GEO

Google increasingly keeps the traffic for itself

For twenty years, search optimisation ran on a simple logic: Google doesn't answer the question itself, it points at pages that do. Rank high, get the visitor.

That logic no longer holds without qualification. Google now shows its own AI answer on more and more queries, assembled from the content of several pages. The user gets what they wanted to know directly on the results page. Clicking the source becomes optional.

For your website that changes the arithmetic fundamentally. You can rank first and still lose visitors, because above you sits an answer built partly from your own content, without anyone arriving at your site.

What the numbers say

The effect is now measurable. Ahrefs compared 300,000 search terms, half with and half without an AI Overview in the results, and set the click-through rates from Google Search Console between December 2023 and December 2025 against each other.

The result: where Google shows an AI answer, the first organic result loses roughly 58 percent of its clicks. The answer sits at the top, the user reads it, and the click never happens.

For businesses that doesn't mean SEO becomes worthless. It means a good ranking alone is no longer enough when the answer sits above it and you don't appear in it.

The principle: how you get into the answer

An AI answer doesn't form like a ranking. Google isn't sorting ten pages and showing the best one, it reads several sources, pulls individual statements out of them, and assembles a text. Whoever's statement got used is the one that gets named.

That implies a different optimisation logic. It's no longer about being the most relevant page for a keyword, it's about providing the clearest answer to a concrete question. A model takes a sentence it can attribute unambiguously and reproduce without interpretation.

In practice that means three things. First: answers instead of marketing language. "We install wall boxes in and around Freiburg, usually within two weeks" is quotable; "your competent partner for all things e-mobility" is not. Second: make facts explicit instead of burying them in prose. Third: structure a machine can read without guessing.

The side effect is pleasant: content an AI can cite cleanly also reads better for humans. Unlike some SEO tricks of the past, there's no conflict of interest here.

What that means in practice

Structured data. What a human infers from context has to be explicitly readable by a machine: services, catchment area, opening hours, prices, responsibilities.

Citable content. Clear statements with numbers and facts instead of marketing prose. A model cites what it can reproduce unambiguously.

A machine-readable summary. An llms.txt file summarises the key facts so AI systems can read them cleanly instead of guessing from the navigation.

Verification. Testing the questions your customers would ask in the common systems, and checking whether and how you appear.