Why your Google ranking says nothing about your AI visibility

We started checking, for a sample of our own customers, whether their Google position predicted whether an LLM would mention them. It does not. Across roughly 400 commercial queries, the correlation was weak enough to be useless for planning.

Two different retrieval problems

Classic search returns a ranked list of documents. You compete for a slot, and the slot is visible: you are #3 or you are not.

An assistant answering the same question does something else. It retrieves a handful of sources, reads them, and writes one answer. There is no list. Either your brand appears in the sentence or it does not, and there is no second page to be on.

That changes what “winning” means:

• Ten blue links reward being one of ten. One synthesised answer rewards being one of two or three.
• Rankings are stable for weeks. Model answers vary between runs of the same prompt.
• A ranking is per keyword. An AI mention is per phrasing — “best rank tracker” and “affordable rank tracking tool” can return entirely different brand sets.

Where the divergence comes from

Three causes we can actually see in the data:

Training data lag. A model’s parametric knowledge is frozen at its cutoff. If you launched in 2025, a model trained earlier only knows you through retrieval — and only if retrieval fires for that prompt.

Citation preference. Assistants over-favour sources that read as reference material: documentation, comparison pages, forum threads with a clear consensus. A landing page optimised for conversion is rarely the thing that gets cited, however well it ranks.

Aggregator mediation. For “best X” prompts, the model often reads a listicle rather than vendor sites. If you are absent from the three listicles it likes, you are absent from the answer even at position #1 on Google.

What to measure instead

Position is the wrong unit. We track four things:

1. Mention rate — of N prompts in your category, what share name you at all.
2. Share of voice — your mentions against each named competitor’s.
3. Sentiment — mentioned as the cheap option, the serious one, or the outdated one.
4. Citation source — which URL the model leaned on, which tells you what to fix.

Only the fourth is actionable on its own. The first three tell you whether the fourth is working.

The uncomfortable part

Being invisible to assistants is not a penalty you can appeal. There is no equivalent of a manual action, no reconsideration request, and no ranking factor list. You get one lever: be the source that is worth citing, in the format that gets cited.

That is slower than buying links. It is also considerably harder to copy.