What an Agent Should Do When ChatGPT Adopts the site: Operator
Promptwatch fanout data shows ChatGPT Search adopting the site: operator overnight on August 8, 2026, with searches per response nearly doubling. The agent move is to refresh the prompt to keyword map.
On August 8, 2026, ChatGPT Search changed how it searches the web overnight. Fanout queries using the site: operator, searches scoped to a specific domain, jumped from 0.37 percent to 16.8 percent of all fanout queries in a single day. At the same time the average number of searches per response nearly doubled.
The site operator fanouts report and the query fanouts report, published August 10 and March 1, 2026, track the shift. For anyone who runs an agent on their AI visibility, the practical question is what an agent should do automatically when the queries ChatGPT generates from a prompt change overnight.
The signal
Promptwatch's site operator report covers the percentage of all ChatGPT Search fanout queries that contain the site: operator, per day. The share hovered between 0.3 percent and 0.5 percent for weeks. It dipped briefly to 0.15 percent on August 3 to 5, which is consistent with a staged rollout or a pre launch experiment. Then it jumped to 16 to 17 percent on August 8. That is roughly a 46x increase in share within a single day.
The query fanouts report covers the average number of fanout queries per ChatGPT Search response, per day. On August 8 the average jumped from about 1.08 to about 1.83, the same day the site: usage surged. The timing is the key signal. The new domain scoped searches come on top of the generic web searches ChatGPT was already running, they do not replace them.
What an agent should do automatically
The first move is to re run your fanout analysis for your tracked prompts and refresh the prompt to keyword map. The queries ChatGPT generates from a prompt changed materially on August 8. Keyword mappings built before the shift are stale. An agent that maps prompts to keywords should re run the map on a schedule, not once, and flag prompts where the fanout depth moved after August 8.
The second move is to treat your own domain as a retrieval surface and make sure your most important pages are crawlable and indexed. A site: scoped query can only surface what search engines have indexed from your domain, so gaps in coverage translate directly into gaps in AI answers. An agent that watches visibility should flag pages that are not indexed as a gap to fix.
The third move is to connect the fanout shift to the citation. A fanout is a leading indicator. The queries ChatGPT runs today are the citations it returns tomorrow. An agent should log the site: queries ChatGPT runs for your prompts and flag the prompts where a site: query targets a competitor, so you can build the page that earns the citation before the next check.
How to do it with Promptwatch
The measurement that turns the population report into a prompt level action is the query fanout tracking in Promptwatch. It tracks query fanout volume, average queries per response, and query length trends over time, based on ChatGPT Search data.
An agent that runs on visibility should wire the fanout report into a Unified Action. When the average searches per response for your prompts crosses a threshold, the action is to re run the prompt to keyword map and flag prompts where the fanout depth moved. That is the automated move that turns a retrieval shift into a fix, not a dashboard you read and forget.
What the report does not tell you
The reports are population level views. They tell you what ChatGPT is doing across the whole web. They do not tell you what ChatGPT is doing for your prompts. For that you need your own fanout data joined to your own visibility data. ChatGPT can adopt the site: operator at the population level and never run a site: query for your prompts, or run them for your competitors.
The report also does not say why ChatGPT adopted the site: operator. A jump that fast, in lockstep across the fanout behavior, points to a model or system prompt rollout on OpenAI's side rather than a gradual behavior change. Platform controlled search behavior can change overnight, which is the part that makes one off audits unreliable and continuous monitoring necessary.
The broader pattern
The site: shift is one instance of a wider change in how AI search retrieves information. The average searches per response nearly doubling at the same time is the other half of the story. ChatGPT now searches both wider, more queries per answer, and deeper, queries scoped to specific domains, than it did before.
For an agent that runs on visibility, that means the retrieval surface is getting more granular. A year ago, the work was to be present in the open web results ChatGPT returned. Now the work is to be present in the open web results and in the domain scoped results. The practical response is to treat fanout depth as a first class signal, not an afterthought.
How to read a shift versus a drift
The shape of the change matters for how an agent acts on it. A share that jumps in a single day is consistent with a rollout, which is a platform decision you watch and adapt to. A share that drifts over weeks is consistent with a gradual behavior change, which is a trend you build against. The two shapes call for different responses, and an agent that treats them the same will overcorrect on a drift and underreact on a shift.
The site: data is the shift shape. The share moved in a single day, and the average searches per response moved with it. That is the kind of change an agent should flag immediately and re run the prompt to keyword map against, because the queries ChatGPT runs today are not the queries it ran a week ago. A shift is a reason to refresh the map now, not next month.
What to watch next
The reports are snapshots through August 17, 2026 for the site operator report and ongoing for the query fanouts report. The open questions are whether the site: share holds, whether the average searches per response keeps climbing, and whether the new domain scoped queries translate into different citation patterns. Those are exactly the questions a fanout view answers for your own prompts.
The dataset Promptwatch publishes is aggregated and non identifiable, and it is refreshed constantly. That makes it useful for spotting population level shifts. It does not replace the need to track your own brand, your own prompts, and your own fanout depth. The two work together: the public report tells you what is happening in the field, and your own tracking tells you whether it is happening to you.