AI search optimization: getting cited, not just ranked

What is Generative Engine Optimization (GEO)?

Generative Engine Optimization is the practice of optimising content and brand presence so AI answer engines cite you inside the answer they generate. The term was introduced in a 2024 research paper by Aggarwal and colleagues at Princeton University and IIT Delhi, published at KDD 2024, which tested nine content modifications against generative visibility metrics.

It differs from SEO in the mechanism, not the fundamentals. SEO aims to rank a document; GEO aims to have a specific passage retrieved, synthesised and attributed. That makes the unit of optimization the passage rather than the page, and the unit of success the citation rather than the impression.

Google's own position is that optimising for generative AI search is still SEO and that the same fundamentals apply — its AI optimization guide explicitly states you do not need llms.txt files, special markup or content chunking. That is accurate. Google also states its generative features are grounded in the same core ranking and quality systems. The practical difference lies in other engines' retrieval behavior, and in the fact that a document read in full by a human may never be lifted at all.

How do AI answer engines choose sources?

AI answer engines use retrieval-augmented generation: they search an index for relevant passages, pass those to a language model, and have it synthesise an answer with citations. Four consequences follow, and each one changes how you should write.

1. Passages must be self-contained. A retriever lifts a chunk of text without the surrounding page. If your answer depends on context from three paragraphs earlier, the lifted version is wrong or incomplete — and a model will not reconstruct it for you.
2. Specificity beats comprehensiveness. "AI Overviews launched in the United States on 14 May 2024" is a retrievable fact. "AI Overviews launched recently" is not. Anchored facts with names, numbers and dates are what gets cited.
3. Freshness is weighted, but not decisive. Engines prefer current sources for current questions, and a page cited last month that has since gone stale tends to be replaced by a fresher competitor.
4. Heading phrasing matters. People ask assistants full questions. A heading phrased as the literal question — "How do I optimize for AI Overviews?" — is far more likely to be matched than "Optimization Strategies".

What actually raises your citation rate?

This is where most GEO advice goes wrong, so it is worth separating what has evidence behind it from what is merely repeated. The 2024 GEO paper tested nine modifications and measured the effect on generative visibility.

Two results in that table are counterintuitive and worth internalising. First, the largest gains went to pages that were not already ranked first — the paper reported roughly +115% visibility from source citations for a rank-five page, while the top-ranked page could actually lose ground. Adding evidence is a leveling play, not a winner's prize. Second, keyword stuffing, the classic 2014 tactic, was actively harmful.

The strongest predictor is not a technical factor

Ahrefs analyzed 75,000 brands and reported a correlation of 0.664 between branded web mentions and appearing in Google AI Overviews, against 0.326 for Domain Rating. In other words, how often and how visibly your brand is discussed elsewhere on the web correlates roughly twice as strongly as your link profile. That is an off-site, authority problem dressed up as an on-site one — and it is why GEO work that stops at the website has a ceiling.

Do I need an llms.txt file?

No, and it is important to be clear about that. Google's official AI optimization guide states directly that you do not need to create new machine-readable files, AI text files, markup or Markdown to appear in its generative features.

llms.txt is a community specification proposed by Jeremy Howard in September 2024. It is a plain-text index at your site root listing your most important pages with one-line descriptions. It is cheap to publish, it does no harm, and some non-Google agent ecosystems look for it.

But it is a discovery aid, not a ranking mechanism. Treat it as insurance with a very low premium rather than a strategy. We publish one on this site and we would not spend client budget on anything more elaborate. The full treatment, including a working example and the honest verdict, is in what is llms.txt?

Which AI crawlers should I allow?

There are now several distinct AI crawlers with different purposes, and treating them as one category is the most common robots.txt mistake. The key distinction is between crawlers that fetch live pages to answer a user's question, and crawlers that gather content for model training.

Our recommendation is to allow retrieval crawlers and to make a deliberate, documented decision on training crawlers. The commercial argument for allowing training is straightforward: you gain citations in future models, you are not paywalled, and you have no content to monetise directly. The argument against is if your content is proprietary or you consider the category valuable enough to compete with. Whichever you choose, write down the reason — a decision with a stated rationale survives the next quarterly argument; one without gets relitigated every few months.

How do you measure AI visibility?

AI visibility cannot be measured with the tools you already have. Search Console does not report it, GA4 does not know about it, and rank trackers do not model it. Measuring it requires sampling the answers themselves.

The method that works is a citation ledger. For a defined set of priority queries — typically the 30 to 100 questions your customers actually ask — sample the answers across Google AI Overviews, ChatGPT Search, Perplexity and Gemini on a recurring schedule, and record two separate numbers.

Track them on a rolling window rather than a single day, because answers are non-deterministic and the cited set rotates. Tools such as Profound, Otterly, Athena and Goodie automate the sampling; if you would rather not pay for one, the sampling is straightforward to script. The discipline matters more than the tool — a spreadsheet updated weekly will beat an expensive dashboard nobody looks at.