Generative Engine Optimisation: What the Evidence Says

Generative engine optimisation (GEO) means shaping your content so AI search tools are more likely to use it in their answers. You will also see it called AI search optimisation or SEO for AI. Google says there is nothing special you need to do for its AI features, and the best-known study of GEO methods is from 2023 and tested engines of that time.

This post covers what GEO is, what that study found, what Google says, what we think is sensible, and what we saw in our own test. It sits under our AI SEO guide, which covers the wider picture. We don’t promise citations, because nobody can.

What is generative engine optimisation?

Traditional SEO aims to get your page ranked in a list of links. GEO aims to get your content used inside an AI-written answer, either as a source or as a named recommendation.

The terms overlap heavily. “Generative engine optimisation”, “AI search optimisation” and “SEO for AI” are mostly used for the same job: making your site easy for AI search products to find, understand and use. If you want a service rather than a definition, our AI SEO agency page explains what we do, and AI SEO in Cornwall covers the local angle.

What the Aggarwal study found

The term GEO comes from Aggarwal et al. (arXiv 2311.09735). The authors built GEO-bench, a set of 10,000 queries, and tested a range of content methods on a GPT-3.5-turbo-based engine and on Perplexity.

They report that adding citations, quotations and statistics improved visibility by over 40%, while keyword stuffing gave little or no improvement.

The caveats matter:

• The engines are old. The study tested systems from 2023 and 2024, not today’s ChatGPT, Gemini or Claude.
• It is the authors’ own setup. Results inside a research benchmark don’t tell you what happens on your queries.
• “Visibility” is their measure. It is not the same as being recommended to a buyer, or getting an enquiry.

So the study is a useful hint about what to try, not proof of what will work for you. We would read its headline figure as “specific, sourced writing seemed to help in that test”, and no more than that.

It is also worth noticing what the study did not test: whether a page that is more visible in an AI answer leads to more enquiries. Our own test, below, shows why that gap matters.

What Google says

Google’s AI features documentation says there are no extra requirements or special optimisations needed to appear in AI Overviews or AI Mode. It also says no special schema.org structured data is needed.

Its AI optimisation guide goes further: structured data isn’t required for generative AI search, and you don’t need special AI text files, markup or Markdown.

That includes llms.txt. It is a proposal for a markdown file that helps agents use a site, and no provider has officially said its search uses it. Anyone selling it as a ranking factor is selling more than the documentation supports.

One part of Google’s page is practical. The snippet controls nosnippet, data-nosnippet, max-snippet and noindex apply to these AI features. If you use them, you are also limiting what AI features can show from your pages.

What helps in practice

Two things are worth keeping apart here: what the study found, and what we think.

From the study: adding citations, quotations and statistics improved visibility in its tests, and keyword stuffing didn’t.

Our view, not a finding:

• Answer the question first. Put the plain answer at the top, then explain it.
• Say where facts come from. Link to the primary source, not a blog repeating it.
• Be specific about who you are and where you work. A model can’t guess which towns you cover or which services you sell.
• Get the basics right. Pages need to be indexable, and the search bots you want should be allowed in robots.txt. Our guide to AI crawlers and robots.txt covers that part.

None of this is exotic. It is the same discipline as good SEO, written down more carefully. A page that states its answer plainly, shows its sources and says who wrote it is easier for a person to trust, and our view is that the same qualities are what a model has the best chance of using well. Our SEO Cornwall page covers the basics for local businesses.

What we saw in our own test

On 2 October 2026 we ran 10 UK buyer prompts, such as “best digital marketing agencies in Cornwall”, across ChatGPT, Claude, Perplexity and Gemini. Outcome was named in 4 of 40 answers: Perplexity 2, ChatGPT 1, Claude 1 and Gemini 0.

In one Gemini answer, our restaurant marketing article was cited as a source 16 times, yet Gemini did not name us as an agency to hire. Being cited and being recommended are different things.

This is a small snapshot from one business on one day, not a study. It can’t tell you how AI engines behave in general. We share it because it shows the problem honestly.

What to do next

1. Read the AI SEO guide for the full picture across Google, OpenAI, Perplexity and Anthropic.
2. Check your robots.txt, using our post on AI crawlers and robots.txt.
3. Rewrite your most important pages so each opens with a direct answer and sources its facts.
4. Run a few real buyer prompts yourself and note what is named and what is cited. Treat it as a snapshot.

If you’d rather have someone do this for you, that is what our AI SEO agency service is for. We don’t offer a guarantee or refund, because no honest agency can promise what an AI engine will say.