How to choose AI tracking prompts with Search Console?

To choose your AI tracking prompts with Search Console, filter your question-style, long and recommendation queries, group them by intent, then rewrite each one as a real question with a context, a constraint and a request for a choice. Your Google queries are not prompts. They reveal your market’s intents, not the way your customers talk to ChatGPT.

I see the same mistake often in audits. A set copy-pasted from Search Console, a clean visibility score, and a measurement that says nothing about what AI engines actually recommend.

The topic is getting urgent. Since August 31, 2026, Google has rolled out its Search Generative AI performance report in Search Console to every website worldwide¹. The report counts your impressions in AI Overviews and AI Mode, but it returns no queries. The Performance report remains your only source of real phrasing.

Here is the method I use with clients: the regex filters, the three transformations and how to balance the set.

Why aren’t your Search Console queries prompts?

A Google query and a question asked to an AI describe two different behaviors. On Google, people type three or four words: “real estate agency nice”. In ChatGPT, they write a full sentence with their situation and their limits: “I want to sell my two-bedroom flat in Nice before summer, which agency would you recommend?”

These two phrasings produce different answers. The short query triggers a generic answer, often a neutral list with no ranking. The long question forces the AI to sort, recommend and justify its choice. That second case is where your brand wins or loses a citation.

Search Console does not give you your prompts. It gives you something better: the real intents of your market, measured on your own impressions rather than estimated by a keyword tool. Your job is to translate those intents into natural questions.

Copying queries as they are distorts the whole measurement. Your AI visibility score rises on phrasings nobody uses, while your competitors capture the recommendations on the real questions. Your prompt set is not a setup detail. It decides the value of every number you will present to your client or your management.

What Search Console shows, and what it hides, in 2026

Search Console remains the best starting point, but three blind spots limit what it can teach you about conversational usage.

The generative AI report returns no queries

In June 2026, Google launched dedicated reports on visibility within generative AI features, for Search and for Discover¹. French industry press reports that they break impressions down by page, country, device and date, with no clicks, no CTR and no queries⁶. You know which pages appear in AI Overviews or AI Mode. You don’t know for which questions.

AI Mode is blended into your Web data

Since June 2025, AI Mode clicks, impressions and positions count toward the totals of the Performance report, with no way to separate them from classic organic traffic³. Some of your long queries may already come from conversational usage. Nothing lets you check it row by row.

Long queries are the most hidden

Google anonymizes queries that are too rare, to protect user privacy. Long, detailed phrasings, the ones closest to prompts, are by nature the rarest. Part of what matters to you will never show up in the export.

These limits don’t invalidate the method. They require you to treat Search Console as an intent signal, never as a ready-to-use prompt list.

The bias nobody mentions: Search Console only sees your own positions

Search Console only lists queries for which your site earned at least one impression. A set built only from it measures your visibility where you are already present. It ignores the intents where you don’t exist yet, which are often the ones where your competitors capture the recommendations.

This bias weighs heavily in GEO. AI engines don’t only cite your site: they rely on comparisons, directories, forums and press articles. At Cockpyt AI, more than 60% of the sites we analyze receive citations whose source is not their own site (Cockpyt AI, 2026). Your AI visibility plays out largely outside your domain, so outside your Search Console.

In my view, a healthy set contains at least two or three questions that Search Console never suggested. I find them in three places: recurring requests from sales and customer service, queries where a direct competitor ranks without you, and comparison questions (“X or Y for this need”). These questions often reveal an absence the fastest.

Which regex filters should you apply in Search Console?

Three regex filters and one native filter are enough to extract the useful queries. In the Performance report, click Add filter, choose Query, then Custom (regex). Extend the date range to 12 or 16 months: on a mid-sized site, three months return too few conversational queries to work with.

The \b trap deserves an explanation. Search Console uses the RE2 engine, where the word boundary only recognizes ASCII characters. In French, a regex ending with où\b finds nothing, because “ù” does not count as a letter. On any accented word, replace \b with \s.

Then add the native Non-branded filter. Launched on November 20, 2025, it automatically separates your branded queries from the rest, including “variations or misspellings of the brand name”, according to Google². It replaces the brand regex everyone used to hack together. It only works on domain properties with enough volume.

Finally, cross-check with metrics. Sort by impressions and spot queries with high impressions and a low click-through rate: people often get their answer directly on the results page. They will look for the same answer in an AI.

How do you turn a query into a prompt in three transformations?

Each intent you keep becomes a natural question through three additions. Before rewriting, group your queries by intent: “real estate agency fees nice”, “estate agent commission nice” and “how much does an agency charge to sell” express the same concern. In an AI, they become one question. Most sites only have one or two dozen truly distinct intents.

• Add the context. Who is asking? An owner, a buyer, a small business, someone in a hurry. Context steers the AI toward a type of answer and provider.
• Add the constraint. A location, a budget, a deadline, a size, an industry. The constraint forces the AI to sort, so to recommend names.
• Ask for a choice. End with “what would you recommend?”, “which ones should I choose?” or “what is the best option for me?”. Without a request for a choice, the AI explains instead of recommending.

The query “flat valuation nice” becomes: “I want to sell my flat in Nice this year, how do I get a reliable valuation and who should I contact?”

Read each question out loud. If it sounds like a keyword surrounded by words, start again. A generative AI can help you rephrase, as long as you feed it your intent groups from Search Console. Don’t ask it to invent the intents themselves: it will produce plausible questions, never measured ones.

Also remember that the AI in turn splits your question into search sub-queries before answering. This is the query fan-out mechanism, which you can observe on the Sources and Query Fan-out page. A good question triggers sub-queries close to your offer.

What size and balance for your prompt set?

Set size depends on your question budget, not on your number of queries. A 5-question set gives a first diagnosis, 15 questions cover the core of an offer, 45 questions are enough for three brands or three markets. On Cockpyt AI, these tiers match the free trial, the Solo plan and the Consultant plan.

Balance matters more than volume. An effective 15-question set follows a split close to this one:

• about ten generic questions from your category, without your brand, to see whether the AI thinks of you;
• three comparison or alternative questions, to measure your place against direct competitors;
• two questions about your brand itself, to check what AI engines say about you and catch errors.

In my view, filling the set with branded questions is the most common trap. They feel reassuring, because you almost always appear, but they inflate the score without saying anything about acquisition. Unbranded questions are the ones asked by people who don’t know you yet.

Test each question on several engines, because ChatGPT, Gemini, Perplexity, Claude and Mistral’s Le Chat don’t recommend the same players. If you must start with a single engine for manual checks, start with ChatGPT: according to SE Ranking’s study of 101,574 French sites tracked in GA4, it generates 74.78% of measured AI traffic in France⁴.

Queries to drop, even with volume

Some queries carry weight in Search Console and have no place in an AI tracking set.

Navigational brand queries. “your brand contact” or “your brand login” teach you nothing. The AI points to your site, you always come first, and your score climbs artificially. The Non-branded filter removes them upfront.

Purely transactional queries. “buy this product online” belongs mostly to search engines and price comparison sites. Recommendation questions upstream of the purchase say more about your place in the answers.

Plain definitions. “what is an exclusive listing” belongs in your content strategy, but an AI’s answer to this kind of question rarely cites a brand. Keep one at most if the topic is central to your business.

Queries outside your offer. A blog post attracting traffic on a side topic doesn’t justify a tracking question. Track what you sell, not what earns you impressions.

Full example: from a Search Console export to five prompts

Here is the method applied to a real estate agency in Nice. The example is fictional, but the queries resemble those I find on this type of site.

Queries surfaced by the filters: “which real estate agency to choose to sell in nice”, “real estate agency fees nice”, “free flat valuation nice”, “how long to sell a flat in nice”, “best real estate agency nice reviews”, “exclusive or open listing to sell fast”.

Intents identified: choosing an agency, understanding fees, valuing a property, anticipating timelines, choosing a listing type.

Final questions kept for the set:

“I want to sell my flat in Nice, which real estate agency would you recommend?”

“How much does a real estate agency in Nice charge to sell a flat, and which agencies are transparent about their fees?”

“How do I get a reliable valuation of my flat in Nice, and who should I contact?”

“How long does it take to sell a two-bedroom flat in Nice, and which agency sells the fastest?”

“To sell quickly in Nice, should I sign an exclusive or an open listing, and with which agency?”

Each question ends with a request for a recommendation, even when the original query was informational. That is deliberate. You are not measuring whether the AI can answer, you are measuring whether it cites you while answering.

When and how should your set evolve?

Review your set every quarter, never every week. Your offer changes, so do your competitors, and AI usage evolves fast, but a metric only has value if it stays comparable over time.

My rule fits in one sentence: replace two or three questions per quarter, no more. Remove those that no longer match your business, add the new intents that appeared in Search Console, and log every change. A curve that rises after replacing half the set proves nothing.

Use each review to rerun your three regex filters on the last three months. New long phrasings often signal a change in usage before it shows in your volumes. Also watch the questions where a competitor appears instead of you: they deserve content or off-site action first.

To set up your first tracking project once your set is ready, the Create your first project guide details each step.

FAQ: choosing your AI tracking prompts

How many prompts should you track to measure AI visibility?

Between 5 and 15 questions are enough for a small business with one main offer. Beyond that, each extra question must cover a new intent. An agency or consultant tracking several clients multiplies this base per project rather than widening a single set.

Can Bing Webmaster Tools help you find prompts?

Yes, as a complement. The AI performance report in Bing Webmaster Tools, launched in February 2026, shows queries grouped by intent and theme, while Google only returns impressions⁵. Volume stays low in France, but the phrasing is closer to conversational usage.

Should you track the same prompts on ChatGPT, Gemini and Perplexity?

Yes. Keeping the same questions across engines lets you compare your results from one engine to another. The gaps between engines are the useful information: they show which sources each AI favors.

Should your prompts include your brand name?

Only a minority of them. Questions with your name check what AI engines say about you and detect false information. Questions without your name measure your ability to be recommended, which remains the main stake.

Can you ask ChatGPT to generate your prompts?

For rephrasing, yes. For choosing intents, no. An AI invents plausible questions unrelated to your market’s real demand. Give it your intent groups from Search Console and only ask it to phrase them naturally.

Is the Search Console generative AI report enough to measure AI visibility?

No. It only covers Google’s AI surfaces and returns neither queries nor clicks. It says nothing about your presence in ChatGPT, Perplexity or Claude, which requires dedicated tracking on a set of questions.

Should a French site track prompts in English?

Only if you sell to English-speaking customers. AI engines don’t recommend the same players depending on the language of the question. A French set measures your visibility with your French customers, which is usually the one that matters.