Key takeaways
- On August 8, 2026, the share of ChatGPT fan-out queries using the
site:operator jumped from 0.37% to 16.8% in a single day, according to Promptwatch Data. That's roughly a 46x increase. - The jump wasn't a substitution. Average fan-out queries per response nearly doubled the same day (from ~1.08 to ~1.83), meaning
site:queries got added on top of generic web searches, not swapped in for them. - Independent research from Peec AI's David Konitzny found similar numbers around the GPT-5.6 default-model rollout: single-fan-out prompts dropped from 94% to 43.5%, and
site:usage rose to roughly 23% of fan-outs. - Because
site:-scoped queries only retrieve what's already indexed on a domain, thin category pages, broken internal search, and unindexed content now cost you visibility in AI answers directly, not just in classic search rankings. - If your keyword-to-heading mapping or fan-out tracking setup predates early August 2026, treat it as stale. Re-run your analysis.
What actually happened on August 8
I want to start with the blunt version, because it's a genuinely unusual data point. For weeks, the share of ChatGPT fan-out queries containing a site: operator sat between 0.3% and 0.5%. It even dipped slightly lower, to about 0.15%, between August 3 and 5, which in hindsight looks like a staged rollout or an A/B test running quietly in the background. Then on August 8, it jumped to 16.8%.
That's not a gradual trend line. That's a light switch. Promptwatch's data (https://promptwatch.com/data/chatgpt-site-operator-fanouts) shows the shift happening within roughly 24 hours, which points to a model or system-prompt change on OpenAI's side rather than any kind of organic behavior drift. Whatever changed, it changed for everyone, all at once.
What's a fan-out query in the first place? When you ask ChatGPT something and it decides to search the web, it often doesn't run just one search. It breaks your prompt into several sub-queries covering different angles, runs them in parallel or sequence, and then synthesizes an answer from whatever it retrieves. Promptwatch's broader fan-out tracking (https://promptwatch.com/data/chatgpt-query-fanouts) shows this average climbing and falling over time, from around 2.15 queries per response in early December down to a lean 1.0 by April, before the August 8 event pushed it back up near 1.83. A site: operator inside one of those sub-queries means ChatGPT deliberately went to a specific domain, not the open web, to pull an answer.
Put plainly: ChatGPT is no longer just throwing a question at the web and seeing what floats to the top. In a growing slice of cases, it's picking a destination and going there on purpose.
It wasn't just Promptwatch's data
I'd be skeptical of a single-source data point this dramatic if it weren't corroborated elsewhere, so it's worth noting that it was. David Konitzny at Peec AI measured the day ChatGPT 5.6 became the default model and found comparable shifts: the share of prompts with only a single fan-out query fell from 94.0% to 43.5%, average retrieved sources roughly doubled from about 12 to 24, and site: operator usage rose from around 0.3% to roughly 23% of fan-outs (documented on Lily Ray's Substack, lilyraynyc.substack.com).
The gap between Promptwatch's ~17% figure and Konitzny's ~23% is worth flagging rather than glossing over. Both numbers sit well above the historical baseline of 0.3-0.4%, but they don't match exactly, and the likely explanation is methodology. Promptwatch and Peec AI extract fan-outs from the real, consumer-facing ChatGPT interface. Chris Long at Nectiv, by contrast, pulls fan-outs straight from OpenAI's Responses API using a Python script, and API behavior can diverge from what a logged-in user actually sees, since the API skips personalization entirely. If you're comparing two GEO tools and their fan-out numbers don't line up, this is usually why. Neither number is wrong; they're measuring slightly different surfaces.
Why OpenAI might be doing this
Lily Ray floated a theory in her August 17 write-up that I find fairly convincing: OpenAI may be using the site: operator, paired with fan-out refinement, as a cheap way to suppress manipulated or spammy content. Scoping a search to a trusted domain, a .gov site, an established review platform, a recognized brand's own pages, functions a lot like Google's E-E-A-T signals, except it's achieved through query engineering instead of a ranking algorithm.
There's a second pattern worth knowing about, documented in the same research: fan-out queries increasingly name specific brands the user never typed. In one small study (57 conversations, single account, July 2026, weighted toward software categories), 21 of 27 initial fan-out queries already contained brand names, and those named brands got cited 68.9% of the time versus just 2.1% for pages that were merely fetched without being named in the query itself. Only about 3.1% of all retrieved pages made it into the final answer at all. That's a small, geographically narrow sample, so treat it as directional, but it lines up with the broader theme: getting fetched isn't the goal. Getting named in the query, or being the domain a site: search points at, is what actually drives citations.
The practical fallout for your tracking setup
Here's where this stops being interesting trivia and starts being an operational problem. If a meaningful chunk of ChatGPT's retrieval now runs through site:-scoped searches, then a site:yourdomain.com [topic] query can only surface what's actually indexed on your site. That reframes a handful of technical SEO basics as direct AI-visibility levers:
- Thin or orphaned category pages stop being a ranking nuisance and start being an AI-answer blind spot.
- A broken internal search or crawl trap that keeps pages out of Google's index keeps them out of ChatGPT's
site:-scoped retrieval too. - Content that exists on your site but was never properly indexed is functionally invisible to this new retrieval path, no matter how good the content is.
Second, because queries per response nearly doubled alongside the site: jump, there are simply more retrieval opportunities per topic now: comparisons, pricing pages, alternatives roundups, and domain-scoped lookups are each separate chances to get pulled in. A page that only answers one narrow angle of a topic is leaving coverage on the table.
Third, and this is the one people skip: re-run your fan-out analysis regularly. Keyword-to-heading mappings built before August 2026 are stale. Fan-out query length has also been shrinking for months, from around 117 characters in early December down to roughly 53 by April, meaning ChatGPT's sub-queries increasingly look like terse keyword searches rather than full questions. Write headings that read like search queries (