, pricing, rules, changelogs, or FAQs;
4. Choose pages from those results to use in the answer and citations.

The official ChatGPT Search help page already confirms that the system rewrites user questions into one or more targeted queries, and follows up with more specific queries as it reviews initial results. What isn't public is how "domain restriction" is represented internally.

Simon Willison speculates that ChatGPT's search tool may accept a structured parameter like search(query, recency, domains), with the interface rendering domain restrictions as the familiar site: syntax. This isn't baseless: OpenAI's public API web search tool already supports filters.allowed_domains.

But the public API doesn't prove ChatGPT's internal tooling is identical. We can confirm that domain constraints appear in volume, but we can't confirm whether they come from GPT‑5.6, a system prompt, the search router, a vendor interface, or simply a change in UI presentation.

Timing offers a clue: OpenAI updated GPT‑5.6 Sol on August 6, aiming to make answers more focused and reliable, with better use of dates, numbers, rules, and sources. Promptwatch's step change arrived two days later. That's a plausible timeline, not proof of causation.

Why this mechanism could improve answers

The most direct benefit of domain-restricted search is that it separates "finding an answer" from "checking it against primary sources."

Take a user asking about the latest price of a product. A general web search might surface outdated prices, republished articles, affiliate pages, and forum threads all at once. If ChatGPT first identifies the official site, then restricts its search to that domain to find the pricing page, help docs, and changelogs, it's far less likely to muddle different versions of the truth.

This flow could deliver three concrete benefits:

1. Reduced source ambiguity: same-name products, outdated pages, and secondhand reposts are less likely to creep into the final evidence;
2. Better in-domain recall: the system isn't stuck with the single most popular result from a search; it can keep digging for docs, rules, FAQs, and changelogs;
3. More controllable source selection: for questions about specs, policies, research, and official rules, the system can prioritize sites closer to the factual source.

The tradeoff is a narrower information base. If the candidate domain list is wrong, the system becomes more efficient at searching the wrong places. If the strategy skews persistently toward official sites and head publishers, independent experts, long-tail sites, forum counterexamples, and genuine user experiences may find it harder to surface in answers.

Different groups face different tradeoffs:

Promptwatch only measured changes in retrieval behavior. It did not test answer accuracy, hallucination rates, search costs, latency, or user satisfaction. So "may improve answers" is the product logic of this mechanism, not a verified experimental finding from this report.

Why GEO competition is expanding from pages to domains

Traditional SEO competes primarily at the page level: whether a page gets indexed, whether it matches a keyword, whether it outranks other pages.

If ChatGPT Search is now selecting candidate sites before digging deeper within them, GEO becomes a two-layer competition:

Layer one: domain selection. Why does ChatGPT consider a website worth searching further? A clear brand–topic association, primary information, sustained topical expertise, and references from other sources may all feed into that judgment.

Layer two: page selection. Once a domain is chosen, can the search engine actually find a page that answers the question? This depends on crawling, indexing, canonicals, stable URLs, internal linking, heading structure, and whether key information is stuck behind a login wall or in non-indexable components.

Only after those layers do you get to whether a page is cited, whether a citation earns a click, and whether that click leads to a signup or purchase. The full funnel is:

candidate domain → crawled page → included in answer → cited → user click → business conversion.

This doesn't mean page-level SEO is dead. It means page competition may now sit behind a domain-level filter. A well-crafted page could go completely unnoticed if its domain never makes the candidate list, leaving no chance for in-domain retrieval.

GEO, then, isn't just about chasing whether a single article gets mentioned by an AI. Sites need a coherent content system that provides primary, verifiable information, so the search system can easily recognize which topics this domain is worth mining further.

What sites should adjust right now

This is not the moment to invent clever new tricks around site:. The mechanism points to four fundamentals.

1. Put key facts on stable, primary pages

Product pricing, feature specs, version differences, research methods, author identity, changelogs, and policy rules belong on your own official website, not scattered across social platforms, PDFs, press releases, or sales conversations.

Each critical fact should have a stable URL, a clear last-updated timestamp, and a defined owner. Even after a domain is selected, ChatGPT still needs to find a page that directly answers the question.

2. Audit crawling, indexing, and internal links

Make sure key pages aren't blocked by robots.txt, login walls, or wrong canonicals. Clean up duplicate pages and dead links. Ensure clear internal links between product pages, documentation, FAQs, changelogs, and comparison pages.

Promptwatch suggests fixing on-site search, but traditional site: queries rely on the external search engine's index, not your site's search box. A good on-site search experience is valuable, but what matters more for this change is crawlability, indexability, and internal linking.

3. Build content coverage around real questions

Users don't just search for a brand name. They ask about price, comparisons, alternatives, compatibility, limitations, how-to, troubleshooting, and latest versions. Your site needs independently retrievable pages that answer these real sub-questions, with consistent terminology, product names, and version info.

That coverage only pays off if the domain, once targeted, still has enough content to complete an answer. Mass-produced keyword pages won't cut it.

4. Track visibility and business results separately

At minimum, separate fetched, cited, clicked, and converted. A page being fetched by a search tool is not the same as being included in an answer. Being included is not the same as being cited. Being cited is not the same as driving traffic and conversions.

Compare GEO baselines separately for before and after August 8. The underlying search strategy has shifted structurally; mixing old and new data risks mistaking a platform adjustment for your own site's optimization success.

For more on practical GEO execution and conversion boundaries, read on at

Related readingHow to get AI to recommend your SaaS product and drive real conversions: A GEO playbook from Ranking on AI's founderReal GEO success is measured in signups and revenue rather than mention screenshots, and it starts by ditching 'what is X' articles to capture high-intent 'best X' queries.→

What we still can't conclude

This change is significant, but it doesn't amount to a fully disclosed new algorithm.

At least four things remain unproven: whether ChatGPT actually uses a domains parameter internally; whether GPT‑5.6 directly triggered the August 8 change; whether domain-targeted search has actually improved answer accuracy; and whether citations are now concentrating among a handful of top-tier sites.

The Reddit development is worth watching. Promptwatch recorded Reddit's average daily share of ChatGPT citations dropping from 3.83% over July 18–August 7 to 0.52% over August 14–17, a relative decline of 86.4%.

The chart actually shows two breakpoints: the first drop on August 8, and a much larger collapse on August 14. That shows the two changes are adjacent in time, but it doesn't prove the site: adjustment caused Reddit's citation decline. Promptwatch itself notes that data collection issues can't be ruled out.

Going forward, what's worth watching isn't just the raw number of site: queries. It's which domains get targeted, whether citations become increasingly concentrated, whether long-tail citation share falls, and whether domain selection ultimately translates into clicks and business outcomes.

The most important takeaway from this mechanism shift for GEO is that AI search competition may no longer happen only between pages—it now begins by deciding whose site to search, then which page within that site to cite. Site strategy must therefore evolve from optimizing a few pieces of content to building a reliable source that is worth being targeted for retrieval.