Prompt Volume Research: Keyword Research for AI Search

Prompt volume research is the AI-era analogue of keyword research: figuring out what your buyers ask AI assistants, in what words, so you can compete for those answers. The catch is structural — keyword research works because search demand is measurable (Google exposes volumes, tools resell them). Prompt demand has no public record. Conversations with ChatGPT are private; no vendor publishes per-prompt volumes.

The difference isn’t just measurability — the queries themselves change shape:

Why this matters commercially is the same story as everywhere in this space: the questions are migrating to surfaces where clicks were already declining, and the brands in the answers get the consideration.

Vendor content is full of invented “top ChatGPT prompts” lists. The credible public evidence is thinner and worth knowing precisely. The best of it is OpenAI’s own large-scale usage study (published September 2025 with an NBER working paper), which classified over a million real consumer conversations.

Two implications for prompt research. First, advisory phrasing dominates: people ask “what should I use for X” and “how do I do Y,” not category-label keywords. Second, information-seeking is a top-three use — the questions your content answers are being asked, verbatim and conversationally, inside assistants right now.

Since you can’t look demand up, you assemble it from sources ranked by how close they sit to real buyers:

1. 1Your own conversations (highest signal). Sales-call recordings, support tickets, onboarding questions, and what prospects type into your site search or support widget — these are literal prompts, in buyer vocabulary, with intent attached.
2. 2Community questions. How your category is asked about on Reddit, in Slack communities, on Stack Overflow — conversational phrasing in the wild, and doubly relevant because assistants themselves lean on these sources (see Reddit and AI citations).
3. 3Search’s conversational edges. People Also Ask boxes, autocomplete on question stems (“how do I…”, “best way to…”), and long-tail question keywords — the subset of search data that already looks like prompts.
4. 4The assistants themselves. Ask ChatGPT or Perplexity “what do people ask you about [category]?” — treat the output as brainstorming, not data, but it surfaces phrasings you didn’t think of.
5. 5Vendor prompt datasets (use with care). Some SEO platforms now sell prompt-volume estimates built from panels and modeling. Methodologies are opaque and unverifiable; useful for relative comparisons, never for absolute volumes.

Then organize by intent, exactly as you would a keyword map: problem prompts (“how do I see which companies visit my website”), category prompts (“best B2B visitor identification tools”), comparison prompts (“X vs Y for SaaS”), and brand prompts (“is X any good”). Problem prompts are usually the volume and the least served by existing category pages.

One dynamic is genuinely new. Assistants answer definitional questions constantly — the emerging-vocabulary questions (“what is warm outbound”, “what is an AI SDR”) where no incumbent owns the answer yet. When a term is young, the first sources to publish a crisp, quotable, consistently-repeated definition tend to become the ones assistants reach for — the original GEO research found quotable, citation-backed content measurably lifted visibility in generated answers.

• Find the young terms in your category — the vocabulary your best customers use before the analysts standardize it.
• Publish the 40–60 word definition early, repeat it verbatim across your pages, and make it the first thing under a question-shaped heading.
• Watch whether assistants adopt your framing — when your phrasing shows up in answers, you own a piece of the category’s vocabulary.

Without volume data, validation is empirical: run the candidate prompts and see what happens.

The keepers become your measurement baseline — the fixed prompt set you probe weekly for citation rate and share of voice, per the methodology in our AI visibility metrics guide. And the sources cited in today’s answers are your competitive teardown: they show exactly what the engines currently consider the best evidence, which is the gap your content has to close (the tactical side of that lives in the GEO playbook).