Prompt research: what people ask engines, and why it is not keyword volume

Prompt research is the work of finding out what people actually ask an answer engine about your subject. It is the successor to keyword research and inherits that page's discipline: start from intent, and treat every volume number as an estimate. It also carries a harder problem of its own. For prompts, the volume data barely exists.

How a prompt differs from a keyword

• Length and form. A keyword is three or four words. A prompt is a sentence, often several, written the way a person would say it out loud.
• Constraints. Prompts name conditions: a budget, a country, a company size, a deadline, an exclusion. One prompt can carry four constraints that would have been four separate searches.
• Position in a conversation. Many prompts are follow-ups. They depend on the previous turn and would make no sense typed into a search box on their own.
• What comes back. A query returns a list to choose from. A prompt returns an answer with a small number of citations, so the outcome is closer to binary than a ranking is, as described in what AEO is.

Why prompt volume data is fundamentally weaker

Search engines report query data back to site owners. Google Search Console shows the queries your site appeared for, with impressions, clicks and average position, which is why this guide treats it as the one measured dataset in the whole discipline.

The AI surfaces do not do this. Google's generative AI performance report in Search Console reports impressions from generative AI features on Search and groups them by page, country, date and device. There is no query dimension in it. You can see that a page of yours was shown inside an AI feature. You cannot see the prompt that produced it.

OpenAI, Anthropic and Perplexity publish no per-prompt volume reporting for site owners at all. The closest primary data is the aggregate usage research the companies publish about themselves. OpenAI's September 2025 paper classifies a representative sample of ChatGPT conversations into categories such as practical guidance, seeking information and writing, and reports each category's share of messages over time. Anthropic's Economic Index aggregates conversations into occupational tasks using an automated pipeline that keeps researchers from reading the conversations themselves. Both publications are deliberately aggregated. Both describe categories of use. Neither reports how often any specific prompt is asked, and neither is intended to.

The consequence is direct. Every prompt volume figure on the market is an estimate, built from a panel, an opt-in browser or app sample, sampled conversations, or inference from search data. None of it comes from the engine's own reporting, because the engines do not publish that reporting. Two products can therefore report different volumes for the same prompt and neither number can be checked against a published source. Cross-vendor prompt volumes are unverifiable from the outside. Impressions in Search Console are a measured quantity. Prompt volume is a modeled one with no public ground truth behind it.

Prompt candidates you can actually trust

1. Your own Search Console queries. Measured rather than estimated, and specific to your site. Filter for question-shaped demand: queries starting with how, what, why, can, best or versus, and the long specific ones. People phrase questions to engines much as they type them into a search box. How to pull and read that export sits on Search Console.
2. Sales and support question logs. Every question a prospect asks before buying, and every question a customer asks afterwards, is a prompt someone will eventually type into an answer engine, phrased in their words rather than in your product's vocabulary.
3. Follow-up questions in your own channels. Comments, webinar Q and A, community threads, replies to email. These capture the second and third turn of a conversation, which is where prompts live and where keyword tools have nothing to say.
4. The expansion the engines perform themselves. Google documents query fan-out, a set of concurrent related queries the model generates to fetch additional results for the user's question. One prompt becomes several sub-queries you never see. The mechanics are covered in how answer engines choose their sources. In practice this means writing out the sub-questions under a prompt is not busywork. It is the shape of the retrieval.

What prompt research is for

Content planning

Prompts name constraints, so the plan should cover the constraint dimensions people actually use: price, size, region, industry, timeline, alternatives, the objection they raise. A page that answers only the head question misses the sub-questions fan-out generates, and those sub-questions are where citations are available. The practical output of prompt research is a list of questions and sub-questions to answer plainly on the page, not a list of phrases to place in it.

Measurement

The second use is the set of prompts you track over time, covered in prompt sets. Selection matters more than size here. A stable, representative set checked repeatedly tells you more than a large set assembled from estimated volumes, because the estimates cannot be validated and the trend in your own set can be.

A prompt list is per market, not global

Prompts in different languages retrieve differently, and the effect is large. A study of language effects on brand visibility in LLM answers found that the language of the query explained 26.5 to 32.0 percent of the variance in the answers, while brand identity explained 1.5 percent. A prompt list translated from English is a list of English assumptions in another language. Build the list from each market's own question sources, as covered in multilingual GEO.

A closing caution

Prompt research is genuinely useful for planning and genuinely unreliable as a volume metric. Both halves of that sentence are load-bearing. Ask any product that reports prompt volumes for its sampling frame: where the prompts were observed, from how many people, in which countries and languages, across which engines, and over what period. A vendor that cannot answer is quoting a number nobody can check. Use prompt research to decide what to write and what to track. Do not use it to forecast traffic.