Google's recipe for AI visibility

covers under half

On 15 May 2026, Google published an official guide to becoming visible in generative AI search. It is well written, it is free, and it is worth reading. It is also written by a party with interests in the matter. We measured for ourselves, on five Norwegian websites. Google's own surfaces accounted for between 35 and 48 percent of AI citations.

Four pieces of advice worth remembering

Google's guide to generative AI search is clear on four points that run straight against much of what the industry has sold over the past year:

On structured data: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.”

On length and splitting: “There’s no ideal page length, and in the end, make pages for your audience, not just for generative AI search.”

And further: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”

On writing for machines: “You don’t need to write in a specific way just for generative AI search.”

On mentions: “Seeking inauthentic 'mentions' across the web isn’t as helpful as it might seem.”

Their conclusion is that “optimizing for generative AI search is optimizing for the search experience, and thus still SEO”.

By Marius Vestlien Hansen
SEO/SEM Lead, Maverix
marius@maverix.no

We measured how much of AI visibility is actually Google

We pulled the number of citations per AI platform for five Norwegian websites: four in travel and hotels, plus our own website as a reference. A citation is a link to the website that was actually used as a source in a generated AI answer.

The pattern is stable across five websites with very different volumes, from 11 to 1,171 citations: Google's surfaces account for around a third to just under half of AI citations. More than half of AI visibility happens elsewhere.

Google's guide is still worth following. It covers less than half of the terrain, though, and Google has good reasons to have no interest in describing the rest.

How to read the numbers

Here are the most important caveats, because this is where most industry articles cheat.

Citations measure something other than traffic and users. Perplexity is the largest single source on three of the five websites in the table above, with 24 to 26 percent of citations. In May 2026, Perplexity had about 1.3 percent of web visits among AI assistants, against 53.9 percent for ChatGPT and 27.9 percent for Gemini, according to Similarweb figures reported by Search Engine Journal. Measured as referrals to websites, Perplexity is larger: StatCounter reports 7.9 percent for June 2026. Perplexity cites aggressively and links a lot, but reaches far fewer people than ChatGPT and Gemini. A high citation count there is worth less per citation than a low one in ChatGPT.

Grok sits at zero everywhere. That is almost certainly a coverage gap in the data and a poor description of reality. We let the number stand as it is, but we build nothing on it.

The sample is an indication. Five websites, mainly from one industry, come from the industry we know best. We present the method so that anyone can reproduce it on their own domain, and we are happy to share the underlying domain names in a conversation.

Read more

Metadata: a hidden force behind digital visibility

Our point is the distribution: it is real, it varies from website to website, and it cannot be read from Google's documentation.

Where Google is right, and where it is silent

We want to be precise about Google's guide, because there are three places where precision changes what you should do.

Structured data. Google writes that structured data is not required. That is correct, and the value is real all the same. Structured data is the only place where you can state, machine-readably and without room for interpretation, that this is an organization with this name, these products, this price and these opening hours. Google has the infrastructure to guess. A model that fetches your page in real time through a RAG lookup has that to a lesser degree. The cost of adding it is close to zero, and the gain is that you remove doubt. It is a poor trade-off to skip.

Splitting and length. Here Google is mostly right, and the industry has muddled it. The advice to “chunk” content has been read as a call for short pages. That is wrong. Splitting is about structure within the page: each section answers one question, under a heading that says what the answer is, so that the paragraph can be lifted out alone and still make sense. The total length of the page is a separate matter. Depth and splitting work at different levels. A long, well-sectioned page gives you the best of both.

And here is where Google and we actually agree: a text split into clear, self-contained sections is also easier for people to read. Writing for people is still right, but the advice should be extended with machine-readable structure. Human readability and machine-readable structure point the same way in almost every case, and where they diverge, machine readability costs you very little.

Mentions. Google warns against “inauthentic mentions”. The warning applies to bought links and link swaps, and it is entirely appropriate. Editorial coverage in credible trade media falls outside it, and the guide says nothing about how such coverage works in systems Google does not own.

Here you should separate two very different mechanisms, because they have very different time horizons:

The first is training data. Coverage that is widely spread and repeated over time can settle into the model's parameters, so that it “knows” who you are without looking it up. It works slowly, you cannot steer it precisely, and it updates only when the model is retrained.

The second is retrieval. When a model fetches sources in real time to answer, the article about you in Dagens Næringsliv, Kampanje or a trade forum competes to be fetched, in the same way your own pages do. It works fast, and this is where the effort actually pays off in the short term.

Merging the two into “mentions program the AI” gives the wrong picture. Keeping them apart gives the right priorities.