Prompt Research methodology
Prompt Research answers one question: which prompts should you track? It takes a business area and a market, and returns a set of topics and prompts that cover what people ask AI platforms about that area, ranked by demand. This page documents how it works, where every input comes from, and what it cannot tell you.
Framework version 2. Last updated November 18, 2026.
At a glance
| Step | What happens | Kind of input |
|---|---|---|
| 1. Adapt the intents | The four GEO intents and their 23 sub-intents are adapted to your business area and market | Framework |
| 2. Derive topics | For each sub-intent, the distinct questions people ask become topics | Generated |
| 3. Validate and size | Each topic is checked for category fit, company fit and language, and Google search volume for that market is attached | Generated, checked · Estimated |
| 4. Draft prompts | Natural questions are written for each topic | Generated |
| 5. Remove duplicates | Prompts too close to each other, or to prompts you already track, are removed | Checked |
| 6. Add high-demand neighbors | Up to ten neighboring keywords with high search volume become topics after the same validation, and compete with the draft for a place in the set | Estimated, checked |
A run covers one business area in one country and one language. You decide which of the proposed topics and prompts to track. Nothing is tracked until an administrator adds it.
The four GEO intents
Prompt Research is intent-first. It does not start from a keyword list. It starts from the reasons people turn to an AI platform, and makes sure each reason is covered before demand decides the ranking.
| Intent | The user wants to | Sub-intents |
|---|---|---|
| Informational | Understand something | Explanatory · Procedural / how-to · Applicability · Simple fact · Currency |
| Brand / entity | Learn about a specific company or product | Identity · Reputation and trust · Positioning · Capability and fit · Comparison and alternatives · Entity facts |
| Commercial | Choose between options | Ranked recommendation · Constrained recommendation · Head-to-head · Alternatives and switching |
| Transactional / agentic | Act, or have the assistant act | Where to get · How to start · Quote and pricing · Demo or trial · Deals · Account actions · Availability · Direct agentic |
Examples of what each sub-intent looks like:
- Informational. “What is [X]”, “how does [X] work”, “who qualifies for [X]”, “what changed with [X]”.
- Brand / entity. “Is [brand] legit”, “what is [brand] known for”, “is [brand] good for [use case]”, “[brand] vs [competitor]”.
- Commercial. “Best [category]”, “best [category] for [use case]”, “[A] vs [B]”, “alternatives to [incumbent]”.
- Transactional / agentic. “Where can I get [category]”, “get a quote for [service]”, “is [category] available in [location]”, “book me [category]”.
Where the framework comes from
The four intents continue the intent taxonomy that search research has used for two decades, re-cut for AI answers:
- Broder (2002), A taxonomy of web search, introduced informational, navigational and transactional intent.
- Rose and Levinson (2004), Understanding user goals in web search, refined these into goal types. Two of our sub-intents follow them: simple fact and availability.
- Chatterji et al. (2025), How People Use ChatGPT (NBER Working Paper 34255), informed the informational sub-intents with observed usage of an AI assistant.
Two changes matter for AI. Navigational intent becomes brand / entity: people no longer ask an assistant to find a website, they ask it what a company is and whether to trust it. And transactional intent extends to agentic: asking the assistant to carry out the task.
GEO is a young field and has no authoritative sub-intent taxonomy yet. The four intents are stable. The sub-intents are our working framework: some follow the research above, some are transposed from search practice, and some are reasoned from what we see in AI answers. We version the framework and note changes on this page.
The pipeline
1. Adapt the intents
The generic framework is adapted to your business area, country and language. A sub-intent such as “Applicability” reads differently for a retail bank than for a software vendor. The adapted framework is stored per company, business area, country and language. The generic one is never overwritten.
2. Derive topics
For each adapted sub-intent, a language model proposes the distinct questions people ask inside it: three to eight topics per sub-intent. Each topic carries a label, a one-line statement of what the user wants to know, and two to four short keyword forms of two to three words each. Topics are written natively in the market’s language. Literal translations of generic English phrasing are rejected. Near-duplicate proposals are dropped at this step: within an intent, a proposal is dropped when its semantic similarity to an earlier one is 0.90 or higher, or when it shares its first keyword form with an earlier one.
3. Validate and size
A second, independent model scores every proposed topic. Three checks decide whether a topic passes:
| Check | Passes when |
|---|---|
| Category fit | The topic is squarely about the business area |
| Company fit | The topic is at least plausibly relevant to your company |
| Language | The topic is in the market’s language |
Topics that fail a check are dropped. The model also scores whether a topic matches its parent intent. That score is recorded with the topic and does not drop it.
Beside validation, each topic’s keyword forms are looked up in Google Ads search volume for that country and language. The draft is then assembled: up to 50 topics, filled round-robin across the sub-intents, so every sub-intent that has passing topics is represented. Search demand explains the scale and what happens when more topics pass than the draft holds.
4. Draft prompts
For each topic, natural questions are written the way a person would ask an AI platform: at least two per topic, three for higher-demand topics.
5. Remove duplicates
Prompts that are near-duplicates of each other are removed (semantic similarity of 0.90 or higher). Prompts that restate something you already track are suppressed (similarity of 0.85 or higher), and the topic is marked as covered, so a run extends your prompt set and does not repeat it.
6. Add high-demand neighbors
Neighboring keywords are looked up in Google Ads as well. The lookup is seeded from the name of the business area and from keyword forms of the topics outside the brand / entity intent, up to two per topic. Up to ten neighbors with high search volume that pass the same validation compete with the draft for a place in the set. The set is trimmed to about 50 topics: draft topics and additions are ranked by fit from validation first and demand second, and the weakest give way. A topic your tracked prompts already cover and the last topic of a sub-intent are never displaced, so a set can end above 50, and never above 60. Additions that stay get their prompts.
Neighboring keywords that belong to an existing topic count toward that topic’s demand, so a topic’s demand can change at this step.
What a run records
Every run records how many topics passed validation and how many were dropped, how many were in the wrong language, how many neighboring topics joined the set and how many draft topics they displaced, and how many prompts were removed as duplicates or suppressed because you already track them. These run figures are not shown in the app today and are available from Sitefire on request.
Search demand
Demand is shown per topic on a five-step scale:
| Bar | Monthly searches in the market |
|---|---|
| 1 | Under 50 |
| 2 | 50 to 499 |
| 3 | 500 to 4,999 |
| 4 | 5,000 to 49,999 |
| 5 | 50,000 and above |
Demand ranks and labels a topic. A topic’s demand counts its keyword forms and the neighboring keywords attributed to it. No search data is shown when no search volume is available for the topic in that market, either because Google Ads has no data for it or because the lookup returned nothing. Sitefire estimates no number of its own for such a topic.
Demand also decides which topics are left out when a run has more than the set holds. When more topics pass validation than the draft holds (50), the draft is filled round-robin across the sub-intents, highest demand first within each sub-intent. When neighbors join in step 6, topics are ranked by fit first and demand second. Within a sub-intent in the draft, and among topics of equal fit in step 6, the topic with lower demand is left out first. A topic without search data counts as the lowest demand.
A worked example
An illustrative trace for the business area “e-bike insurance” in the United Kingdom, in English. The demand band and the neighboring keyword are examples, not measured values.
| Stage | Result | Kind |
|---|---|---|
| Intent | Commercial | Framework |
| Sub-intent | Constrained recommendation: “best [category] for [use case]“ | Framework |
| Topic | E-bike insurance for commuters | Generated |
| Validation | Category fit yes, company fit yes, language English. Intent match recorded | Generated, checked |
| Demand | Bar 3, 500 to 4,999 monthly searches for the topic’s keywords | Estimated |
| Prompt | ”What is the best e-bike insurance for someone who commutes daily in London?” | Generated |
| Duplicate check | No tracked prompt reaches 0.85 similarity, so the prompt is proposed | Checked |
| Neighbor | ”E-bike theft insurance” has high search volume, passes validation and joins the set as a topic | Estimated, checked |
Where each input comes from
Sitefire keeps its inputs apart. Where the source is known, it is stored with each topic (site keyword, keyword expansion, keyword suggestion, related keyword, entered by a user, generated by a model, or a Prompt Research run) and with each prompt (the user who added it and the topic it belongs to). A topic that already has a source keeps it when you add the same topic from a Prompt Research run.
| Kind | What it is | Refresh |
|---|---|---|
| Observed demand | Your own Google Search Console and Bing Webmaster queries, and visits AI platforms refer to your site (Google Analytics 4) | As your connected accounts update |
| Observed platform behavior | Query fan-outs: the web searches an AI platform runs while it answers a tracked prompt | Daily, with every answer |
| Estimated demand | Google Ads monthly search volume per country and language, retrieved through DataForSEO | With every Prompt Research run and import; administrators can refresh it once a month |
| Generated | Topics and prompts drafted by Prompt Research | Per run |
| Keyword-derived | A keyword from your site or Search Console can seed a topic; the prompts generated for it keep the link to that keyword | On import |
| Customer-entered | Prompts you write or import yourself | When you add them |
Query fan-outs and your own search data are separate signals in the product. They do not feed a Prompt Research run automatically. Sitefire has recorded more than 6 million fan-out queries. The patterns in them, such as the current year, the market’s name or words like “best” and “official” that the model adds on its own, inform how Prompt Research words its prompts. See Query Fan-Out and our research on why AI models cite different sources.
Why we do not use clickstream prompt panels
Some tools size demand with “prompt volumes” bought from consumer panels: samples of what opted-in users type into AI platforms, scaled up to a population estimate. We chose not to.
Prompts are not keywords. A panel prompt is often several sentences long, depends on the conversation before it, and is shaped by personalization. Two people with the same need rarely type the same prompt, so counting identical prompts across a sample and scaling the count does not produce a volume you can verify. Search demand has its own limits (see below), but it comes from a source you can check against your own Search Console data.
Markets and languages
A run is scoped to one country and one language, and every topic and prompt carries both. Research is done natively per market, so local terminology decides the wording and that market’s own search demand decides the ranking. Sitefire supports 218 countries and 72 prompt languages. Search demand depends on Google Ads: where it has no data for a topic in a country and language, or the lookup returns nothing, the topic is labeled No search data.
Limits
- Search demand is a proxy. Google search volume measures demand on Google, not on AI platforms. No vendor has access to the AI platforms’ own query logs. Sitefire reports no per-prompt AI volumes.
- Figures are rounded. Google Ads volumes are rounded and grouped by Google, which is why we show bands and not confidence intervals.
- Trend, not forecast. Rising or falling demand describes the past twelve months. It is not a prediction.
- Thin data on specialist topics. Institutional, B2B and small-market topics often have little or no search volume. They are labeled and rank low. When a run has more topics than the set holds, low demand counts against a topic, so such topics can be left out. Whether to track them rests on the intent framework and your own judgment.
- Generated prompts need review. Topics and prompts are drafted by language models and checked by a second model. Review them before you track them.
- Runs are limited. Five Prompt Research runs per workspace and month are included by default. Administrators start a run.
Related
- KPIs for how search demand weights the Visibility Score
- Prompt Research, weekly opportunities, and new connectors for the release notes