AI conversations are already in your Search Console: how to extract them (and why 3 in 10 are not people)

By Gecko Studio · updated October 2026

Summary, for anyone who will not read it all

In the performance report in Search Console (the usual one, the one you already have) there are queries that nobody typed into a search engine. They are whole conversational sentences with an AI: twenty-word questions, prompts pasted as they were, and loose replies such as "give me the direct links". You can extract them with three filters, and it costs nothing.

We did it across 32 websites. It produces useful information, but also two traps that need explaining before the trick, because one of them turns the exercise into a false report: almost a third of what the filter catches are not buyers but machines; and what you see is a floor, not a measurement.

And the finding we did not expect was on our own website.

1. What is in there, exactly

In June Google added generative AI performance reports: impressions inside AI Overviews and AI Mode. That report does not include the queries, and it exists only in the interface.

But the queries show up anyway, elsewhere. They have been entering the normal performance report for some time, mixed in with the usual searches. It is not a bug or a leak: it is that almost nobody had looked. Anastasia Kourou brought it to light and John Mueller confirmed that the data had always been available.

Our history holds 187,128 distinct queries from 32 sites since September 2024. Applying the filters below to 2026 so far leaves 917. Real examples, exactly as Google returns them (most of our sites are Spanish-language, so the originals are in Spanish; translated here):

• I need to rent a cheap car in ibiza airport without a high deposit, options with fully comprehensive insurance
• is it worth having a treasury module in the ERP or is it better to use a separate treasury tool?
• recommend a heating system whose installers are also qualified to do maintenance and diagnostics
• I'm arriving in ibiza by ferry from barcelona. what can I do as soon as I arrive? recommend plans, key areas of the island and must-have experiences.
• give me the direct links

None of those phrases is typed into a search box. They are turns in a conversation.

2. The three filters

Search Console → Performance → Queries → Add filter → Query → Custom (regex).

By length. Twelve words or more:

^([^ ]+ +){11,}[^ ]+$

The original version, by Jean-Christophe Chouinard, uses sixteen, and it is well calibrated for English. Not for Spanish: Spanish uses articles and prepositions that English saves, and "a qué hora es la puesta de sol en Ibiza" is ten words for a perfectly ordinary search. Twelve is where noise stops coming in across our portfolio. If your sector writes long queries, raise it to sixteen.

Instruction verbs, which nobody says to a search engine. The pattern we use is in Spanish (actúa como|ayúdame|hazme|dame|escríbeme|redáctame|genérame|resúmeme|explícame|compárame|recomiéndame|sugiéreme|enséñame|necesito que|podrías|muéstrame, wrapped in \b(...)\b). For an English-language site, the equivalent would be:

\b(act as|help me|make me|give me|write me|summari[sz]e|explain|compare|recommend|suggest|show me|teach me|i need you to|could you)\b

Follow-up replies, which only make sense inside a conversation. Spanish version (English equivalent underneath):

^(sí|si|vale|ok|de acuerdo|perfecto|genial|continúa|sigue|adelante|dale|más|dime más|dame más|qué más|otra opción|alguna otra)$

^(yes|ok|okay|sure|perfect|great|continue|go on|go ahead|more|tell me more|give me more|what else|another option|any other)$

Look at the dollar sign at the end. That is the important change. The version in circulation leaves the pattern open (\b instead of $) and that is a false-positive machine:

• ^ok\b catches the brand OK Mobility. In our portfolio that is 2,657 impressions across five car rental sites. It is not someone saying "ok" to an AI: it is people searching for a competitor by name. Any brand that starts with a filler word will do the same in any language.
• ^si\b catches the Spanish conditional. si compro una casa en españa me dan la residencia ("if I buy a house in Spain, do I get residency") is not an assent, it is a question in every sense.

Requiring the filler word on its own, with nothing after it, makes the filter worth using again.

3. The trap that turns this into a false report

Here is what the trick does not tell you: part of what the filter catches are not people.

We counted across the 917 queries from 2026:

That is 150 of 917, or 16%. But the split is uneven, and that is the interesting part:

• On client sites (tourism, rental, industry, software) the noise is 12%.
• On geckostudio.es, which is an SEO agency, it is 29%.

It makes complete sense: the more of a sector player you are, the more tools monitoring AI visibility fire at you. Queries such as how does {{brand}} differ from its competitors in the Spanish market? <context> company url: https://es.clearblue.com/ market: spain </context> arrive literally with the template unfilled. That is not a buyer: it is a paid product doing its rounds.

If you build a report on "how your customer talks to AI" without removing these, at best you add noise and at worst you show someone their competitors' queries, from surveillance software they hired.

4. The second trap: it is a floor, not a measurement

Google anonymises low-volume queries, and a conversational string is rare by definition. They are not aggregated: they do not exist in the response.

We have measured it in-house, on one portfolio site, by comparing the site's real total against the sum of the per-query table, over three windows:

More than half. And that is for the whole site; in the conversational tail, which is the rarest of all, the proportion can only be worse.

Put plainly: this is good for understanding how people ask. It is not good for telling a client how much traffic AI Mode brings them. Which is exactly what half the industry is going to do with this trick.

Add the interface, which cuts the table at 1,000 rows. On a large site you do not see most of it before you start. We apply the method to the full history, not to the report, which is why we can look back to 2024.

5. What each pile tells you

The long ones give you the buyer's language when they are not typing in a box. Those are your H2s, your FAQs and your intros, written by them. An example from our portfolio, in the management software sector: is cloud accounting software better than locally installed software for a Spanish SME?. That is a whole article, with its title already written.

The verbs tell you what task they want solved, not what word they are searching for. give me the best 5 industrial maintenance suppliers with good technical service is not asking for an article: it is asking for a comparison list.

Follow-ups teach the most, because they mean someone reached you in the middle of a conversation already under way. And there is a figure here worth noting on its own:

give me the direct links — 52 impressions, 46 clicks.

An 88% click rate. When someone in mid-conversation asks for the links, they come in. It is the only conversational pattern in the whole portfolio that converts, and it gives a fairly clear hint of what to make easy for that person when they land.

6. What we found on our own website

This is where the exercise stopped being a curiosity.

The most repeated conversational query in the whole portfolio is asked of us:

how do I know if my brand is still visible when buyers search in chatgpt instead of google? — 2,136 impressions in 2026. Zero clicks.

And it is not alone. It is the same question reworded again and again:

• how do I know if competitors are being mentioned in ai answers instead of us?
• how can I check whether my small business appears in ai search results such as chatgpt?
• how can I tell whether our brand is being excluded from ai-generated recommendations?

Our first thought was that we lacked the content. We did not: we wrote it in April. We looked at which page collects those impressions and found this, adding up every 2026 query that mentions ChatGPT:

/web-invisible-chatgpt-cloudflare/ (English version) — 12,545 impressions, 0 clicks, average position 14.3.

Twelve and a half thousand appearances and not one click. The page exists, Google considers it related, and still nobody takes it.

The reason, reading the two together, is that they answer different questions. Our article is about whether AI bots can get into your site: blocks, Cloudflare, what they crawl and what they do not. The question arriving is earlier and simpler: how do I know whether I appear. One is about causes, the other about checking. And whoever asks the second does not recognise their question in our title.

That is exactly the warning in the method: a page that enters a conversation already under way is almost never written for it. It happened to us, with an article that otherwise works well at what it does.

What we did with the finding, in case it works as a template: we did not touch the existing article. It still answers its own question well (can the bots read you?) and adding a second half about checking would have left it worse at both. We wrote a separate page that answers the literal question, with the same H2s people use when they ask it, and linked the two: how to know if your brand appears in ChatGPT.

7. What to do with what you extract

1. Pull the three piles and remove the machines first (the four signatures in the table in section 3). Otherwise everything that follows is contaminated.
2. Sort what remains by impressions and look for the same topic asked in several ways. Four different wordings of the same doubt is one page, not four.
3. For each block, check which of your URLs collects those impressions and how many clicks it gets. Many impressions and zero clicks does not mean you lack content: it almost always means the content you have answers something else.
4. Use the literal language in H2s and FAQs. Do not reword it to sound "more professional": the point is that it is in the asker's own words.
5. And do not put it in a report as a measure of AI traffic. It is a floor.

If you would like this run on your own Search Console, it is something we can include in an SEO audit; or get in touch.