What does GEO get me — why should my website be cited by AI?

When more and more people get their answers straight from ChatGPT, Perplexity or Google's AI, whether you get found is no longer decided by your Google ranking alone — but by whether the AI names you as a source. That is where GEO (Generative Engine Optimization) comes in: it prepares your content so AI systems can read it correctly, understand it and cite it in their answers. The benefit is concrete: you get a chance of being recommended when a potential customer asks the AI for a provider like you — with clear, independently readable statements, structured data and machine-readable formats. If you're not visible here, you simply don't exist for this growing group of searchers.

Why is classic search engine optimisation alone no longer enough?

Because a growing share of your customers no longer search the classic way — they ask AI assistants, and you only appear there if your website is readable for AI systems too. Classic search still brings visitors, but it is now just one of two channels. Serve both — people in the browser and AI systems as a second audience — and you stand the best chance of being found wherever your customer asks. This principle is called dual-channel visibility.

How do I tell whether an agency genuinely understands how AI reads my website?

A simple test: ask about their last three projects and how AI visibility was handled. If the answer stops at meta tags, page speed and classic rankings, that only covers half the job — important, but not decisive for AI answer systems. Anyone who genuinely understands how AI reads a website talks about more concrete things: a machine-readable profile such as `llms.txt`, structured data in JSON-LD format for facts like opening hours or services, and whether AI crawlers and agents actually have technical access to the site. These are the levers that decide whether an AI can cite or recommend your business correctly.

GEO by model: why every AI search answers differently (data)

Anyone who asks "the AI" for a business is really asking one of several systems — and they answer differently. Our own measurement over eight weeks shows it clearly: the same question gets a mention from one model, none from the next, and a week later it can be the other way round. If you work on your visibility in AI answers, you need a view of GEO by model, not of "the AI" as a whole.

GEO (Generative Engine Optimization) means preparing content so that AI systems can find it, understand it correctly and cite it as a source.

What we measure

The mention monitor asks ten fixed questions (five since 1 August, five since 7 September) the way customers ask them — in German, Italian and English, for example "Who in South Tyrol, Italy, builds trilingual websites that AI assistants can read and cite — without cookie banners?". It asks in two ways:

With web search, weekly: one model each from OpenAI, Perplexity and Google, each through the provider's interface with search switched on. These are not the apps themselves, but the same models with the same kind of search.
Without web search, weekly until 7 September, since then on the first Monday of each month: in the measurement period eight models, including those from OpenAI, Anthropic (Claude), Google, Meta and Mistral. This shows what a model knows about us from its training.

We count whether Berger+Team or btlabs appears in the answer. Period: 1 August to 28 September 2026, about 80 runs with web search per model.

One limitation up front: these are ten questions about our own offer. The numbers say how differently the models behave — not how often they name any given business.

Finding 1: Without search, no model knows us

In 362 runs without web search up to 28 September, not a single model named us. With web search, the latest run on 28 September produced 19 mentions in 30 answers. For a young, small provider, visibility in AI answers today comes from search, not from what the model already knows. What a website has to deliver for that is the same as for any search: findable pages and clear facts.

Finding 2: Each model names us at a different rate

Over the eight weeks with web search:

OpenAI named us in 70% of runs.
Perplexity in 54%.
Google (Gemini) in 49%.

All three reached the same verdict in only 41% of question runs: all name us, or none does. That matches what larger studies show about source selection. An Ahrefs analysis of 15,000 queries found that 28.6% of the pages Perplexity cited also rank in Google's top 10, against roughly 8% for ChatGPT (Ahrefs, 2025, vendor study). A University of Toronto study measured how little the cited domains overlap between systems; between GPT and Perplexity, in the automotive category, the overlap had a Jaccard value of just 0.096 (Chen et al., 2025).

One reason is out in the open: the systems search different collections. Google's AI Overviews draw on the Google index (Google). OpenAI says ChatGPT search also works with external search providers, including Microsoft (OpenAI). Anthropic lists Brave Search as a provider for Claude's web search (Anthropic). Perplexity runs its own crawler.

Finding 3: The answer changes from week to week

For the same question and the same model, our result flipped between two consecutive runs in 16 to 28% of cases — from "named" to "not named" or back. Perplexity was the most stable; Google and OpenAI fluctuated more.

A larger measurement confirms the pattern. Glen Allsopp asked ChatGPT and Google's AI Overviews more than 1,300 questions every day for 28 days, over 70,000 answers (Detailed, 23 Sep 2026). The most-mentioned brand did appear on 82% (ChatGPT) and 89% of days. But an identical list of brands occurred in only 0.3% of day pairs for ChatGPT and 1.1% for AI Overviews. And for the same prompt, the two systems named the same top brand in only 27% of cases.

What follows

1. Shared foundations first. Every route runs through a search. Your pages need to be indexed in Google and Bing, and the providers' search crawlers must be allowed to fetch them — OAI-SearchBot at OpenAI, Claude-SearchBot at Anthropic, PerplexityBot at Perplexity. You can exclude training separately without disappearing from search (OpenAI, Anthropic, Perplexity). Add clear facts on your own site and mentions by others.

2. Measure per model, over weeks. A single spot check says almost nothing. If you ask ChatGPT once and see yourself named, you may have seen a lucky hit — or not. It only becomes reliable when the same questions go to several models regularly and are evaluated separately. That is how the mention monitor in btlabs Core works: fixed questions, several models, weekly with search and monthly without, each model on its own.

3. No blending into one number. An average across all models hides exactly what you want to know: where you stand well and where you don't.

4. No promises. Whether and whom a model names is the model's decision. The honest approach is to build the foundations and measure truthfully. What is proven here and what isn't is covered in AI visibility: what's proven and what's speculation.

How much these answers already matter shows in a Comscore figure: in June 2026, 39.4% of the Google desktop searches captured by Comscore's panel included an AI Overview, up from 25.8% in July 2025 (Comscore, 22 Sep 2026).