AI Search Prompts in Google Search Console (2026)

Somewhere in the last year, “check your AI prompts in Search Console” became standard advice. You will find it in newsletters, conference decks, LinkedIn posts and at least one Chrome extension built specifically to do it. The pitch is irresistible. Google is quietly logging the questions people ask AI assistants, and if you paste the right regex into the query filter, you can read them.

Then in June 2026 Google shipped an official AI report in Search Console. Which should have settled the question. Instead it made the gap obvious, because the report Google built gives you the one thing the hack could never give you, and withholds the one thing the hack was famous for.

I run this on a real site every week. Below is what each source genuinely contains, where the “see your prompts” trick came from, and a second fingerprint sitting in my own Search Console that almost nobody talks about. The short version is that you cannot read AI prompts in Search Console, and it matters less than you would think.

Can you see AI search prompts in Google Search Console?

No. Not in the official report, and not reliably anywhere else.

This is worth stating flatly because the whole category of advice around it is built on a softer claim that quietly hardens as it gets repeated. “These queries look like prompts” becomes “these are prompts” becomes “here are your AI prompts” by the third retelling. The first version is true and useful. The third is not something Google’s data can support.

The generative AI performance report is a Search Console report that counts how often links to your site appeared inside Google’s generative AI features, broken down by page, country, device and date, with no query data attached. That last clause is the whole story, and it is Google’s design choice rather than an oversight.

What does Google’s generative AI performance report show?

Google announced it on the Search Central blog on June 3, 2026, and rolled it out to a subset of properties first, starting in the United Kingdom before widening. It covers AI Overviews and AI Mode on Search, plus generative features in Discover.

Here is the entire surface, taken from Google’s own Search Console help documentation.

Four dimensions and one metric. That is the report.

The omissions are not accidental. Search Engine Land’s coverage of the launch put the click question plainly, noting that Google will not be telling site owners how many searchers click from AI responses through to their sites. Google’s line was that it is “continuing to work with website owners to understand what insights will be most helpful” and will add metrics “over time”. That is a roadmap sentence, not a commitment.

So the official answer to “how am I doing in AI search” is now a number of appearances, and nothing about what those appearances were worth or what question triggered them.

Where did the “see your AI prompts” trick come from?

It came from a genuine leak, and the origin explains why the technique feels so much more authoritative than it is.

In late 2025, analytics consultant Jason Packer documented something strange in Search Console query reports. Private ChatGPT prompts, some over 300 characters, were appearing as search queries on sites that had nothing to do with them. Not paraphrased. The actual text people had typed into a chatbot, including the sort of thing you would only type somewhere you believed was private.

The cause was mundane and specific. A hidden parameter in some ChatGPT sessions triggered a real-time web search, and the prompt text rode along inside URLs that Google’s systems then scanned and logged into site owners’ reports. As TechSpot reported at the time, OpenAI acknowledged the problem and said it had fixed an error that affected a small number of queries. Packer’s point was sharper than the bug itself. Nobody clicked share. The prompts were simply misrouted.

That is the seed. For a window, AI prompts really were visible in Search Console, verbatim, by accident. The technique that grew out of it survived the fix.

Does the 10-word regex actually find AI prompts?

It finds long queries. Long queries correlate with AI use. Those are two different statements, and the distance between them is where people go wrong.

The pattern most commonly shared traces to Chris Long, writing in Search Engine Land in February 2026. You open Performance, filter on Query, choose Custom Regex, and paste this:

^(?:\S+\s+){9,}\S+$

That matches any query of ten or more words. The reasoning is good. People type keywords into search boxes and sentences into chatbots, so unusually wordy queries skew toward prompt-shaped behavior. Long himself framed it as a way to surface queries that read like prompts, not as a label Google applies.

The limits are real and worth holding onto:

• Google does not tag query origin: There is no field anywhere in Search Console that says a query came from an assistant. Every method is inference.
• Plenty of humans type long queries: Voice search, panicked troubleshooting and anyone who has learned that specific searches work better all produce ten-word queries with no AI involved.
• The anonymization threshold hides the tail: Search Console omits rare queries entirely, and prompt-shaped queries are rare almost by definition, so the ones you can see are a biased sample of a sample.

Use it. It is genuinely interesting. Just do not build a report for a client that presents the output as a prompt count, because it is a proxy wearing a label it did not earn.

The fingerprint nobody talks about: search operators

Here is the part I have not seen covered anywhere, and it is sitting in my own Search Console.

Filter your queries for -site: instead of filtering for length. On contextbolt.com, that returns rows like this one, which pulled 11 impressions at an average position of 4.7 in a recent 28-day window:

"ai answer engines ai user agents" -site:reddit.com -site:twitter.com
-site:x.com -site:wykop.pl -site:tripadvisor.com -site:youtube.com
-site:yelp.com -site:booking.com -site:facebook.com -site:instagram.com
-site:tiktok.com

And this one, 12 impressions at position 5.25:

"cline" "ai" "claude" "vs code" "agent" "coding" "developer"
-"airline" -"baseline" -"deadline" -"online" -"pipeline" -"timeline"

Look at what the second query is doing. It is forcing seven exact-match terms, then excluding six words that merely contain the letters “cline” so the engine cannot drift onto airlines and deadlines. That is token disambiguation. No human has ever sat down and typed six negative exact-match exclusions to stop a four-letter product name colliding with “timeline”.

Both queries share a shape that the word-count method misses entirely:

• Quoted phrases: Forcing exact match on multiple terms at once, which is how a machine constrains a result set it intends to parse.
• Long exclusion chains: The same block of social and aggregator domains stripped out, because the caller wants primary sources rather than forum threads.
• No natural language at all: Some of these are under ten words and would never trip the regex, despite being obviously automated.

Two honest caveats, because this is inference too. Google does not tell you who issued these, so “automated” is the safe claim and “ChatGPT specifically” is not. And SEO tools, rank trackers and researchers also use operators, so a share of these are people with a search-operator habit rather than agents. What makes the pattern interesting is not certainty about the sender. It is that the queries are structurally impossible to have been typed casually, and they are ranking you at position 4 to 5 on questions nobody in your keyword research would have found.

Which source should you use for what?

The two methods answer different questions, and most confusion is people using one to answer the other’s question.

That “neither” row is the important one. If somebody hands you an AI visibility report with a value attached, ask which field it came from, because Google does not publish one. Our own AI Overview click loss numbers came from measuring the gap on the normal performance report over time, not from any AI-specific metric, and that remains the only honest route to a value estimate.

How to read your AI search data without lying to yourself

Four steps, about twenty minutes, repeatable monthly.

1. Check whether the report exists for you yet: Open Search Console and look for the generative AI section in Performance. Rollout is staged, so its absence means you are not in the wave yet, not that you have no AI visibility.

2. Pull the pages dimension first: Impressions sitewide is a vanity number. The pages breakdown tells you which of your URLs Google’s AI features actually reach for, and that is directly actionable, because it names your citation-worthy content.

3. Run both query filters and label them honestly: The ten-word regex for prompt-shaped queries, and a -site: contains-filter for operator-shaped ones. Record them as “long queries” and “operator queries”, not as “AI prompts”. Future you will thank present you.

4. Compare against your normal performance report: Same pages, same window. If a page is climbing in AI impressions while its ordinary clicks flatten, that is the trade being made in public, and it is the single most useful comparison available right now.

The reason this stays manual for most people is that Search Console’s interface makes cross-referencing two reports over a rolling window genuinely tedious. This is the job we built Search Console access for our agent around, so you can ask for the comparison in one sentence instead of exporting two CSVs and reconciling them in a spreadsheet.

Should you use the AI content blocking toggle?

Google shipped a control alongside the report that lets you keep your content out of AI Overviews, AI Mode and the generative features in Discover. It does not harm your core Search rankings, which is the reassurance most people are looking for.

Read the trade before you touch it. Google’s own framing is that sites which opt out will not receive traffic or impressions from those generative features. So you are not removing yourself from the summary while staying in the results. You are removing yourself from the surface entirely, including the citation link that sends people to you.

For a small site, that is almost always the wrong trade. The publishers with a real case for opting out are the ones whose entire product is the answer itself, where a summary genuinely substitutes for the visit. If your pages exist to bring someone into a product, being named in the answer is the point. This is the same calculation that runs through the AI crawlers list, and the mistake in both places is treating a visibility question as a copyright one.

What this means for how you measure AI search

Step back and the picture is clearer than the discourse suggests.

Google has decided that site owners get to know whether they appear in AI answers, and not what was asked or what it earned. That is a deliberate settlement, and it is unlikely to move much, because the prompt text is the most commercially valuable data in the entire stack and the leak of 2025 showed exactly how badly it goes when it escapes.

Which makes the practical answer less about extraction and more about coverage. You cannot reverse-engineer the questions, so you write pages that answer questions completely and are cheap for a machine to lift. That is the whole of answer engine optimization and it has not changed because a new report shipped.

The one thing I would change in most people’s routine is the operator filter. It takes ten seconds, it is structural rather than probabilistic, and on my own property it surfaced queries at position 4 that no keyword tool would ever have suggested, on phrasings I would never have written. That is a content brief arriving in your inbox with no research budget attached. The prompts are not coming. This is the closest thing available, and almost nobody is looking at it.

What to do this week

Open Search Console and check whether the generative AI report has reached your property. If it has, pull the pages dimension and write down the three URLs Google’s AI features reach for most, because those are your proven citation assets and they deserve better internal links than they currently have.

Then run the operator filter. Search your queries for -site: and read whatever comes back. On most sites it will be a handful of rows. Read them anyway. They are the closest thing to a machine telling you, in its own syntax, exactly what it was trying to find when it landed on you.