People type their problem into ChatGPT, not a keyword. Yet many brands begin their AI approach with a list of prompts stretched from their keyword list: “best [product]”, “[service] cost”, “[brand] review”. Logical, because that is how we did it with search engines. But it measures the wrong thing. A prompt is not a longer keyword, it is a question or a problem in plain human language. Ignore that difference and you are optimising for searches nobody phrases that way.
A prompt is not a keyword
The difference is in the form. With a search engine you type a few keywords and puzzle the rest together yourself. With an AI you lay down your whole problem and let the model do the work. You see that back in the data. Research by Pew Research (68,879 searches, 2025) shows that only 8% of one or two word searches triggered an AI summary, while for ten words or more it rose to 53%. And 60% of searches that started with a question word (who, what, when, why) got an AI answer. Longer, question-driven phrasing drives AI answers, short keywords don’t.
In its own research on more than 80 million clickstream records, Semrush measured that a ChatGPT prompt without the search function averages 23 words. People talk to an AI the way they would talk to an advisor, not the way they fill in a search bar.
A prompt is not a keyword. It is your customer’s problem, in their own words.
Giacomo Perticara, founder and SEO strategist at GRP Digital
The AI splits your question up anyway
There is another reason why “ranking for a single prompt” doesn’t work: the AI doesn’t do that either. Google describes in its own documentation the “query fan-out”: AI Overviews and AI Mode split one question into several related searches across subtopics, and build one answer from them. One customer question thus becomes a handful of searches internally. Steering on the position of a single prompt is then as pointless as aiming for one keyword while the customer types a whole story. We covered that mechanism separately in why ranking for a single keyword is no longer enough.
What you should actually base your prompt set on
Not on your gut, and not on the default list your AI tracker supplies. Base it on real data and real customer problems. Google’s Search Console has shown, since 2026, impressions from AI features at page level, and Bing Webmaster Tools even shows the grounding queries: the searches the AI generated internally to fetch your content. That is real, source-driven data. Add to that what your customers actually ask: look in your customer service, your live chat, your contact forms. That is where the problems sit in your audience’s own words, exactly as they put them to an AI.
The art is not to collect as many prompts as possible, but to pull the patterns out of them. Link them to your product lines, your service lines and the stages of the customer journey. Thirty to forty meaningful prompts that cover real customer problems are worth more than a list of three hundred stretched keywords.
Steer on patterns, not on individual rankings
Another difference with classic SEO: at source level you get no neat ranking per prompt back. Google’s AI report gives impressions, no query breakdown, and Bing explicitly calls its grounding data a sample. Obsessing over the position of individual prompts is therefore a dead end. The question that does matter: are you consistently shown for a topic, or not? Mentioned once and not the next time is normal noise. Consistently appearing or consistently missing is the signal.
As Roxana Stingu, SEO lead at Alamy, put it in the panel by Aleyda Solis said: people build prompts that are merely extended keywords, when you should start with the problem the customer is trying to solve. At GRP Digital we therefore flip the order: first the customer problem, then the prompt.
Frequently asked questions
Do I then no longer need keywords at all? You do, they stay important for classic SEO. But for AI visibility they are the starting point, not the end product. Translate your keywords into the problems and questions behind them.
How many prompts should I track? Rather thirty to forty meaningful prompts that cover your product and service lines and the customer journey than a huge list. Quality and coverage over volume.
Where do I get real prompts? From AI Overview impressions in Search Console, grounding queries in Bing Webmaster Tools, and above all from your own customer contact: support, chat and forms.
Sources: Pew Research (query length and AI summaries), Google Search Central (query fan-out), Google Search Central (gen-AI impressions), Bing Webmaster Tools (grounding queries) and Semrush (prompt length, own clickstream research).
