Sooner or later someone phones you offering «AI optimisation». The list is
nearly always the same: an llms.txt file, markup of its own for AI, the
site’s text chopped into small pieces, and a package of mentions on portals and
directories. Four line items, four invoices.
Google keeps a public guide on exactly this subject, updated in July 2026, and it is unusually direct. Of the four line items, three are denied there in writing, one line each. The fourth is played down by the company itself.
What follows is a reading of that guide and of the help pages that come with it. The links are all at the end, so you can check every sentence without taking anyone’s word for it.
At the end there is what is left: boring, checkable work, and the only work that decides whether a page is even eligible to be cited. Nothing here promises you it will be cited, because nobody controls that.
AI Mode has spoken Portuguese in Portugal since October 2025
AI Mode reached Portugal on 7 October 2025, in Portuguese, announced on Google’s own Portuguese blog: «A implementação em Português em Portugal começou hoje» (the rollout in Portuguese in Portugal began today). AI Overviews, the generated block that appears above the results, were already here before that: Google announced their arrival in Portugal, in Portuguese and in English, on 25 March 2025.
In practice, when a client of yours searches for a service today, they will probably read an answer written by a model before they see the first blue link.
The same Google text adds a useful observation: early AI Mode users are asking questions «two or three times longer» than traditional searches. Anyone writing for loose keywords is writing for the search engine of 2015.
Google says there are no additional requirements
If someone tells you there is a second discipline to buy, Google’s documentation says otherwise: «There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.»
There are two requirements, and they are written down in the same family of pages.
- Ordinary search eligibility: «To be eligible to appear in Google Search generative AI features, a page must be indexed and eligible to appear in Google Search with a snippet, meeting Search technical requirements.»
- A box that is already ticked: «In addition to Search technical requirements, a site must be included in Search generative AI features in Search Console to be eligible for display in Google Search generative AI features.» It is a control you can switch off, and we come back to it further down.
Put another way: there is no second discipline to buy. The work that puts a page in ordinary search is the same work.
And Google warns, on the same page, that even that is not guaranteed: «Indexing and serving are not guaranteed.»
Google Search ignores the llms.txt file
The guide’s sentence is this one: keeping an llms.txt file
«neither harms nor helps your site’s visibility or ranking in Google Search,
because Google Search ignores those files».
Keeping one is legitimate, if some other service uses it. What is not legitimate is selling it as a requirement.
Notice why this line item turns up so often: it is a text file that takes ten minutes to write and that nobody can check is doing anything. If it is on the proposal in front of you, ask the question any budget line has to survive, the one we have already written out in full in what a website quote has to name: which specific system reads this file, and what does it do with it.
There is no schema.org markup built for AI
The guide is literal: «Structured data is not required for generative AI search and there is no special schema.org markup that you need to add.»
Structured data still has a use, and the use is the usual one: making a page eligible for a rich result in ordinary search, when that kind of result exists for that content.
What does not exist is markup created for language models. A proposal that names it is selling something the vendor’s own documentation says you do not need to add.
Chopping the text into pieces does not help AI
Another line from the guide: «You do not need to break content into very small chunks for AI to understand it better.»
This is the line item that does the most damage, because it is not only money. Following the opposite advice produces a site with dozens of thin pages, each with three paragraphs and the same call to action.
That is a different and more expensive problem, one we have already dealt with in why we do not build 30-page websites. A page that answers a whole question well is worth more than six that each answer a sixth of it.
Bought mentions are worth less than they look
The guide does not forbid them, but it does answer: chasing «inauthentic “mentions” across the web is not as useful as it might seem». In the same document it leaves a warning that covers this whole conversation: «Be wary of third-party tools that promise ranking success or claim to use “internal” Google metrics.»
The practical distinction is easy to make on your own. A mention that would have existed anyway, because a client wrote about you, because the trade association listed you, because a newspaper quoted you, is a mention. A mention bought from a portal that accepts any company by bank transfer is an invoice.
What takes a site out of the answers is the site itself
Two things can do it, and both are on your side. The first is the directives that prevent the snippet:
nosnippet: «do not show a text snippet or video preview in the search results for this page».max-snippet:0: limits the snippet to zero characters.data-nosnippet: marks parts of a page as not quotable.
If the page cannot show a snippet, it does not meet the first requirement from the previous section.
The second is the control Search Console now has for generative AI features. A site that opts out ends up like this, in Google’s words: «It will not receive any traffic or impressions from these features.»
The same page says the control «is not used as a ranking or inclusion signal affecting other parts of Search». Switching it off takes you out of AI answers and leaves the rest alone.
That gives you a free check, to do today, before you pay for anything: open the source
code of your site’s home page and search for nosnippet and
max-snippet.
It happens on sites migrated in a hurry or left in pre-launch mode. If the directives are there by mistake, no paid optimisation will have any effect while they stay there.
Ranking first is not enough to be cited
In an Ahrefs study published on 2 March 2026, covering 863,000 results pages and 4 million addresses cited in AI Overviews, the split was this:
- 37.9% of the cited addresses also appeared in the top ten result blocks for the same search.
- 31.2% came from positions 11 to 100.
- 31.0% came from outside the first hundred.
Counting only the blue links, and leaving the other blocks out, the top number is practically the same, 37.10%, but the other two move apart: positions 11 to 100 drop to 26.20% and pages outside the first hundred rise to 36.70%. On the stricter count, even more cited pages came from outside the first hundred positions.
It is a third-party study, not a Google measurement, and it measures AI Overviews rather than all of AI Mode. Even so, the direction is clear: ranking first is not enough, and not ranking first is no bar.
The mechanism the guide itself describes explains why. It is called
query fan-out, and Google defines it as «a set of related searches,
generated simultaneously by the model, to request more information and get additional
relevant search results that answer the user’s query».
The model is not answering the question the user typed. It is answering several sub-questions it wrote itself. A page can get in through one of those sub-questions, which it happens to answer well, without being the strongest on the main question. And a page that is first on the main question may answer none of them.
Our reading of this, and it is a reading, not an instruction from Google: it is worth writing one question at a time, in the words the client uses, and answering it in the next two or three sentences before you develop it. That is how this article is laid out, one question per section.
One caveat the guide itself makes, and it is worth keeping in view: creating separate
content for every possible search variation, including the sub-questions of
query fan-out, in order to manipulate rankings or generated answers,
breaks Google’s scaled content abuse spam policy.
The useful move is the opposite: one page that answers a whole question, rather than one page for every way of phrasing it.
The Search Console report shows impressions, not clicks
Search Console has a report of its own for generative AI features. Google’s help defines an impression as the number of times links to your site were shown to a user in a generative AI feature in Search.
The dimensions available are pages, countries, devices and dates.
What the report does not have matters more than what it has: no clicks, no click-through rate, no position, and none of the search queries.
It is also not on every property, because Google described the rollout as phased. If you open your Search Console and cannot find it, that is not a mistake of yours.
The commercial consequence is direct. If someone shows you a dashboard with your «AI position» and charges you monthly for it, you are looking at a number Google itself does not publish. It is worth rereading, here, the guide’s sentence about third-party tools that claim to use internal metrics.
OAI-SearchBot weighs on the answers, GPTBot on training
ChatGPT is a different system and Google’s documentation does not govern it.
What OpenAI documents, on its crawlers page, is OAI-SearchBot: sites that
block it «will not be shown in ChatGPT search results», though they may
still appear as browsing links. OpenAI’s own recommendation is to allow that
crawler in robots.txt.
Notice the shape of the answer, because it is the same as Google’s: there is no
sign-up, no submission and no purchase. There is a door that can be open or shut, and it
is your robots.txt that opens it.
That is another free check: type your site’s address in the browser followed by
/robots.txt, in the manner of aminhaempresa.pt/robots.txt, and
read what is there. If you find a Disallow for
OAI-SearchBot, someone made that decision for you, and it may have been
whoever built the site, without telling you.
A Disallow for GPTBot is a different thing, and it can be a
deliberate decision. OpenAI documents the two crawlers as independent, and that second
one only says the content should not be used to train the models. It does not shut the
door on search answers.
None of this guarantees a model will cite you. It only guarantees that it can.
The documentation that exists does not separate GEO from SEO
GEO is the commercial name the market gave to optimising for generated answers. On the documentation that exists, it is not a separate discipline: Google’s guide answers it with the sentence we have already quoted, there are no additional requirements and no special optimisations necessary.
Where there is a real difference, it is small and technical: other assistants have
crawlers of their own, with rules of their own in robots.txt, as with the
OpenAI case above. That is one line of configuration, not a monthly service.
If the proposal in front of you splits SEO and GEO into two line items, the question to ask is not hostile, it is simply specific: what is the concrete task that exists only in the second, and how is it checked once it is done.
What is left is boring and checkable
Not much, and it is always the same:
- Being indexed and eligible for a snippet, and not being excluded from generative AI features in Search Console. Those are the only two requirements Google names.
- Keeping the doors open in the files that control the crawlers.
- Writing for people, one question at a time, in the words they use.
- If you are a local business, keeping the business profile filled in and correct: the guide itself says that using Google Business Profile and Merchant Center can help products and services appear both in AI answers and in the other results.
Being indexed is the first half of this and has checks of its own, in order: what to do when your site does not appear on Google runs through them one by one.
If the rest of the budget you were shown is made of new files, invented markup, chopped text and bought mentions, you are paying for three things the technology vendor says in writing are not needed, and a fourth it plays down itself.
And one last piece of honesty, because it is what separates this article from the ones that will tell you otherwise: nobody, us included, can promise you that your site will be cited in an AI answer. Google writes, in its own documentation, that indexing and serving are not guaranteed.
What can be promised is that nothing on the site blocks eligibility: that the doors stay open, that the checks get done, and that anyone can confirm them afterwards.
If you have an «AI optimisation» proposal in front of you and you want an honest reading of what is in it, book a 15-minute call. It does not have to be us you work with.
Sources
- Requirements, llms.txt, schema, chunks of text, mentions,
query fan-outand Business Profile: Optimizing your website for generative AI features. - The absence of additional requirements and the sentence on indexing and serving: AI features and your website.
- Definitions of
nosnippet,max-snippetanddata-nosnippet: Robots meta tags. - The generative AI features control in Search Console: Search generative AI control.
- The generative AI features performance report: Generative AI performance report.
- AI Mode in Portugal, in Portuguese, on 7 October 2025: Google’s Portuguese blog.
- AI Overviews in Portugal, on 25 March 2025: Google’s announcement.
- Citation and position percentages: Ahrefs, AI Overview citations and the top 10, 2 March 2026.
OAI-SearchBotandGPTBot: OpenAI, Bots.
Consulted on 22 August 2026. Read the quoted passages at the sources above: those pages carry the exact wording, and they are the ones that count.