Free AI SEO Keyword Research: A No-Cost Method That Uses Real Data

The method has three inputs and one output. The inputs are your Google Search Console query export, your Bing Webmaster Tools AI Performance export, and a general AI assistant. The output is a short list of topic clusters, each with the buyer questions behind it.

The data does the demand work. The assistant does the sorting work. Keep those jobs separate and the method holds up.

Search Console's performance report lists the queries that earned your site impressions in classic Google Search, with clicks, click-through rate and average position. That is real demand from real people, filtered to queries where Google already considers you relevant.

Bing Webmaster Tools adds something no other free source offers. Its AI Performance report, in public preview since February 2026, shows grounding queries: the phrases the AI used when retrieving content it then cited in Copilot, Bing AI summaries and some partner integrations. Microsoft notes that the grounding-query data is a sample. As of the latest evidence we have, it has no API, so you export it from the interface.

The assistant then groups hundreds of raw queries into themes a human can act on. That's the part generators are good at. For why generated keyword lists are ideas rather than demand, see AI keyword generators: ideas, not demand data.

Each free source answers a different question. None of them answers the one most teams want answered: which queries put you inside an AI answer on every engine.

The biggest gap is the retrieval query. OpenAI's help page says ChatGPT search typically rewrites your query into one or more targeted queries before sending them to search providers. Google's Gemini API documentation says grounding automatically generates one or more search queries. So the words a buyer types and the words the engine searches are often different. A keyword list built from typed queries is a proxy for an input you can't observe.

Organic rank is a partial proxy too. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the first 10 results, in a March 2026 study. That means ranking matters, and also that most citations came from elsewhere. Treat your Search Console positions as one signal, not a forecast of AI citations.

For the deeper difference between typed keywords and what engines actually use, read keywords vs prompts in AI search.

It helps to split AI search into four stages: what the model learned in training, whether crawlers can reach your page, whether your page is retrieved for a query, and whether the model uses and cites it.

Keyword research only touches the third stage, and only indirectly. Here's the honest split.

In your control

• Which topics your site covers, and how directly each page answers the buyer question behind it.
• Stating specific facts (a price, a number, a one-line definition) on pages that crawlers can fetch, so an engine has a sentence it can quote.
• Access: whether AI crawlers can fetch the page, and whether the key content is in the HTML before JavaScript runs.

Not in your control

• The queries the engine generates from a buyer's prompt.
• Where your page sits in the model's list of retrieved sources.
• Whether the engine searches at all. Retrieval is per query and discretionary on every major platform.

A critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) found topical relevance and context position were the most reproducible levers, and that no technique showed a stable, cross-platform effect on discoverability. Topical relevance is the part keyword research can inform. Position is not something you can set.

Plan for an afternoon the first time. After that, repeat it each quarter.

1. Export Search Console queries. Use the last three months of the performance report. Keep query, impressions, clicks and position.
2. Export Bing grounding queries. Open the AI Performance report in Bing Webmaster Tools and export or copy the grounding queries. If you haven't verified your site in Bing, do that first. It's free.
3. Note your AI Overview footprint. Open the Search Console generative AI report and list which pages earn AI Overview or AI Mode impressions. This tells you which topics Google's AI already surfaces you for, without queries attached.
4. Merge and deduplicate. Put everything in one sheet with a column for source. Drop branded queries into their own tab.
5. Cluster with an AI assistant. Paste the list and ask it to group queries by underlying topic and buyer intent (learn, compare, buy, fix). Ask it to name each cluster in plain words and to flag queries it can't place. Tell it not to add volumes or new keywords.
6. Check the clusters yourself. Merge clusters that describe the same buyer need. Split ones that mix intents.
7. Write buyer questions per cluster. Turn each cluster into two or three questions a real buyer would ask an AI assistant. The prompt research method covers this step in detail.
8. Map clusters to pages. For each cluster, name the page that should answer it. Gaps become your content plan.
9. Check which clusters trigger AI Overviews. Search a sample of queries from each cluster, or use the free check for which keywords trigger AI Overviews.

The free checker on this site can then run up to three of your buyer questions and show whether AI engines name you. It's a snapshot, not ongoing tracking, but it closes the loop between your clusters and real answers.

The assistant is a sorting tool. It's good at spotting that "cost of invoicing software for freelancers" and "cheap invoice app for contractors" are the same need. It's poor at anything that needs data it doesn't have.

A few rules keep its output useful:

• Give it your data, not a seed word. Clustering your exported queries grounds the work in real demand. Asking it to brainstorm from a single keyword produces plausible lists with no evidence behind them.
• Ban numbers. If the assistant returns search volumes or difficulty scores, discard them. It has no access to a demand dataset, so those figures are made up.
• Ask for its reasoning. Have it explain why each query belongs in its cluster. Wrong reasoning is easy to spot and fix.
• Keep raw queries attached. Every cluster should list the source queries under it, so you can trace any decision back to real data.

Say you run a hypothetical bookkeeping firm in Leeds. Your Search Console export holds 400 queries. The assistant might group them into twelve clusters, such as year-end accounts, VAT registration and switching accountants. That's a plan you can act on. A list of 400 separate keywords is not.

Most errors come from treating free or generated data as something it isn't.

• Trusting invented volumes. AI assistants and many free generators output numbers that look like demand data. They aren't. Use first-party impressions instead.
• Collecting every phrasing. Google's guide to its generative AI features says you don't have to worry about capturing every long-tail variation or every way someone might seek your content. Cover the topic well once.
• Rewriting pages for AI. The same Google guide says you don't need to write in a specific way for generative AI search. Studies that tested citation-focused rewrites found them ineffective or harmful.
• Tracking an AI rank. Google's generative AI report shows impressions only, so there is no position to check a rank against. Free AI rank tracking explains what to measure instead.
• Reading one engine as all engines. Bing grounding queries describe Copilot and Bing surfaces. They say nothing reliable about ChatGPT, Gemini, Claude or Perplexity.
• Ignoring access. A perfect cluster plan doesn't help if AI crawlers can't reach the page, or if the answer only appears after JavaScript runs. Check access before you write.
• Testing logged in. Memory and custom instructions can change what an assistant searches for. Test buyer questions logged out, with memory off.

If you want to check access and answers together without paying, the do-it-yourself AI visibility audit walks through it.

Is there a free AI keyword tool that shows search volume? Not a trustworthy one built on AI alone. A language model has no demand dataset, so any volume it prints is invented. The most reliable free demand data is your own: impressions in Google Search Console and grounding queries in Bing Webmaster Tools. Those only cover queries where you already appear, but every number in them is real. Use an AI assistant to organise that data, not to produce it.

Can Search Console show which queries trigger AI Overviews for my site? Not at query level, as of October 2026. Google's generative AI report in Search Console shows impressions for AI Overviews and AI Mode by page, country, date and device, with no queries, clicks, CTR or position. You can see which pages earn AI impressions, then cross-check by searching queries from your classic performance report to see which ones show an AI Overview.

Why are Bing grounding queries worth the effort? They are the only first-party view of the queries an AI system actually ran when it retrieved and cited content. Every other source shows what people typed. Engines like ChatGPT and Gemini rewrite prompts into their own searches, so typed queries are a proxy. Bing's data covers Copilot, Bing AI summaries and some partners, it's a sample, and it says nothing about other engines, but it's real retrieval input.

Should I optimise pages for the exact prompts buyers type into ChatGPT? Use prompts to understand needs, not as strings to match. OpenAI says ChatGPT search typically rewrites a query into one or more targeted searches, so the exact wording rarely reaches the search provider. Google says you don't need to cover every phrasing for its AI features. Build pages that answer the underlying question directly, with specific facts stated plainly in the HTML.

How often should I redo this keyword research? Quarterly is a sensible rhythm for most small teams. Three months of Search Console data is enough to show stable patterns, and Bing's AI Performance report has been expanding since its February 2026 preview. Rerun the clustering when you launch a product, enter a new market, or notice a cluster's impressions shift. Keep earlier exports so you can compare clusters over time.