Search Console and AI search: what it shows, what it hides

Search Console is the only free, first-party record of how a search engine treats your site, and for twenty years it was close to a complete one. That changed quietly. A growing share of what it reports now happened inside a generated answer, and nothing in the interface tells you which share.

This is a guide to reading it anyway: which AI activity is in the data, which is absent, and which questions it cannot answer no matter how you filter.

AI Overviews are in your numbers, unlabelled

The single most important fact about Search Console and AI search is that the AI data is already there and is not separated out. Google includes impressions and clicks from AI Overviews — and from AI Mode — in the Performance report under the Web search type, aggregated with ordinary blue-link results. There is no filter, no breakdown, no dimension, and no API field that splits them.

So the honest summary of what your Performance report shows is: every appearance in Google Search, whether that appearance was a ranked link, a citation inside a generated answer, or both on the same query, summed.

That has an immediate consequence for how the report reads. Two very different events now produce the same row. A query where you ranked fourth and were clicked looks like a query where you were named in an AI Overview and the reader got their answer without leaving. Same impression, different outcome, indistinguishable in the data.

What position means now

Average position was always a weaker number than people treated it as. For AI Overviews it changes meaning outright: a link inside a generated answer is reported at the position of the overview block, not at a rank within the block’s own list of sources. Being the first source cited and the fourth source cited are, as far as this column is concerned, the same.

This is worth saying plainly because position is the metric most often used as a proxy for visibility, and it has become the metric least able to carry that weight. If your average position improved while clicks did not, one available explanation is that more of your impressions are now inside overview blocks near the top of the page rather than links further down it.

The pattern to look for

You cannot isolate AI Overview data, but you can look for its fingerprint: impressions flat or rising while clicks and click-through rate fall, on informational queries.

Compare a recent period against the same length a year earlier, group by query, and sort by the largest drop in CTR among queries whose impressions held up. Question-shaped queries — how, what, why, best, vs — will be concentrated at the top of that list, because those are the ones generated answers are most often triggered on. Navigational and branded queries will mostly be untouched.

That is a correlation and it is open to other readings: a competitor may have moved above you, the SERP may have grown a new feature, seasonality exists. Treat it as the question rather than the conclusion. It is still the most useful cut of this data available to you.

What Search Console does not cover at all

Everything that is not Google. There is no Search Console for ChatGPT, for Perplexity, for Claude, or for Copilot. No impression count, no citation list, no query data, no verification process. For the engines that a meaningful share of research questions now go to, the first-party channel does not exist.

What you get instead is downstream and partial:

• Referral traffic in your analytics, when an assistant links you and someone clicks. This undercounts badly, since the whole point of a generated answer is that it often needs no click, and it tells you nothing about the answers where you were named and nobody followed the link.
• Crawler hits in your server logs, which show that an engine fetched a page but not what it said about you afterwards.
• Asking the engines directly, on a schedule, and recording the answers. This is the only method that observes the thing itself, and it is what measuring AI visibility actually means.

A tool that claims to report your AI search visibility from Search Console data alone is reporting Google, and only the part of Google that resulted in a page view.

Crawl stats show Google’s fetchers and nobody else’s

The Crawl stats report under Settings is genuinely useful — response codes, fetch purpose, file types, host status over time — and it is entirely about Google’s own crawlers. It says nothing about OAI-SearchBot, PerplexityBot or ClaudeBot, because Google has no reason to know they visited.

So if the question is whether answer engines can reach your site, Search Console can only answer a third of it. Your robots.txt decides who is allowed — see which crawlers to allow and block — and your server logs are the only record of who actually came. A sudden spike in 5xx responses or a host status warning in Crawl stats is still worth acting on immediately, since a site Google struggles to fetch is usually one everybody struggles to fetch.

Setting it up so the data is usable

Pick the domain property

Search Console offers two property types and the choice is not cosmetic. A URL-prefix property covers exactly one protocol, one host and one path prefix: https://www.example.com/ and https://example.com/ are two separate properties with two separate datasets. A domain property — written sc-domain:example.com — covers the domain, its subdomains, and both protocols in one place.

Take the domain property. It costs one DNS TXT record and it removes an entire class of confusion where half your traffic is in a property you forgot you had. This matters more than it sounds if your site redirects between hosts, which is the failure described in the canonical tag that points at a redirect: a URL-prefix property on the host that redirects away will look mysteriously empty, and the explanation is that you are verified on the wrong side of a 308.

Expect a lag, and build for it

Performance data is not live. The most recent two to three days are still being assembled, so a query that ends today reliably returns fewer rows than it should, and a query for today returns none.

We hit this while building our own Search Console integration, and it is worth naming because of how it presents. An empty result from a valid credential against a correctly verified property is indistinguishable, from the caller’s side, from a broken integration. Our check script therefore ends its date range three days back and says in the output when zero rows is a real answer rather than a failure. If you automate anything against this API, do the same — lag it by three days and make the empty case say so.

The other half of that lesson: check each step separately. A Search Console integration can fail with a valid token, a 200 response and no data, simply because the service account was never added to the property under Users and permissions. Authentication succeeding is not the same as access existing.

The three cuts worth your time

Most Search Console use is aimless. These are the views that repay the visit.

Queries where impressions rose and CTR fell. The closest thing available to an AI Overview report, described above. Reviewed quarterly rather than weekly, since the comparison needs enough data to mean anything.

Pages ranking between positions five and fifteen on question-shaped queries. These are the pages close enough to be considered as a source and not quite being chosen. They are the best candidates for the editorial work in making a passage quotable, because the engine has already decided the page is relevant.

Pages with impressions and zero clicks, ever. Usually one of three things: a title that does not match the query’s intent, a page answering a question fully in the snippet, or a page that should not be indexed. All three are worth knowing, and the third is worth fixing first.

What cannot be known from this data

A short, honest list, because the gaps are where people invent numbers.

• How often you appeared in an AI Overview. Folded into Web impressions, unrecoverable.
• Whether a given impression was a link or a citation inside an answer.
• What any engine said about you. Search Console reports appearances, never content — and whether an answer was favourable is a different measurement entirely.
• Anything at all about ChatGPT, Perplexity, Claude or Copilot.
• Which of your competitors were cited alongside you, which is the comparison that makes a visibility number mean anything. See competitive analysis for AI search.

None of that makes Search Console less worth having. It is first-party, it is free, it covers the largest single source of search demand, and its crawl and indexing reports catch problems nothing else will. It is simply no longer a complete picture of how machines find you, and the part it is missing is the part growing fastest.

For scale: our own domain property, four weeks after launch, reported a single click against three impressions. That number is honest and it is also nearly useless — it is too small to read a trend from, and it says nothing about the three engines not represented in it. A new site’s first evidence of AI visibility almost never arrives through this channel.