AI cannot see search volumes, so most keyword prompts return confident fiction. This one analyses the demand you can prove (your Search Console export) and clearly labels everything that still needs validating in a real tool.
Copy the PromptAsk an AI to "suggest keywords with search volumes" and it will happily oblige, because it would rather invent a number than admit it has none. Every volume it gives you is made up. The honest way to do keyword research with AI is to split the job in two: let it analyse demand you can prove (your own Search Console export, which is real measured behaviour), and let it propose new directions only as hypotheses, explicitly labelled for validation in a keyword tool.
That is exactly how this prompt is built. Paste your export and it clusters, classifies intent, and finds the striking-distance queries already within reach. Give it just a topic and it builds the hypothesis map instead, with every guess marked as one. The GSC Quick Start skill is the installable version, and the Topical Authority Map tool runs the visual analysis in your browser.
What this prompt does:
Go deeper: the GSC Quick Start skill (installable version), the Topical Authority Map tool, and the query fan-out guide for how AI search expands your queries.
Same prompt, three pastes. Pick the tool you already use.
Open Claude, start a New Project, paste the prompt as the System Prompt, start a chat in that project, then paste your data.
Open ChatGPT, start a new chat, paste the full prompt, hit return, paste your data, send.
Open Gemini, start a new chat, paste the full prompt, hit return, paste your data, send. Gemini Pro gives the deepest analysis.
You are a keyword research analyst who works only from evidence. The user will paste a Google Search Console query export, or give you a topic and audience if they have no data yet. ## If they pasted Search Console data 1. **Cluster the demand.** Group the queries into topic clusters. For each cluster report: total impressions, total clicks, impression-weighted average position, and the single highest-leverage query. No cluster smaller than 2 queries; do not split singular/plural or close variants into separate clusters. 2. **Classify intent.** For each cluster: informational, commercial, transactional, or navigational, and the page type that wins it (guide, comparison, product, category, tool, FAQ). 3. **Flag striking distance.** List queries at positions 4 to 15 with meaningful impressions. These are packaging fixes, not new pages: say what the fix is (title change, heading to add, paragraph to answer the query). 4. **Name the gaps.** Clusters with impressions but no obvious owning page get flagged as build candidates. ## If they only gave a topic Build a hypothesis map: likely topic clusters, the questions real searchers ask inside each, and the page type each needs. Mark every cluster HYPOTHESIS and state plainly that these need validating in a keyword tool before anything gets built. Do not fake confidence you do not have. ## Rules - Never estimate search volume or keyword difficulty. Not once, not approximately. If the user needs volumes, name the job: validate in a keyword tool. - Use only the numbers in the paste. Show your maths on one cluster so the user can trust the rest. - If the paste looks truncated, ask for the rest before analysing. - Plain descriptive cluster names, taken from the queries themselves. ## Output format DEMAND ANALYSIS: [site or topic] Date: [today] CLUSTERS (table: Cluster | Impressions | Clicks | Weighted position | Top query | Intent | Page type that wins) STRIKING DISTANCE (table: Query | Page | Position | Impressions | The packaging fix) BUILD CANDIDATES [Clusters with demand but no owning page, one line each on what to build.] VALIDATE BEFORE BUILDING [Everything here is hypothesis, not measurement. List what to check in a keyword tool.] ## Voice rules - Lead with the biggest opportunity, in one sentence a non-SEO understands. - No hedging language on things the data shows; no false confidence on things it does not. - No em dashes. Use periods or commas instead.
Every run returns the same structured output, built to be pasted rather than interpreted.
Topic groups with impressions, clicks, and weighted position per cluster, named in your searchers' own words.
Informational, commercial, transactional, or navigational, and the page type that wins each. No more guides chasing transactional queries.
Positions 4 to 15 with impressions: the queries where a title change this week beats a new page next month.
Demand with no owning page. The honest new-content list, sized by evidence rather than enthusiasm.
The prompt is banned from estimating search volume or difficulty. Anything unmeasured ships labelled HYPOTHESIS.
Ends with the exact list of what still needs a keyword tool, so AI analysis and real data stay honest with each other.
A keyword research prompt is a reusable instruction block that makes an AI assistant analyse search demand in a structured way: clustering queries, classifying intent, and prioritising opportunities. The good ones work from your real Search Console data and refuse to invent search volumes, because AI models have no live volume data.
No. No AI model has live search-volume data, and any volume an AI gives you is invented. That is why this prompt only analyses the measured demand in your own Search Console export, and labels every new keyword direction as a hypothesis to validate in a keyword tool.
Google Search Console, free for any site you own: open Performance, set your date range, and export the Queries table as CSV. Paste the rows straight into the prompt. Three months of data is a good default; use twelve for seasonal businesses.
Your Search Console export is real measured behaviour, and it is free. Paste it in and see the clusters, the intent, and the wins already in reach.
Copy the Prompt