What Is Generative Engine Optimization (GEO)?

Where the term GEO came from

GEO has a birth certificate. In November 2023, six researchers led from Princeton posted a paper called "GEO: Generative Engine Optimization", later revised and presented at a major academic conference. They defined a generative engine as a search tool that retrieves sources and then writes a single answer from them, and they asked a practical question: can a publisher change its content so the engine uses more of it?

Everything sold since under the labels GEO, AEO or LLM SEO borrows from that question. The honest way to use the term is as a name for the work of getting included in synthesized answers, not as a separate discipline with its own rulebook.

What the paper tested and what won

The authors built GEO-bench, a set of 10,000 queries drawn from nine sources and spanning 25 domains. About 80% of the queries were informational. They rewrote source pages nine ways and measured how much of the generated answer drew on the rewritten page.

The headline is "up to 40%" better visibility. That is the best case across tactics and queries, not an average. On Perplexity, the one deployed engine the authors also tried, the best tactics showed gains of up to about 37% on one of their measures, and keyword stuffing did worse than doing nothing.

What the paper cannot tell you

The research is useful and old, and both facts matter.

• The engine was a 2023 model. The main experiments gave gpt-3.5-turbo the top five Google results and asked it to write an answer. Today's products use different models, different retrieval and, in Google's case, query fan-out.
• Two engines in total. One was the research setup, one was Perplexity.
• Visibility was measured by the authors' own scoring. Their measures count how much of an answer relies on a source and how prominent it is. They are reasonable, but not what a customer sees.
• Rankings were out of scope. The authors say they did not measure effects on search rankings because ranking algorithms are a black box.
• Results varied by domain. A tactic that helped for one topic did less for another.

The authors themselves say methods may need to adapt as engines change, the way SEO did.

What GEO adds to SEO in practice

Google's own guide says optimizing for its generative features is optimizing for Search, and that you do not need special files, markup or rewritten content. Taken together with the research, that leaves three practical additions.

1. Specific, sourced content. Put the number, the date, the named source and the exact condition into the sentence. This is where the paper's winning tactics and Google's call for non-commodity content point the same way.
2. Access for bots outside Google. ChatGPT, Claude and Perplexity each run search-related crawlers that you can allow or block separately. How to rank higher in ChatGPT search covers the OpenAI side.
3. Measurement that accepts variance. AI answers change between runs. Ahrefs notes that AI Overviews are probabilistic, with citations and related queries shifting from one query to the next.

A worked example

This example is illustrative. The clinic and the figures are invented to show the edit.

Before: "We offer a wide range of high-quality physiotherapy services at competitive prices for patients across Limassol."

After: "A first physiotherapy appointment at our Limassol clinic lasts 45 minutes and costs EUR 55. It includes an assessment, a written plan and, where needed, a referral letter for your insurer. Treatment sessions after that are 30 minutes."

The second version gives an engine a quotable passage with a price, a duration and a list of inclusions. If the figures are real and you can cite them, a reader and an assistant both get what they came for. The first version could belong to any clinic in any city.

Turning the nine tactics into edits for a service business

Most of the tactics in the paper need translating before you can use them. This is our reading of how each one applies to a local or professional services site. It is an interpretation, not something the paper proves.

The consistent thread is evidence. The three best tactics each make a claim easier to check, and the worst one adds words without adding anything to check.

Claims you will meet, and the verdict

• "GEO needs its own special markup." Not supported. Google says no special schema is needed for its AI features.
• "Optimize for each assistant separately." Partly sensible for crawler access, but the content work is shared.
• "A tool can score your content for AI citation." Scores can flag missing facts and weak structure. No tool can know how a given engine will rank a page, and Google states that third-party tools have no access to its internal systems.
• "GEO delivers results in weeks." No source we found supports a timeline. Treat dates as red flags.

Next step

Pick your five most valuable pages and rewrite one paragraph on each so it contains a specific figure, a named source or a precise condition. Then read GEO vs SEO vs AEO to see which of the other labels you can safely ignore, or look at the full AI search optimization guide if you want the whole map. Our AI search optimization service applies the same steps to your site.