GEO and AI search: what Google actually says about optimising for it

Over the past two years a new service category has appeared: GEO, generative engine optimisation, or AEO, answer engine optimisation. In proposals it looks like a separate discipline with its own tools, its own metrics and its own fee. Before buying it, it is worth reading what Google itself publishes on the subject. Its documentation is unusually direct here.

In short

• Google's documentation states that from its perspective, optimising for generative AI search is optimising for the search experience, and thus still SEO.
• Generative features are rooted in the same Search ranking and quality systems, not a separate "AI index".
• Google explicitly lists things you do not need to do: llms.txt and other special files, chunking content, rewriting for AI, chasing inauthentic mentions, and over-focusing on structured data.
• Eligibility comes from the basics: the page must be indexed, meet the technical requirements and be eligible to show with a snippet.
• There is an official measurement source — the generative AI performance report in Search Console.

What Google says plainly

Google's guide to optimising for generative AI features contains a dedicated passage about these terms. Its substance: "AEO" and "GEO" describe work focused on visibility in AI search experiences, but from Google Search's perspective, optimising for generative AI search is optimising for the search experience, and thus still SEO. The same passage suggests evaluating third-party GEO advice or services against Google's general guidance on assessing outside SEO advice.

The technical reason is straightforward. Generative features use the same index and the same ranking and quality systems. An answer is assembled by two mechanisms: retrieval from the index to ground the response, and query fan-out, where the model generates additional related queries and fetches more results. Both depend on what is already in the index. There is no separate "AI ranking" that could be optimised on its own.

Five things Google says you do not need to do

This list is useful beyond the technical detail. Almost every item on it is something being sold as a service.

• llms.txt and other special files. Google states you do not need to create them because Google Search does not use them — they neither help nor harm. Keep one for other systems if you like, but not as a Google visibility measure.
• Chunking content. Google states there is no requirement to break content into tiny pieces, and no ideal page length.
• Rewriting content for AI systems. Google states you do not need to write in a particular way: its systems understand synonyms and meaning, so you need not capture every phrasing variant.
• Chasing inauthentic mentions. Google states that seeking such mentions is not as helpful as it may seem, because the same quality and spam systems apply.
• Over-focusing on structured data. Google states structured data is not required for generative AI search and there is no special schema. It is still worth using for other reasons, such as rich results.

Where the real risk sits

The most interesting part is not what fails to work but what can actively hurt. Google's documentation states that creating separate content for every possible query variation — including fan-out queries — primarily to manipulate rankings or generative responses violates its scaled content abuse spam policy. So the popular GEO tactic of covering every phrasing with its own page is not merely ineffective in the long run; it can fall foul of policy.

The second risk is paying for work nobody can measure. Google separately warns against third-party tools that promise ranking success or claim to use "internal" Google metrics, because no third-party tool has access to its internal ranking or AI systems. If a GEO proposal rests on such a metric, that metric belongs to the supplier, not to Google.

What did genuinely change

It would be wrong to conclude that nothing changed. Three things did, and all three are useful.

• Measurement. Search Console has a generative AI performance report, so visibility in AI features has stopped being guesswork. This is the first time the subject can be discussed with data rather than intuition.
• The content bar rose. Google describes the difference between commodity and non-commodity content: general listicles anyone could write, against writing that carries genuine experience and a point of view. That was always true, but it matters more now, because an answer can supply the general information without you.
• Eligibility is stated clearly. A page must be indexed, eligible to show with a snippet, meet the technical requirements, and the site must be included in generative AI features in Search Console. Google separately notes that meeting every requirement still guarantees neither crawling, nor indexing, nor serving.

Our position, which Google's documentation does not state

What follows is no longer a citation but Devnora's own assessment, and it is worth separating. In our view, no GEO-specific technique has yet been publicly shown to have an independent, durable, cross-platform effect that can be separated from the effect of ordinary SEO work. That does not mean visibility in AI answers is unimportant — it matters. It means the effectiveness of the methods sold as GEO is currently a claim rather than a measured fact. This is why we treat GEO as part of SEO and do not sell it as a separate service.

What to actually do

• Check the basics: are pages indexed, are any blocked, do they respond quickly enough, is there duplication.
• Set up Search Console and enable the generative AI report, so you hold your own data rather than a supplier's metric.
• For each page, answer which question it settles and how its answer is not general knowledge. If there is no answer, the page's problem is not AI.
• Stop planning pages around phrasing variants. One good page per topic is both safer and more effective.
• If you already have an llms.txt or similar file, keep it — just do not treat it as a Google visibility measure.

Illustrative scenario, not a description of a client project

Imagine a company offered a GEO programme: an llms.txt file, rewriting content into chunks, additional schema, and twenty short pages per topic to "cover AI queries". By Google's own documentation the first three make no difference to Google Search, and the fourth — mass pages for phrasing variants — can fall within the scaled content abuse policy. The same budget could have produced a handful of pieces nobody else is able to write. This is an example of how we assess a proposal, not a promise about a particular outcome.

What cannot be promised

No supplier can promise you will be cited in an AI answer. The composition of an answer is decided per query by the system, and Google notes that meeting every requirement still does not guarantee being shown. The honest promise is a different one: remove the technical obstacles, write content that does not exist elsewhere, and measure the result with the official source. Everything else is guesswork that somebody decided to bill for.

If you want to check whether a GEO plan you have been offered contains anything ordinary SEO work would not do, send it over — we will point out which items make no difference according to Google's documentation, and which are genuinely worth doing.