AI optimisation: what’s being sold and what’s actually true

Artificial intelligence optimisation — that is, GEO, AEO and other acronyms — is marketed as a separate discipline with its own techniques. This is not true. Google states explicitly in its own guidelines that optimisation for generative search is search engine optimisation — and the only study in the field conducted with a comparable control group found that methods specifically targeted at AI search are mostly ineffective and often even harmful. This article examines what has been measured, what has been guessed, and what is being sold without any evidence.

Google says so itself

Let’s start with a source that is more authoritative than any research: Google’s own documentation. The page ‘Optimising your website for generative AI features’ (updated 10 July 2026) states, word for word:

From Google Search’s perspective, optimising for generative AI search is optimising for the search experience, and is therefore still SEO.

On the same page, Google also states that no new files are required: “You don’t need to create new machine-readable files, AI text files, markup, or Markdown to appear in Google Search.” And regarding structured data: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.”

This isn’t marketing spiel but developer documentation maintained by Google to ensure that website creators know how to proceed correctly. When a service is marketed on the basis that AI search requires separate technology, the seller’s view differs from that of the company whose search engine is being discussed.

To what extent is AI optimisation the same as standard SEO?

There’s one popular figure doing the rounds in response to this question: 75–80 per cent. It’s broadly correct but untenable as a figure, and the reason is an interesting one: there are two different answers to the question, depending on what is included in the calculation.

What is measured	Result
How many AI-generated responses contain at least one source that also appears in the search results?	81–94 %
What proportion of all individual citations come from the top of the search results?	12–54 %

Both are correct answers. They simply answer different questions, and in marketing materials they are used interchangeably.

Even more instructive is what happened to one particular figure. In July 2025, Ahrefs measured that 76.1 per cent of citations in Google’s AI-generated answers came from the top ten results. This figure was widely reported. In March 2026, the same company measured the same thing again and found the figure to be 37.9 per cent. It cited two reasons for this: its own parsing method had improved, meaning the old figure was partly due to a measurement error — and Google had started breaking questions down into sub-questions, meaning that sources were drawn from searches that the user had never actually entered.

For AI systems other than Google’s own, the share is significantly smaller. When Ahrefs ran 15,000 queries through ChatGPT, Gemini, Copilot and Perplexity and compared the sources with Google’s top ten results, the overlap was 12 per cent.

How should this be phrased correctly?

Almost every Google AI response relies on at least one page that also appears in standard search results. Around a third of individual sources come from the top ten results, whilst for other AI systems the figure is around a tenth.

And here’s what really matters in practice: search engine optimisation is a necessary condition for AI visibility, but it is not sufficient. Google’s AI Overviews and AI Mode retrieve their sources from Google’s own index using a standard Googlebot. ChatGPT’s search requires access via its own bot. A page that hasn’t been indexed won’t appear in any of these. However, there is no evidence whatsoever that SEO ‘accounts for 80 per cent’ of AI visibility — the overlap figure is not a measure of explanatory power.

Four things that are sold without proof

1. llms.txt

A file intended to inform language models about the website’s content. In May 2026, Ahrefs analysed 137,210 domain names: 97 per cent of llms.txt files received no requests at all during the entire month. The largest single group of readers were SEO audit tools (21.7 per cent). Bots that generate search results accounted for 1.1 per cent.

Google’s John Mueller wrote in June 2025: “FWIW, no AI system currently uses llms.txt”, adding that this can be seen in his own server logs. The file isn’t useless — it’s a neat piece of infrastructure for developer documentation and code agents. But no brand is going to win the AI race because of a text file.

2. ‘AI schema’ and structured data as a solution for AI visibility

This is the only proper control group experiment in the field. Ahrefs tracked 1,885 pages that added JSON-LD markup between August and March, and compared them with 4,000 control pages from different domains. Result: AI Mode +2.4 per cent, ChatGPT +2.2 per cent, AI Overviews −4.6 per cent. It’s all just noise.

Otterly.ai carried out an even more illustrative experiment: it provided a unique answer based solely on structured data, not on visible text, and asked seven AI platforms to retrieve it. Not a single one found it. A third study (searchVIU) reached the same conclusion: in real-time searches, the models read the visible HTML, not the JSON-LD.

Structured data is still worthwhile — it produces enriched search results and is part of a clean implementation. It just shouldn’t be marketed as a solution for AI visibility.

3. Hidden text and instructions written in the language template

The honest answer here is a bit uncomfortable: technically, it works. Researchers at ETH Zurich demonstrated this using the production versions of Bing and Perplexity, and a corresponding result has been published in a peer-reviewed paper at the EMNLP conference in 2024.

Yet it cannot be sold, for three reasons. Google’s spam policy explicitly prohibits hidden text and cloaking. Researchers at ETH model the situation as a prisoner’s dilemma, where the benefit disappears as soon as the method becomes widespread. And thirdly: this constitutes an attack against a third party, not an improvement to the client’s website.

4. Mass production of content

Services are being sold in this sector for $299 a month, promising “30+ articles automatically”. Google’s definition of ‘scaled content abuse’ specifically covers this. The industry press has reported cases where a provider presented a client with rising AI visibility figures whilst that same client’s organic traffic plummeted — in one instance by 66 per cent.

So what makes the difference?

The screen is thinner than the description suggests, but it’s not non-existent. Ranked from strongest to weakest:

Designer	Display	What is it based on?
Bots can access the site	Strong	OpenAI states: websites blocked by its search bot do not appear in ChatGPT’s search results. The same applies to Google.
Relevance	Strong	252,000 test runs, six models, 18 variables modified one at a time (SIGIR 2026)
Third-party brand mentions	Reasonable	75,000 brands: mentions correlate around three times more strongly than links
Organic positioning	Reasonable	The top position is cited in 43 per cent of responses, whilst 20th place is cited in only 7 per cent
Comparative, list-style content	Reasonable	63% of citations pointed to list-style articles (~400 million citations)
Server-side rendering	Reasonable	None of the major AI crawlers execute JavaScript
Structured data	No impact	Control group test, 1,885 + 4,000 pages
llms.txt	No impact	97 per cent of files unread

Note what’s at the top of the list: standard technical proficiency and mastery of the subject matter. Nothing that’s been invented because of language templates.

Where does AI search really make a difference?

There are differences, and they are genuine — they’re just not the ones that are being sold.

• There is no ranking; it is a distribution. When the same query is run twice, around 80 per cent of the sources change. The probability of two identical brand lists is less than 1 in 100.
• The query is broken down into parts. Google searches for sources based on sub-queries that the user did not type. The ranking of an individual search term is less important; the breadth of the topic is more important.
• Visibility is about the brand, not the page. Search engine optimisation focuses on the URL, whilst AI-powered search focuses on the business.
• There are many machines, and they are almost entirely separate. The similarity between the source sets of different AI machines is less than 0.2. ‘AI visibility’ within a unit is a flawed concept.
• There are three types of bots. Training, search indexing and user-triggered searches are different things. Practical implication: Blocking GPTBot will not remove you from ChatGPT’s search results, nor will blocking Google-Extended remove you from AI Overviews.

Why no metric is reliable

This is the most awkward part of the whole subject, and it’s worth knowing before you buy an ‘AI visibility report’.

A variance analysis published in July 2026 examined the causes of variation in AI responses. Data: 12,933 responses, 20 brands, 8 languages, 3 models. Results:

A source of variety	Share
The same question repeated	34,8 %
Language of the survey	32,0 %
Brand × context	29,6 %
Template used	1,7 %
The brand itself	1,6 %

The researchers’ conclusion: the reliability of a single response for brand ranking is 0.010 — effectively zero. And even the maximum sample size of 3,600 respondents only reaches a score of 0.365, which is insufficient to draw conclusions about a single company.

Another study measured this differently: around 65 per cent of sources change from one day to the next. A third study found that the same model, using the same question but with a different mindset, produced an overlap of only 25.6 per cent in the domains cited.

The practical implication is clear: visibility share across tens or hundreds of searches is a sensible metric. A single screenshot is not, nor does ‘investment in AI search’ exist as a concept. When a tool promises such a thing, it is promising a metric that the research literature considers impossible.

Two figures that change the proportions

Around two-thirds of ChatGPT’s responses do not retrieve anything from the internet. In Semrush’s click-stream panel of 200 million users, a web search was triggered in 34.5 per cent of queries, and this figure has fallen from 46 per cent. All work aimed at identifying a source therefore relates to a minority of responses.

Server logs do not measure visibility. According to Cloudflare, around 80 per cent of indexing by AI bots is carried out for training purposes, and searches triggered by user queries account for less than 5 per cent. Heavy bot traffic in the logs therefore indicates that your content is being crawled — not that it is being shown to anyone.

What hasn’t been measured

In all honesty, here’s a list of things nobody knows:

• To what extent can search engine optimisation account for AI visibility? The overlap figures do not represent degrees of explanation.
• Is the effect of third-party mentions causal? No intervention trial has ever been conducted.
• Does AI visibility generate clicks, leads or sales? Not a single study measures the end result.
• How many of the visits influenced by artificial intelligence go completely untracked when the user later searches by name or types in the address directly?
• How do Google, OpenAI or Perplexity actually select their sources? All ‘citation factors’ are observations made from outside the black box.

This is where an honest journalist differs from a salesperson. AI visibility is likely to have an impact that current metrics fail to capture — and that is precisely why no one should be selling it as a percentage figure.

Frequently asked questions

Is GEO different from SEO?

Not as a separate job category. Google defines optimisation for generative search as search engine optimisation in its own documentation. There are genuine differences — sources vary from one search to the next, the query is broken down into sub-queries, and visibility relates to the brand rather than an individual page — but these are not a separate technique that can be purchased as an add-on.

Is it worth adding an llms.txt file to the site?

It can be added, but it isn’t worth paying for and you shouldn’t count on it. In a study of 137,210 domain names, 97 per cent of files received no requests at all in a month, and the largest group of visitors were SEO tools. According to John Mueller of Google, no artificial intelligence system currently uses it.

Does structured data help it to appear in AI responses?

Not to any significant extent. In an experiment conducted with a control group (1,885 test pages, 4,000 controls), the effect remained within the margin of error, and on one platform it was negative. Structured data is still worth doing properly, as it produces enriched search results — but it is not a tool for AI visibility.

Can AI visibility be measured reliably?

Partially. The visibility share, calculated from tens or hundreds of visits across multiple languages and platforms, is a sensible indicator. A single screenshot tells us nothing: in the analysis of variance, the reliability of a single response was 0.010. There is no such thing as ‘ranking in an AI search’, so it cannot be measured either.

Should AI bots be allowed on the site?

If you want to appear in AI-generated responses, yes — and search bots in particular. OpenAI states that websites which have blocked its search bot do not appear in ChatGPT’s search results. Please note that the training bot and the search bot are different things: blocking GPTBot will not remove you from ChatGPT searches, and blocking Google-Extended will not remove you from AI Overviews. Many websites block these accidentally due to a plugin’s default settings.