AI Keyword Research: Useful for Clustering, Limited as Demand Data

AI keyword research means handing seed terms to a language model and asking it to do the slow parts of the job. It does three of them well.

• Expansion. Give it "dental implants Austin" and it returns dozens of related phrasings, questions and adjacent topics in seconds.
• Clustering. Paste in a few hundred keywords from a search tool and it groups them by shared meaning, which is tedious by hand.
• Intent labelling. It can tag each phrase as informational, commercial or navigational, and suggest which page type fits.

The limit is just as clear. A model generating keywords is predicting plausible text. It has no access to how often anyone searches a phrase, and it can't tell you whether a phrase it invented is typed by anyone at all. Demand numbers still come from a search data provider, not from the model. Our guide to AI keyword generators covers that split in more detail.

So use AI for breadth and structure. Use measured search data for prioritising. Keep the two columns separate in your spreadsheet so nobody mistakes a generated idea for a validated one.

Classic keyword research assumes the words a person types are the words the engine searches for. AI search engines break that assumption.

As of October 2026, the documented behaviour looks like this:

• ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to its search providers, according to OpenAI's help centre. OpenAI also says saved memories may be used when it rewrites a query.
• Google AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics.
• The Gemini API's grounding with Google Search automatically generates one or more search queries and runs them itself.

This matters because retrieval runs on the rewritten queries, not on the buyer's prompt. A page that matches the exact prompt wording may never be fetched, and a page that matches a subtopic the engine invented may be.

Retrieval also isn't guaranteed to happen. ChatGPT decides per query whether to search. AI Overviews often don't appear at all. When an engine answers without searching, it relies on what the model learned before its knowledge cutoff, and no keyword work reaches that answer this quarter.

The practical upshot: any list of keywords or prompts you build is a proxy. It approximates an input that only the engine sees. That's still useful, as long as you treat it as a sample rather than a map. If it helps to see the contrast laid out, read answer engines vs search engines.

A common reflex is to use AI to generate hundreds of near-identical phrasings and write a page, or a paragraph, for each. Google has said directly that this isn't necessary.

Google's guide to optimizing for generative AI features says you don't need to write in a specific way for generative AI search, and that "you don't have to worry that you don't have enough 'long-tail' keywords or haven't captured every variation of how someone might seek content like yours."

The research points the same way. The original GEO paper (Aggarwal et al., 2023) found that keyword stuffing decreased its visibility metric, scoring 17.8 against a 19.5 baseline. That study ran on a lab engine with sources already loaded into context, so treat it as a direction, not a forecast for live engines. Our comparison of lab GEO studies and live engines explains why that caveat matters.

What has more support is topical relevance. A 2026 critical survey of GEO studies (arXiv 2607.14035) found that topical relevance and context position are the most reproducible levers, while generic heuristics transfer poorly. Even that finding comes from pages already retrieved, so it is moderate evidence, not a guarantee. The question does staying on topic help AI choose my page goes through it.

The move: use AI expansion to discover distinct buyer problems, then collapse the variants. One clear page per problem beats twenty pages per phrasing.

Keyword research feels like control because you choose the list. Most of what decides whether you appear in an AI answer sits elsewhere.

The left column is mostly access work, and the evidence behind it is the strongest available. Most AI crawlers other than Googlebot don't execute JavaScript, so text that appears only after scripts run may be invisible to them. A robots.txt block on OAI-SearchBot removes you from ChatGPT search answers, whatever your keywords are.

Stating a specific fact on a fetchable page is the one on-page step with a positive experimental signal: it gives the engine a sentence it can quote. Keyword density is not on the list.

Here's a sequence that uses AI where it helps and keeps you honest where it doesn't.

1. Write down the buyer problems first. Ask sales or support what people ask before they buy. Five to fifteen real problems beat five hundred generated phrases.
2. Expand each problem with AI. Ask a model for related questions, comparisons and objections. Mark everything it produces as "unvalidated."
3. Pull demand data. Run the expanded list through a search data tool. Note which terms show real volume and which show none. Our guide to AI keyword research tools lists what to look for.
4. Cluster and collapse. Ask AI to group the validated terms by meaning. Each cluster becomes one page or one section, not a page per variant.
5. Turn clusters into buyer prompts. Write the questions a buyer would type into ChatGPT or Perplexity. The prompt research method walks through this step.
6. Check which terms trigger AI Overviews. Not every query shows one, and the AI Overview keyword check shows which of yours do.
7. Fix access before content. Confirm retrieval crawlers can fetch the page, the key facts are in the raw HTML, and the page isn't set to noindex or nosnippet.
8. Write the page. Answer the problem directly, state the specific facts, and stop when the question is answered. Google says there is no ideal page length.
9. Look at real answers. The free checker on this site shows whether AI engines name you for up to three buyer questions. It's a snapshot, so repeat it after changes rather than reading one result as a trend.

Since AI engines rewrite prompts, the most valuable data is the queries they actually ran. Very little of it is published.

• Bing Webmaster Tools has an AI Performance report that shows grounding queries: the phrases the AI used when retrieving content cited in Copilot and Bing AI answers. Microsoft notes this data is a sample. As of August 2026, the Bing Webmaster API documentation did not cover it, so you read it in the interface.
• Google Search Console has a generative AI performance report for AI Overviews and AI Mode. As of the evidence we hold, it shows impressions only, with no clicks, click-through rate or position.
• OpenAI, Anthropic and Perplexity APIs return the sources a search retrieved, but they don't show you what happens inside the consumer apps for your buyers' prompts.

If you have a verified Bing property, export the grounding queries and compare them with your keyword list. The gap tells you how far your proxy sits from the engine's real input. Everything else, including any prompt tracker, remains a sample of prompts someone chose.

These are the errors that waste the most time in AI keyword research.

• Treating generated keywords as demand. A model can invent a phrase nobody types. Validate before you build pages.
• Writing a page per variant. Google says you needn't capture every variation, and thin near-duplicates at scale can fall under Google's scaled content abuse policy.
• Mass-generating pages from the list. Google defines scaled content abuse as many pages generated to manipulate rankings without helping users. A public 2026 experiment by OtterlyAI saw two sites of roughly 1,000 AI-written posts each removed from Google's index with no manual action reported.
• Optimising for the exact prompt wording. ChatGPT rewrites prompts and Google fans them out. Cover the problem, not the string.
• Reporting an "AI rank" for a keyword. Google exposes no position data for its AI features, so there's nothing to check such a number against.
• Skipping access checks. A perfect keyword map does nothing if a firewall returns 403 to OAI-SearchBot or your pricing loads only after JavaScript runs.
• Hiding keywords for AI crawlers. Hidden text and cloaking aimed at AI bots are manipulation. They risk penalties and can mislead the engines that do read them. Don't do it.

For a wider view of where AI tooling helps across SEO tasks, see AI SEO software categories.

Can ChatGPT replace a keyword research tool? Not on its own. ChatGPT can expand seed terms, cluster lists and label intent quickly, but it has no search volume data and can suggest phrases nobody types. Use it alongside a search data provider. Let the model generate breadth and structure, and let measured demand decide which clusters get pages first.

Are keyword lists useless for AI search? No, they're proxies. AI engines rewrite prompts into their own searches, so you never see the exact retrieval queries. A keyword list still captures the problems buyers have and the words they use. Treat it as a sample of likely inputs, and where you can, compare it with real grounding queries from Bing Webmaster Tools.

Should I optimise content for the prompts people type into AI? Optimise for the problem behind the prompt, not the prompt string. OpenAI says ChatGPT typically rewrites a query into one or more targeted queries, and Google can fan a query out across subtopics. A page that answers the underlying question clearly, with specific facts in plain HTML, fits more of those rewritten searches than a page tuned to one phrasing.

Does keyword density matter for AI citations? There's no evidence it helps, and some that it hurts. The original GEO paper found keyword stuffing lowered its visibility metric compared with the baseline, though that was a lab setup with sources already in context. Google's generative AI guide says you don't need to write in a special way for AI search. Topical fit has more support than repetition.

How do I know which keywords trigger AI Overviews? Check them directly. AI Overviews don't appear on every query, and appearance varies by location and language. Run your priority terms through a checker that records whether an overview showed and which sites it cited. Remember Google requires a page to be indexed and snippet-eligible before it can appear as a supporting link.