Key findings
- The share of site: queries grew about 9-fold from summer 2025 to summer 2026.
- Just over a third of the websites studied had such searches. On about 5 out of 100, searches target the site’s own domain, up from about one a year earlier.
- Almost nobody clicks: the click-through rate is below 0.2%. That speaks against people as the senders.
- The pattern fits measurements showing that ChatGPT has searched with site: far more often since 2026. Search Console cannot prove who sent them, though.
What site: queries are
With “site:” you search only one website. Typing site:beispiel.at öffnungszeiten into Google returns results from beispiel.at only. People rarely use this operator, mostly specialists checking something.
For an AI system, though, it is very handy. Instead of searching the whole web and then checking which source is right, it can look directly where the reliable answer should be: with the manufacturer, the authority or the provider itself.
Where ChatGPT gets its information
ChatGPT answers many questions with its own web search and rewrites the question into targeted queries for it. That is how OpenAI describes it on its official help page on ChatGPT search: the question is turned into one or more specific searches sent to search partners. After the first results, ChatGPT may follow up with further, more specific searches.
These background searches are called fan-out queries. They decide which websites ChatGPT gets to see at all. The sources linked in the final answer come from exactly these searches.
Why there are suddenly so many site: searches
Since 2026, ChatGPT has narrowed its fan-out searches with site: more and more often. SEO expert Lily Ray summarised the measurements so far:
Lily Ray suspects ChatGPT uses this to avoid low-quality sources: for facts and prices it checks with the provider, for health and legal questions with official bodies. She stresses that attributing these searches in Search Console is an assumption. So we looked at real data.
What our data shows
Until the end of 2025, site: queries were rare. From spring 2026 they rise sharply. In summer 2026 there were about 66 site: queries per 100,000 queries, a year earlier about 8.
site: queries grew about 9-fold in one year
site: queries per 100,000 queries of three words or more, per quarter. Search Console data from 300+ websites, focus on Austria.
Values as a table
| Quarter | per 100,000 queries | websites with site: | site: on own domain |
|---|---|---|---|
| Q3 2025 (Jul to Sep 25) | 7.6 | 10% | 1% |
| Q4 2025 (Oct to Dec 25) | 6.1 | 14% | 0.3% |
| Q1 2026 (Jan to Mar 26) | 14.7 | 17% | 2% |
| Q2 2026 (Apr to Jun 26) | 41.8 | 22% | 5.6% |
| Q3 2026 (Jul to Sep 26) | 66.1 | 25% | 4.8% |
In absolute terms these are few searches. What matters is the trend: it starts exactly when ChatGPT switched to this way of searching according to the measurements, and it does not only affect large websites.
More and more websites see such searches
Out of 100 websites, this many had at least one site: query in the quarter.
Jul to Sep 25
10 out of 100
Jul to Sep 26
25 out of 100
Targeted searches on the site’s own domain
Share of websites whose own domain was searched with site:, per quarter.
Two patterns, two senders
Not every site: search is the same. The data shows two very different forms.
One domain plus a specific term, often with “official”. For facts, prices and terms the system checks directly with the provider.
- site:beispiel-shop.at rückgabe frist (return period)
- site:musterbau.at förderung wärmepumpe official (heat pump subsidy)
- preise jahreskarte site:beispiel-museum.at (annual pass prices)
A name in quotes, keywords and long chains of “-site:”. This looks more like tools that monitor companies and news.
- "beispielfirma" übernahme OR kooperation -site:newsportal-a.at -site:…
- "musterverein" kritik -site:musterverein.at -site:…
The examples are made up and only show the form. They are not from the data and do not name real websites. German search terms are translated in brackets.
What Search Console does not tell you
Search Console shows what was searched, but not who searched. That part of these searches comes from ChatGPT is plausible: the timing fits the measurements, the click-through rate is close to zero, and words like “official” appear. Another part very likely comes from tools that monitor companies, or from SEO tools.
What you can do now
site: searches on your domain are a rare opportunity: you see which facts an AI system is looking for on your site. Use it.
- 01Check Search ConsolePerformance, search results, filter “Query”, “Queries containing” and enter
site:. The regular expression^(?!.*-site:).*site:shows only targeted searches without exclusion chains. Set the period to 16 months (help on the performance report). - 02Read what is searched on your domainThe terms after
site:your-domain.comshow which facts an AI system looks for: prices, deadlines, opening hours, technical data, subsidies. Is the answer easy to find on your website? - 03Provide facts as readable textPrices, dates and terms belong on the page as normal text, not only in images, PDFs or content loaded later by JavaScript. A short, clear answer at the top helps people and machines.
- 04Allow OAI-SearchBotCheck robots.txt, firewall and host: if the ChatGPT search crawler is blocked, your website cannot appear there as a source, however good the content is.
- 05Make clear which domain is the official oneAI systems guess the domain when it is unclear and are sometimes wrong. The same company name and domain on all profiles, in directories and in the imprint help.
- 06Read the numbers correctlysite: queries bring impressions but hardly any clicks. Without filtering them out you see a falling click-through rate that has nothing to do with real visitors. Exclude them from reports and track them separately.
How we measured
We counted all queries containing the “site:” operator and related them to all queries in the same period. That way it does not matter that some websites have more data than others.
- Counted: every query with “site:” in a month on a website. The value per 100,000 refers to all queries of three words or more in the same quarter.
- Own domain: if the domain after “site:” is the domain of the website whose data we analyse, it counts as a targeted search on the own website.
- Cross-check: without the five websites with the most site: searches, the value also rises many times over.
- Privacy: only rounded totals and shares across many websites. No website is identifiable, the examples are made up.
