ChatGPT is now using the site: operator to search specific domains: what does that mean for SEO?
Our agency's take on ChatGPT running site: searches against specific domains, what it changes for SEO work, and how we check it with real fan-out data.
When a client forwards us a screenshot saying "ChatGPT is now using site: to search specific domains, should we panic?", our answer is always the same. Do not panic, and do not reorganize a site on the strength of one report. Check what ChatGPT does for your prompts, then decide.
We have not independently confirmed how often ChatGPT does this, or what index is behind it, and we would be wary of anyone who claims to know. What we can do is explain what it would mean if true, and show how we test it.
What the operator implies
A site: query limits results to one domain. If ChatGPT writes one as part of the searches behind an answer, it has already decided where to look. The decision may come from the brand named in the prompt, from a source it saw earlier in the same answer, or from what it already knows about who is authoritative on a topic.
Two consequences follow. First, being a domain the model thinks of is now part of the game, which is mostly built off-site through mentions and citations. Second, once the model is inside your domain, the result depends entirely on how well your own pages answer the narrow query. Competitors are not in that result set.
What we change on client sites, and what we do not
Nothing about this makes us abandon the basics. If anything it makes the basics more visible.
- Page titles and headings that say what the page answers. A narrow scoped query rewards a clear page over a clever one.
- One page per question. If three pages half-answer the same thing, a scoped search may surface the wrong one.
- Crawler access. We check robots.txt and server logs for the real bots: GPTBot, OAI-SearchBot and ChatGPT-User for OpenAI, plus ClaudeBot and PerplexityBot for other engines. Blocking the wrong one by accident is more common than people expect.
- Third-party presence. If fan-outs scope to review, comparison or directory domains, we want the client present and accurate there.
What we do not do is rewrite site architecture around a behavior we have seen in a handful of screenshots.
How we test it
We use Promptwatch because it exposes the pieces in one place, and this check is a good example.
- Choose the prompts. Fifteen to twenty that matter to the client's pipeline, covering brand, category and comparison questions.
- Pull the fan-outs. Promptwatch tracks query fan-outs per model. For ad hoc checks the free GEO Inspector Chrome extension shows ChatGPT fan-outs and the difference between sources fetched and sources cited, and it runs locally in the browser.
- Mark the scoped ones. We note any
site:query and the domain it names. - Compare against citations. The citation analytics show which domains and pages are actually cited, and Page Tracker follows specific URLs, ours or a competitor's.
- Confirm access. Agent Analytics provides IP-verified AI crawler logs, so spoofed user agents are filtered out. We look for ChatGPT-User and OAI-SearchBot hitting the pages that should answer, and read the crawl-to-citation path and citation rate per page. Crawler logs begin on the Professional plan.
The result is a table of prompts, scoped domains and whether the client appears. It takes an afternoon, and it replaces a debate with a list.
Reading the results
Three outcomes come up.
The operator names the client's domain. We make sure the target page is indexed, well titled and fetched by the ChatGPT bots, and that nothing on it is hidden behind scripts or gates.
The operator names someone else's domain. We treat it as a placement target. That could be a trade publication, a review site or a community page, depending on the category.
The operator never shows up. Good. It means this is not a factor for that client yet, and we put the time into content and citation work instead. We re-run the check later, because model behavior changes.
A note on measuring the effect
After changes, we track visibility score and share of voice for the same prompt set, and the AI-referred visits and conversions from Promptwatch's visitor analytics. Position in answer matters as much as presence, so we watch it too. Be careful about causation. If a client's visits rise the same month ChatGPT changes how it searches, a single change on our side cannot take the credit. We say so in the report.
No tracking tool can guarantee a recommendation, and we do not promise one. What we can promise is that the client will know what the model actually searched for, which pages it fetched and whether those pages got cited. That is a better foundation than any hot take about operators, and it is the reason we run client programs on a platform that records all of it.