Why keyword research doesn't work for GEO — and what actually works

The myth of the universal keyword

Since the rise of Google, keyword research has become second nature for every digital marketing professional. Identify the queries your prospects type, assess their volume and competitiveness, create optimized content around those terms — this logic has dominated online visibility for two decades.

It still works. For Google.

The problem arises when you apply this same logic to AI engines — ChatGPT, Perplexity, Gemini, Claude, Grok — assuming they work the same way. They don't. And conflating the two disciplines produces strategies that optimize for the wrong system.

How Google uses keywords

Google is fundamentally an indexing and ranking engine. Its algorithm analyzes web pages, indexes them by content, and ranks them in response to specific queries. A keyword in an H1 title, consistent semantic density, a volume of backlinks pointing to a page — all signals Google interprets to decide which page deserves position 1 for a given query.

Keyword research maps the space of these queries: what terms your prospects type, how often, with what buying intent. A BOFU strategy — Bottom of Funnel — targets high commercial-intent queries where the user is ready to buy or compare. It's a rigorous and effective approach in the Google universe.

How LLMs select brands — a radically different logic

Large language models don't rank pages. They generate answers.

This distinction seems trivial. It is fundamental.

When someone asks ChatGPT "what's the best HR management solution for a 50-person company," ChatGPT doesn't consult an index of pages ranked by keyword relevance. It generates a synthetic answer drawing on two main sources: patterns from its training corpus, and in web-enabled versions, pages it retrieves in real time.

In both cases, the selection mechanism isn't the keyword. It's the entity.

An LLM reasons in terms of entities — brands, organizations, people, concepts — and their relationships. When it generates a recommendation, it selects the entities it recognizes as reliable, consistent, and authoritative in the requested category. Not those with the best score on a keyword.

The three real selection criteria of LLMs

What AI engines actually evaluate comes down to three dimensions that keyword research doesn't directly address.

The first is entity recognition. Does the AI know your brand? Does it describe it accurately and consistently across all the platforms where it has encountered it? A brand whose description varies across sources — "digital agency" on its website, "transformation consulting firm" on LinkedIn, "tech startup" on Crunchbase — sends an ambiguity signal that LLMs interpret as unreliable.

The second is source authority. Which independent platforms mention your brand? G2, Clutch, Trustpilot, Wikipedia, recognized industry publications — these sources carry considerable weight in LLM selection because they represent third-party validation that the model considers objective. A blog post optimized for the keyword "HR software SME" doesn't produce the same signal as a complete, reviewed listing on G2.

The third is content citability. Is your content structured so that an LLM can extract a direct answer from it? A page that opens with "We're passionate about HR innovation" will never be cited. A page that opens with "Our solution manages leave, payroll, and reviews for companies between 20 and 200 employees, with native integration with major HRIS platforms" will be extracted and cited — because it directly answers a probable question.

What the BOFU strategy misses in the GEO context

The BOFU strategy works for Google because it targets users in the decision phase who type specific queries. It generates qualified traffic and converts well.

Applied to GEO, it creates two structural problems.

First: LLMs don't respond to keywords — they respond to intents expressed in natural language. "Best HR software SME" is a Google query. "I have a 50-person company and I'm looking for a solution to automate my HR management — what do you recommend?" is an LLM query. Both cover the same need but the selection logic is different. Optimizing for the short keyword doesn't prepare you to be cited in the long answer.

Second: a keyword-centered content strategy typically produces pages that talk about your offering. A GEO strategy produces pages that answer your prospects' questions. The difference in angle is total — and it shows immediately in how an LLM processes the content.

What to do instead

This doesn't mean keyword research is useless in a GEO strategy. It remains relevant for identifying the topics around which to produce citable content. But it must be supplemented with three approaches that classic SEO doesn't address.

First, map the questions your prospects ask AI engines — not the keywords they type into Google. These two spaces partially overlap but diverge on comparison queries, recommendations, and contextual questions. Test directly on ChatGPT, Perplexity, and Gemini the 20 most likely questions from your ideal customers. Note who appears. Analyze why.

Second, build your entity before optimizing your content. If LLMs don't recognize your brand as a consistent, reliable entity, no content strategy will produce lasting results. Consistency of description across your website, third-party profiles, and editorial mentions is the prerequisite for everything else.

Third, structure your content for extraction, not for clicks. Effective GEO content is direct, factual, organized as answers to explicit questions. It doesn't try to generate clicks — it tries to get cited. This is a complete paradigm shift from the SEO conversion logic.

The red flag to watch for

When an agency offers you a "GEO strategy" based primarily on keyword research and content production optimized for Google queries, ask one simple question: how do you measure my actual appearance in ChatGPT, Perplexity, Gemini, Claude, and Grok answers?

If the answer is vague, if it references Google rankings rather than citations in AI answers, if it doesn't include a measurement tool with documented scores — you're dealing with SEO rebranded as GEO.

GEO is measured differently from SEO. It's optimized differently. And it's sold honestly — with starting scores, target scores, and a documented method in between.