What is prompt research?
Prompt research is the practice of identifying the AI prompts your buyers, customers and evaluators are typing into ChatGPT, Perplexity, Claude and Google AI Overviews - then deciding which of those prompts your site should own. It is the GEO equivalent of keyword research.
Traditional keyword research starts from a list of phrases people search on Google. Prompt research starts from a list of full-sentence questions people ask AI assistants. Both ultimately map to pages, but the inputs and the success criteria are different.
How are prompts different from keywords?
Prompts are longer, more specific, and usually contain a buyer context that keywords leave out. "Product analytics" is a keyword. "Best product analytics tool for a 15-person B2B SaaS using Snowflake" is a prompt. The prompt has done half the qualification before it arrives.
Three structural differences matter for GEO:
- Length. Prompts average 15-30 words; keywords average 2-4.
- Context. Prompts carry buyer attributes (team size, stack, geography, budget).
- Output. A keyword returns 10 ranked URLs. A prompt returns one synthesised answer with 3-5 cited sources. There is no second page.
2-3 words. Volume measurable in Search Console and SEO tools. Targets a SERP with 10 blue links.
Sentence-length. Volume is invisible. Targets a single synthesised answer with 3-5 cited sources.
Why should you do prompt research and keyword research?
Because the audiences overlap but the surfaces don't. Google still drives more pipeline than AI search for most B2B SaaS today, and the same buyer who types a keyword into Google on Monday will paste a prompt into ChatGPT on Tuesday. Skipping keyword research leaves money on the table; skipping prompt research means you only show up for the lower-intent half of the funnel.
The right approach is to do both in the same workflow: pick a topic, find the keywords that anchor the SEO ranking, find the prompts the keyword expands into, and build one page that satisfies both. For a primer on the SEO half, our metadata guide and the wider on-site optimisation overview cover the page-level mechanics.
How do you find the prompts your buyers actually use?
Volume data for prompts is not yet public, so prompt research is a triangulation exercise rather than a lookup. Use as many of these sources as you can and look for prompts that show up more than once:
- Sales calls. The first three questions a prospect asks on a discovery call are almost always the prompts they typed beforehand. Pull them from Gong, Fathom or your transcripts.
- Support and onboarding. Read the last 50 support tickets. Each one is a real-world prompt your existing customers couldn't get answered by your site.
- Reddit and HN. Search "recommend [category]", "alternatives to [competitor]" and "anyone using [stack]". Threads with 20+ comments are prompt goldmines.
- "People also ask" boxes. Google's PAA is the closest public proxy for AI follow-up prompts. Each PAA entry is a prompt you can own.
- Direct AI prompting. Ask ChatGPT: "List 20 questions a head of growth at a B2B SaaS would ask about [category]." Then run each through Perplexity and see which sources get cited. Those sources are your competitive set.
- Your own logs. If you have an on-site search or chat, every query is a prompt. Most teams ignore this data.
How do you map prompts to page types?
Once you have a shortlist of 30-50 prompts, sort them by intent. Awareness prompts need a guide or definition. Comparison prompts need a listicle or alternatives page. Validation prompts need a case study or pricing FAQ. Decision prompts need a structured pricing page.
A single page can satisfy multiple prompts as long as the answers are extractable. A good comparison page, for example, can answer "X vs Y", "alternatives to X" and "is Y worth the price" in three different sections of the same URL.
How do you prioritise which prompts to own?
Score each prompt on three axes:
- Commercial intent. How close to a buying decision is this prompt?
- Current visibility. Run the prompt today in ChatGPT and Perplexity. Are you cited, mentioned, or invisible?
- Effort to win. Do you have a page that almost answers this, or do you need to build from scratch?
The first prompts to fix are the ones with high commercial intent where you are invisible but already have a page that nearly answers it. That's the work that moves pipeline fastest. Our guide on the free GEO audit walks through the scoring system in detail.
How often should you redo prompt research?
Quarterly is enough for most B2B SaaS teams. Prompts shift slower than keywords because they track buyer language, not search-algorithm trends. The two events that should trigger an unscheduled rerun are a new product launch and a new competitor entering your category.
What mistakes kill prompt research?
- Treating prompts as long keywords and writing pages padded with question phrases.
- Only running prompts in one AI tool. ChatGPT, Perplexity, Claude and Google AI Overviews cite different sources for the same prompt.
- Skipping the "current visibility" check, so you build pages for prompts you already win.
- Ignoring keyword research entirely. Google still indexes the pages AI cites; SEO and GEO compound.
For the next step, see how to find out if your SaaS is showing up in AI search - it gives you the exact prompts to score yourself against once your research is done.