AI Keywords: How to Choose the Prompts to Track for Your Brand (2026 Method)

Prompts are the new keywords

In SEO, everything starts with keyword research. In GEO, everything starts with the prompt set: the list of questions you regularly test on ChatGPT, Perplexity, Gemini and the other engines to find out whether your brand appears in the answers. It is the first step of any AI visibility measurement, and it determines the quality of everything that follows: a poorly built prompt set produces a reassuring, false dashboard.

The keyword comparison has its limits, though, and understanding them prevents the main mistakes. A keyword is a typed expression ("seo agency lyon"); a prompt is a formulated question ("what is the best SEO agency in Lyon for an SMB?"). A keyword has a measurable search volume; a prompt doesn't, since nobody publishes ChatGPT's logs. A keyword returns a results page where a hundred sites exist; a prompt returns an answer where two to five brands exist. Selection is therefore even more strategic than in SEO: every tracked prompt is a bet on what your customers actually ask.

The method: 5 categories to cover your market

A good prompt set covers five distinct intents. The proportions below are what we recommend for a set of 30 to 50 prompts; adjust to your business.

1. Recommendation prompts (40 to 50% of the set)

The engine room: questions where a customer asks the AI to choose for them. "Best [service] in [city]", "which software for [need]", "which [product] brand do you recommend for [profile]". These prompts carry most of the commercial value, because the answer is a direct prescription. Decline them by segment: by city if you are local, by use case if you are software, by product line if you are e-commerce.

2. Comparison prompts (15 to 20%)

"[You] vs [competitor]", "alternative to [leader]", "choosing between [A] and [B]". These questions are asked every day, and AIs make and unmake shortlists in their answers. Include the comparisons where you naturally appear, but also those where you should appear and don't yet: those are your conquest targets.

3. Problem prompts (15 to 20%)

The questions your customers ask before they know they need you: "why doesn't my site appear in ChatGPT", "how do I know if AI recommends my restaurant". Being cited in these answers places you upstream in the buying journey, where competitors aren't in the conversation yet.

4. Brand prompts (10 to 15%)

"What is [your brand]?", "[your brand] reviews", "how much does [your brand] cost?". They don't measure your conquest but your accuracy: what AIs say about you when asked directly. This is where pricing, offering or positioning errors get detected, common and costly ones. Keep their share limited: a set dominated by brand prompts measures your ego, not your market.

5. Distrust prompts (5 to 10%)

Negative or skeptical questions: "is [your brand] reliable?", "scams to avoid in [your industry]". Not pleasant to track, but they are the ones that reveal a reputation problem in the answers before it becomes visible anywhere else.

Where to source your prompts, concretely

Four sources, in order of yield. Your existing SEO queries first: rephrase your converting keywords into natural questions; Search Console is your best starting point. Your customer conversations next: the questions asked in demos, pre-sales and support are exactly what those same people now ask AIs. The AIs' own suggestions: ask ChatGPT "what questions would a customer ask before choosing a [your category]?"; the answers are often good candidates. Community spaces finally: forums, groups, Reddit, where questions are already written in natural language.

Test each candidate once before adding it: if the answer never cites any brand (too generic a question) or systematically cites the same untouchable giants, the prompt will consume tracking without producing insight.

The mistakes that ruin a prompt set

• The showcase set. Only tracking prompts where you already appear: the dashboard is flattering and teaches nothing. A good set contains a majority of conquest prompts.
• The ego set. Too many brand prompts: you measure your awareness, not your share of the market inside the answers.
• The frozen set. A prompt set gets revised: every quarter, remove the prompts that went silent and add the market's new questions. Queries about Google's AI Mode didn't exist two years ago; tomorrow's don't exist yet.
• The monolingual set. If you sell in French and English, track both: AI answers differ sharply by language, and a competitor can dominate you in one language and not the other.
• The bloated set. 500 prompts tested once a quarter are worth less than 30 prompts tested daily. Repetition beats volume, because AI answers are non-deterministic: only the average makes the data.

How many prompts, how often

For most brands, 30 to 50 well-distributed prompts cover the essential territory. Below 20, blind spots dominate; beyond a hundred, tracking becomes noise unless organized by segments. The ideal frequency is daily with a tool, monthly at minimum manually; the indicators to compute on those measurements (citation rate, share of voice, accuracy) are detailed in our guide to core GEO KPIs, and the optimization checklist tells you what to do with the results.

Frequently asked questions

What is a tracked prompt? A customer question tested at regular intervals on AI engines to measure whether your brand appears in the answer, against which competitors and with which information. It is GEO's equivalent of a tracked keyword in SEO.

How many prompts should you track? 30 to 50 for most brands, split between recommendation (about half), comparison, problem, brand and distrust. Better a tight set tested daily than a huge set tested rarely.

How do I know what questions my customers ask AIs? There is no "Search Console for ChatGPT": you reconstruct by deduction. Rephrase your SEO queries into questions, mine your sales and support conversations, and ask the AIs themselves what questions a buyer in your category would ask.

Do prompts replace keywords? No, they add to them. Your SEO keywords remain the raw material of your prompts, and SEO remains GEO's entry requirement: the two disciplines reinforce each other.

How often should you revise your prompt set? A quarterly revision is enough in most cases: remove silent prompts, add the market's new questions, reflect product launches and the new competitors AIs are citing.

Conclusion

The prompt set is to GEO what keyword research is to SEO: the foundation work that conditions everything else. Five categories, 30 to 50 questions, a majority of conquest, one revision per quarter: the method fits in one sentence, and discipline makes the difference. To see immediately what AIs answer on your first strategic questions, the free Sapian Metrics audit tests your brand on 8 engines in minutes: the fastest way to turn your first prompt set into a baseline snapshot.