# Keyword and Prompt Research for GEO: From Keywords to Prompts

**Author:** John Morabito (Founder, /winston)
**Published:** June 14, 2026
**Reading time:** 12 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/keyword-and-prompt-research-for-geo/

Keyword research does not disappear in generative engine optimization. It changes shape. The unit stops being a keyword you rank for and becomes a prompt you get cited in. Here is how to run research for the AI engines: turning keywords into prompt families, building a buyer-intent prompt set, mapping it to pages, and using the same set to measure citation share.

## The short version: research the prompt, not the keyword

In GEO, the thing you research is the prompt a buyer types into an AI engine, not the keyword they type into Google. A keyword is a compressed phrase built for a ranking algorithm. A prompt is a full, natural-language question built for a conversation. "Best running shoes" is a keyword. "What are the best running shoes for someone with flat feet who runs three times a week" is a prompt, and it is the shape the engine actually answers. Your research job is to collect those prompts across your category, tag them, map them to pages, and then measure how often you show up as a cited source across the whole set. That last number, citation share, is the scoreboard. Everything below is how to build the set that scoreboard runs on.

## What actually changes from keyword research

Traditional keyword research and GEO prompt research rhyme, and the muscle memory carries over. But three things move, and if you skip the shift you end up optimizing for a results page that the AI answer is quietly replacing.

| Dimension | Keyword research (SEO) | Prompt research (GEO) |
|---|---|---|
| Unit | A keyword or phrase | A full natural-language prompt |
| Intent signal | Compressed into a short string | Stated openly in the question |
| Success metric | Average position, ranked clicks | Citation share across the prompt set |
| Deliverable | Ranked keyword list | Tagged prompt set mapped to pages |
| Competitor view | Who ranks above you | Who the engine names instead of you |

The through-line: keyword tools still tell you where demand lives, and you still use them. They just become the seed for the prompt set rather than the finished product. The bigger philosophical break, why the two disciplines are not the same job with a new label, is the argument in the complete GEO audit methodology (https://www.winstondigitalmarketing.com/playbooks/the-complete-geo-audit-methodology/), which puts prompt research inside the full audit it belongs to.

## From keywords to prompts

The cleanest way to start is to treat every keyword you already care about as the seed of a prompt family. A single keyword rarely maps to one prompt. It fans out into the handful of ways a real person would ask an engine about that thing.

Take the keyword "project management software." A buyer does not type that into ChatGPT and stop. They ask: what is the best project management software for a small agency, is Asana or Monday better for a design team, which project management tools have the best free tier, what do people on Reddit actually recommend for project management. One keyword, four prompts, four different intents, four different pages that could answer them. Your job is to make that fan-out explicit rather than leaving it implied inside a keyword.

Start with your existing keyword data because it tells you where the demand and the money already are. Pull an export, a Search Console query list, or a rankings file, and weight the prompt set toward the themes that drive the business. Then expand each high-value keyword into its prompts. The keyword data is the seed. The prompt set is what grows from it.

## Building the buyer-intent prompt set

A prompt set is only useful if it mirrors how buyers actually ask, which means it has to be bigger and more honest than the ten obvious questions a team lists from memory. Build it in three passes.

**Pass one: cover the buyer journey.** Write prompts for every stage. Awareness prompts ("what is X", "how does X work"), consideration prompts ("best X for Y", "X vs Z"), and decision prompts ("is Brand good", "where to buy Brand", "how much does X cost"). Most teams over-index on one stage. The set should span all three because the engines answer all three, and the pages that win differ at each.

**Pass two: cover every competitor.** For each named competitor, write a comparison prompt in both a neutral form and the community-opinion form ("Brand A vs Brand B", "Brand A vs Brand B reddit"). Comparison prompts matter most because that is where an AI answer writes a verdict and ships it to a buyer who never visits your site. If you are not present in your own comparison answers, you are losing the deals you are closest to winning.

**Pass three: cover the awkward questions.** Price, alternatives, complaints, "is it worth it", "cheaper than". These are the prompts your competitors avoid answering, which is exactly why the engine cites whoever does. A published range with honest caveats gets cited almost by default. The full step-by-step build, including the seven sub-theme patterns and the tagging schema, lives in the GEO prompt research playbook (https://www.winstondigitalmarketing.com/playbooks/geo-prompt-research/); this section is the why, that one is the how.

## Prompt families and fan-out

Organize the set into families so it stays legible as it grows. A family is a keyword theme plus the pattern applied to it. The same handful of patterns fan out across every theme in your category:

- **Comparison:** "[Brand] vs [Competitor]", plus the community variant. The highest-stakes family.
- **Best-of:** "best [category] for [use case]", "top [category] [year]".
- **Cost:** "how much does [thing] cost", "is [thing] worth it", "cheaper alternative to [Brand]".
- **Definitional:** "what is [concept]", "how does [thing] work". These map to glossary and explainer content.
- **Buy / access:** "where to buy [Brand]", "[Brand] near me", "how to get started with [thing]".
- **Trust / proof:** "is [Brand] legit", "[Brand] reviews", "[Brand] complaints".

Fan-out is the act of running each pattern across each theme so no theme is left with a single lonely prompt. A category with eight themes and six patterns is not eight prompts; it is closer to fifty before you add the long-tail variants. That volume is the point. A twenty-prompt audit never surfaces the family where you are invisible. A few-hundred-prompt set does.

## Mapping prompts to pages

The prompt set becomes a build plan the moment you assign each family to the page that should own it. Group the set by family and intent, then attach each group to the single best-positioned page, or flag the gap where no such page exists.

- A **comparison family** maps to a comparison page, one that names competitors honestly and answers the "vs" question in the first passage.
- A **cost family** maps to a service or pricing page with published ranges, not a page that dodges the number.
- A **definitional family** maps to a glossary entry or explainer, one extractable definition per concept.
- A **trust family** maps to proof: case studies, author bylines, connected entity schema.

Families with no matching page are your content backlog, ranked by how much business the underlying demand represents. This is where prompt research stops being a report and turns into a calendar. Each mapped page then gets written to be lifted, not just read, which is the on-page craft covered in how we write content AI engines actually cite. And because Perplexity is the most citation-transparent engine, it is the fastest place to check whether a newly mapped page is winning its family, using the method in how to rank in Perplexity (https://www.winstondigitalmarketing.com/playbooks/how-to-rank-in-perplexity/).

## Using the prompt set to measure citation share

The reason to build the set with this much care is that it doubles as your measurement instrument. Once the prompt set is stable, run every prompt through the engines you care about, ChatGPT, Perplexity, Google AI Mode, Gemini, and record which domains get cited on each answer. Citation share is the percentage of your prompt set where your domain appears as a cited source, cut per engine and per family.

That single number, tracked weekly against a fixed set, is the GEO scoreboard the way average position was the SEO scoreboard. It tells you not just whether you are visible but exactly which families you own and which you are absent from, which is the same gap list that drives your content backlog. Run it by hand at first; the discipline of the fixed set matters more than the tooling. When you are ready to productize the measurement, treat the prompt set as the input to a recurring instrument rather than a one-time report.

The honest scope note: building the seed list and fanning out a first prompt set is an afternoon of focused work, and any team can do it. Where an agency earns its fee is the ongoing loop: keeping the set fixed enough to trust week over week, running it across every engine, mapping the gaps to a real content calendar, and shipping the pages that close them. That loop is what we run for clients through [our generative engine optimization service](https://www.winstondigitalmarketing.com/services/generative-engine-optimization/). If a vendor pitches "GEO keyword research" that hands you a keyword list with no prompt set and no citation-share plan, they have relabeled SEO and hoped you would not notice.

## Where this fits

Prompt research is the first move in a wider generative-search program: research the prompts, map them to pages, write pages that get cited, and measure citation share on repeat. The full audit that wraps this step is in the complete GEO audit methodology (https://www.winstondigitalmarketing.com/playbooks/the-complete-geo-audit-methodology/), the deep build of the set itself is in GEO prompt research (https://www.winstondigitalmarketing.com/playbooks/geo-prompt-research/), and the engine you should learn to measure on first is in how to rank in Perplexity (https://www.winstondigitalmarketing.com/playbooks/how-to-rank-in-perplexity/).

## Frequently asked questions

**What is prompt research in GEO?**
Prompt research is the generative-engine version of keyword research. Instead of collecting the short strings people type into Google, you collect the full questions people ask AI engines like ChatGPT, Perplexity, Google AI Mode, and Gemini about your category, then treat that set of prompts as the map of where you need to get cited. A keyword like best running shoes becomes a family of prompts: what are the best running shoes for flat feet, is Brand A or Brand B better for marathon training, which running shoe brands are actually worth the money. The prompt set, not the keyword list, is the thing you build content against and measure against.

**How is keyword research different for GEO?**
Traditional keyword research optimizes for a ranked position on a results page. GEO research optimizes for being one of the sources an AI answer names. The unit changes from a keyword to a prompt, the intent is expressed in a full natural-language question rather than a compressed phrase, and success is measured as citation share (how often you appear as a cited source across a fixed prompt set) rather than average position. You still use keyword tools to find demand and seed the work, but the deliverable is a tagged prompt set mapped to pages, not a ranked keyword list.

**How many prompts should a GEO prompt set contain?**
Enough to sample the real spread of how buyers ask, which is usually far more than the ten or twenty obvious questions teams start with. A useful set for a single category typically runs a few hundred prompts built from a handful of prompt families and expanded through fan-out. The exact number matters less than the coverage: you want every buyer-intent theme represented, every named competitor covered in a comparison prompt, and enough volume that a weekly citation-share reading is stable rather than noisy.

**How do you map prompts to pages?**
Group the prompt set by theme and intent, then assign each group to the single page best positioned to answer it, or flag the gap where no such page exists yet. A comparison family maps to a comparison page, a how-much-does-it-cost family maps to a pricing or service page with published ranges, a what-is family maps to a glossary or explainer. Prompts with no matching page become your content backlog. The map is what turns prompt research from a report into a build plan.

**Can I use my existing keyword data to build a prompt set?**
Yes, and you should. An existing keyword export, a Search Console query list, or a rankings file tells you which themes already drive demand, so you weight the prompt set toward the categories that matter to the business instead of guessing. You expand each high-value keyword into its prompt family, add the comparison and buyer-intent variants that keyword tools rarely surface, and tag everything by theme and intent. The keyword data is the seed; the prompt set is what grows from it.

Service: https://www.winstondigitalmarketing.com/services/generative-engine-optimization/
Audit: https://www.winstondigitalmarketing.com/contact/#audit