ResearchJune 15, 20269 min read

Google made GEO official: legitimation with rules

On 5 June 2026 Google named GEO and AEO as legitimate service categories in its documentation. The useful reading is not demotion, it is legitimation with rules.

On 5 June 2026, Google published updates that changed the status of Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO) in its documentation: they stopped being terms outside its vocabulary and became legitimate service categories, side by side with classic SEO. Most of the industry read it as a demotion, "so it is just SEO after all". The more useful reading is the opposite: it is legitimation with explicit rules about what counts and what does not.

What changed: three documents, two dates

In three weeks, Google published three things that matter:

15 May 2026, a standalone technical guide, Optimizing your website for generative AI features on Google Search, inside a new Search Central section called Generative AI fundamentals. The first time Google formalises, in a dedicated document, how to appear in the answers generated by its search.
5 June 2026, two pieces on the same day: the new page Google Search's guidance on using third-party SEO tools, services, and advice (a warning about the market of tools and consultancies) and the update of the Do you need an SEO? guide to explicitly name AEO and GEO as a legitimate service category, alongside SEO.

It is the first time Google has recognised the market's vocabulary and described, in print, what it considers real work and what it considers an empty promise.

What Google says does not work

The list is short, direct, and particularly useful because it cuts through half the industry in one go. Google says you do not need to do any of these things to appear in AI answers in its search:

Create machine-readable files: llms.txt, ai.txt, special markup or structured Markdown.
Break content into small chunks for the AI to digest.
Write in a specific style for AI.
Apply schema.org dedicated to AI answers: there is no special markup, it is not a requirement.

And it explicitly warns against three practices:

Fabricating inauthentic mentions to manipulate generated answers: fake reviews, paid profiles, staged threads.
Believing services that sell themselves as approved by Google or acceptable to Google: no such programme exists.
Trusting third-party tools that guarantee performance. In the guide's direct translation: these tools do not have access to Google's internal ranking data and cannot guarantee performance.

The hard part: llms.txt

This is worth taking head on. In this section there is an article explaining how to create an llms.txt and, in June 2026, Google said it is not necessary. At first sight, that looks like a contradiction.

It is not, and the position was already in the article itself: llms.txt was always described as low-cost hygiene, not as a proven citation lever. Google's update formalises that nuance, it does not contradict it. It still makes sense to publish an llms.txt for three more modest reasons:

It documents the structure of the site for crawlers, auditors and internal teams, the equivalent of a human-readable sitemap.xml.
Engines other than Google may come to use it, some retrieval platforms and specialised agents already consume it, and the cost of having it is nil.
Two hours of work, with no recurring cost. The worst case is that it makes no difference; the best case is that it does.

What we do not do, and never did, is promise citations in ChatGPT, Claude or Gemini on the basis of a well-written llms.txt. Whoever promises that is selling what Google has just said nobody can guarantee.

What works, in Google's reading

The guide's through line is simple and familiar: unique, valuable, crawlable content plus real authority. SEO fundamentals still matter because the AI features rest on the same core ranking and quality systems. There is no separate AI engine; there is a search engine whose answers are, in some cases, generated by AI.

The mechanics, in two keywords worth knowing:

RAG (Retrieval-Augmented Generation): the AI does not invent from training; it fetches results from the index in real time and generates the answer anchored in those sources, with links.
Query fan-out: the AI reformulates the original question into several related sub-questions, searches for each, and synthesises. Hence the importance of covering a topic with lateral depth, not just with one excellent page.

The practical path to appearing in generated answers is, in essence, the same path that always made SEO work, only with a different bar for the content: it needs to be extractable and citable, not optimised for the click. There is no technical shortcut.

The measurement changed too

On 3 June 2026, two weeks after the technical guide, Google launched the first Search Generative AI performance reports inside Search Console. For now, on a subset of sites in the United Kingdom and with no click data at launch. But it is the beginning of AI visibility measurement inside Google's own infrastructure, instead of only through third parties.

The practical implication for any honest GEO operation:

Until now, the work inevitably depended on external tools, all with a natural conflict of interest because they sell the product they measure.
From now on, part of the measurement has a primary reference from the source. For clients in the United Kingdom, that is already operational. For the rest of Europe, it is a matter of months.
We keep using internal monitoring and external benchmarks, now with the Search Console AI report entering the stack as it opens to more markets.

Scope: this is Google. The others are not all the same.

A critical reading note few articles make: the documentation Google published describes Google's search (Search, AI Mode, Gemini). It does not describe, and Google does not control, ChatGPT, Claude, Perplexity, or other generative engines. Each has different indexing, training sources and mechanics.

The llms.txt Google dispenses with may be useful to a Perplexity agent today, or to an engine that does not yet exist. Generalising "the AI" as if it were all Google is a common and expensive mistake. Honest GEO work is done engine by engine: what serves each one, what each one rewards, what each one penalises.

Final reading

Google did three things in three weeks: it gave the work a name (GEO and AEO legitimated as service categories), it said what does not count as work (magic llms.txt, special schema, guarantees) and it started measuring it (Search Console AI reports). It is the best possible validation for whoever does this seriously, because it finally distinguishes practice from pitch.

For us, almost nothing changes. We keep doing the same: extractable content, real authority, honest measurement, a clear scope engine by engine. The difference is that there is now a shared vocabulary with the source, and it is harder for the less serious part of the industry to keep selling smoke.

Sources: Google Search Central documentation, May and June 2026 (Generative AI fundamentals, Do you need an SEO?, Google Search's guidance on using third-party SEO tools, services, and advice). Market numbers in the destaque.ai study on AI visibility of Portuguese B2B SaaS 2026.