ChatGPT does not answer most questions straight from memory. When a prompt needs current information, the model breaks it into a set of background searches — “fan-out queries” — runs them in parallel, and stitches an answer together from whatever comes back. Watching how those background searches have changed over the past year turns out to be one of the more useful ways to understand what OpenAI is actually optimizing for, and several independent researchers have now converged on the same pattern.

SEO researcher Lily Ray has spent months tracking this behavior using a mix of tools — Peec AI, Profound, the Resoneo Chrome plugin, FanoutFox, and Google Search Console — and combined her own observations with published research from several other analysts. The consensus: ChatGPT is running far more searches per prompt than it used to, scoping a growing share of them with the site: search operator, and increasingly narrowing its sources to a short list of domains it already trusts.

Retrieved versus cited: a distinction worth knowing

Two words get used interchangeably in AI search discussions that actually mean different things. Retrieved means a page ChatGPT fetched while running its fan-out queries. Cited means a page that made it into the visible answer as a link. Those two numbers are currently moving in opposite directions: the number of retrieved URLs per response is rising, while the number of unique domains cited per response is falling. More pages are being considered; fewer are getting credit.

That gap matters when you’re reading any AI-search study — check whether it’s measuring retrieval or citation before comparing its numbers to another one, because the two rarely agree.

What independent research is finding

Several analysts have measured this behavior separately, with different tools and different methodologies, and their numbers point the same direction:

• David Konitzny at Peec AI tracked the rollout of ChatGPT 5.6 and found the share of prompts using only a single fan-out query dropped from 94.0% to 43.5%. Average retrieved sources roughly doubled, from about 12 to 24, and prompts requiring a second fan-out iteration jumped from about 5% to 33.5%. Use of the site: operator went from appearing in roughly 0.3% of fan-outs to about 23%.
• Chris Long at Nectiv compared roughly 4,000 prompts on ChatGPT 5.6 “Sol” against a 2025 baseline, finding average fan-out queries per prompt rose from 2.17 to 7.61, with the longest observed query chain going from 4 searches to 29. “Site:,” “official,” and “gov” ranked among the most common terms, with site: appearing in 64% of queries — a higher figure than Peec AI’s, likely reflecting a different collection method (API versus consumer interface).
• Olivier de Segonzac and the Resoneo team found unique domains cited per response dropped from 19 to 15 after the 5.3 update — the citation side of the same retrieval-versus-citation gap.
• Suganthan Mohanadasan analyzed 57 conversations and 3,554 retrieved pages and found that 21 of 27 initial fan-out queries already contained brand names the user never mentioned, across 11 of 13 product categories tested. Brands named in that first query were cited 68.9% of the time; pages that were merely fetched without being named upfront were cited just 2.1% of the time. Only about 3.1% of all retrieved pages made it into a final answer.

Ray’s own testing turned up a similar pattern: ChatGPT’s reasoning-tier models fire off ten or more searches for a single prompt, including multiple site: queries, while faster/cheaper tiers often run just two or three. The type of domain targeted tracks the type of question — opinion and product-review prompts lean on Reddit and specific subreddits, YMYL questions skew toward .gov and .org domains (in one batch of roughly 20 legal-topic prompts, every citation came from a .gov domain), and spec-and-pricing questions go straight to the relevant brand’s own site.

A theory: the site: operator as a spam filter

Ray draws a direct comparison to Google’s E-E-A-T framework. Evaluating whether a random page on the open web is trustworthy is hard and expensive to do at ChatGPT’s scale — hundreds of millions of queries a day. Narrowing fan-out searches to .gov domains, established review platforms, recognizable brands, and official sources sidesteps that problem: instead of judging an unknown page’s credibility in real time, the model simply looks in places it already considers reliable.

The pattern also shows up in page-type data. Product pages now make up 16.39% of retrieved pages, per Peec AI’s numbers, overtaking listicles — and the formats losing retrieval share (listicles, how-to guides, comparison pages) happen to be exactly the formats most commonly targeted by manipulative GEO tactics. Read together with the site: patterns, it looks like a deliberate move away from exactly the kind of self-serving content that has been a known SEO gaming vector for years — a shift covered in more detail in ChatGPT already knows who it’ll recommend before it searches.

The risk: guessing the wrong domain

Heavy reliance on site: searches only works when the model has correctly identified which domain belongs to a brand. Research from Malte Landwehr and a 2025 Netcraft study both found cases where it hadn’t. Landwehr documented ChatGPT repeatedly running site:census.com for the startup Census, whose actual domain is getcensus.com — census.com is an unrelated parked domain. He found the same misfire pointing at other parked domains, including lago.io, persona.id, lightfield.ai, and mesa.com. Netcraft’s study found roughly a third of brand login links generated by LLMs pointed to domains the brand didn’t own, and about 29% pointed to unregistered or parked domains — with smaller, less-established brands the most exposed.

That is a real risk for brands without a strongly reinforced official domain: someone could register the parked domain ChatGPT keeps guessing at and quietly feed the model wrong information.

Site: searches are showing up in Search Console

Site owners checking Google Search Console and Bing Webmaster Tools have found thousands of impressions for site: queries against their own domain with close to zero clicks — a pattern consistent with bot activity (LLMs, tracking tools, scrapers) rather than human searchers, and one that has been increasing over time. There’s no dedicated “ChatGPT Search Console,” so this remains inference rather than confirmed attribution, but it’s worth checking: in Search Console’s Performance report, filter Query by “Queries containing” site:; in Bing Webmaster Tools, search the query table the same way.

Bing has moved further here than Google — its AI Performance Report, launched inside Webmaster Tools in February 2026, separates AI citations from traditional search and surfaces the “grounding queries” behind Copilot’s answers directly.

What this means in practice

• Keep investing in search rankings. Because ChatGPT still leans on major search indices to populate its fan-outs, ranking well in ordinary search results continues to feed what the model retrieves.
• Make key facts readable as plain HTML. Pricing, specs, and support details need to be text the model can actually parse — not locked behind client-side JavaScript rendering or images.
• Reinforce which domain is official. Consistent, unambiguous signals about your real domain — including “Official Site” language where it reads naturally — are the best defense against a model guessing wrong, a risk detailed further in how ChatGPT actually picks its sources.
• Don’t chase individual fan-out queries. They’re long-tail and vary by user and session. It’s more useful to aggregate the topics fan-outs keep returning to and rank well for those broader terms in traditional search.
• Check whether your brand shows up at all. Run the prompts that matter to your business several times and look at ChatGPT’s initial searches, not just the final answer. If your brand never appears in that first round, the fix is reviews, comparisons, and coverage that build category recognition — not a technical patch.
• Take Reddit seriously, carefully. Separate research has found ChatGPT discards the vast majority of Reddit pages it retrieves — yet Reddit still tops most-cited-domain lists, because it gets pulled so often that even a small survival rate adds up. That dynamic is explored further in ChatGPT often retrieves but rarely cites Reddit pages. Artificially seeding Reddit with brand mentions is not a safe workaround — Reddit has been actively removing that kind of manipulation.

The broader signal is that this behavior has moved down from premium, reasoning-tier models to the free default that the large majority of ChatGPT’s users are on, following the rollout of 5.6. Whatever the free tier does when it searches is now close to what most people experience — which is covered in more depth in how ChatGPT’s default and premium models search the web differently.

Frequently asked questions

What is a ChatGPT fan-out query?

It’s one of several background searches ChatGPT runs automatically when a prompt requires current information. The model deconstructs the prompt into multiple targeted searches, runs them in parallel, and synthesizes its answer from the combined results.

Why does ChatGPT use the site: operator so much now?

Multiple independent studies show a sharp rise in site:-scoped fan-out queries, especially toward .gov domains, established review platforms, and official brand sites. The leading theory is that it functions as a low-cost way to avoid retrieving low-quality or manipulated content, similar in spirit to how Google’s E-E-A-T guidance elevates authoritative sources.

Does this mean SEO doesn’t matter for AI search?

The opposite. Because ChatGPT still draws on traditional search indices to populate its fan-outs, ranking well in ordinary search results remains one of the clearest ways to influence what the model retrieves and ultimately cites.

How can I tell if ChatGPT already knows my brand?

Run the prompts relevant to your business several times and check whether your brand name appears in ChatGPT’s initial fan-out queries, not just its final answer. Tools like Peec AI, Profound, or the Resoneo plugin can surface the actual queries the model runs rather than just the response text.