All insights
Fundamentals21 Nov 20266 min read

What AI search actually rewards (and what it ignores)

Answer engines are not ranking pages, they are assembling answers. That changes which parts of your site matter, and which habits from SEO are now wasted effort.

Most AI visibility advice is SEO advice with a new noun. That is a mistake. A search engine ranks a page; an answer engine composes a response from several sources, then names a few of them. The unit of competition is the sentence inside the answer, not the position of a link.

What gets pulled into an answer

Models and retrieval layers look for passages that answer a question directly, in language close to how the question was asked, with enough specificity to be quotable. In practice, four properties keep showing up:

A direct answer near the top. A page that opens with 600 words of context before answering is a page that gets skipped. State the answer, then explain it.

Specific claims with attribution. "Many companies report improvements" is unusable. A sentence such as "In our audit of [number] B2B sites, [number] were absent from the first answer for their own product category", filled in with real figures, is citable, and it survives summarisation because it still means something when shortened.

Structure a machine can follow. Clear headings that match real questions, short paragraphs, and lists where a list is genuinely the right form. Not for the crawler's sake, but because extraction is easier from clean structure.

Consistency about who you are. If your site says you are a consultancy, a directory says you are an agency and a profile says you are a platform, the model has no stable entity to describe. It will describe you vaguely, or describe a competitor instead.

What matters much less than it used to

Keyword density, exact-match anchors, and word count targets are largely noise in this context. So is publishing volume without specificity. A single page that answers a real commercial question with sourced evidence outperforms twenty pages written to cover a topic.

Also overrated: chasing every model. The prompt behaviour and sourcing differ between ChatGPT, Gemini, Perplexity, Copilot and AI Overviews, but the underlying requirement is identical. Fix the substance once, then measure per engine.

The uncomfortable part

Being cited is not the same as being liked. If your published position is vague, an engine will summarise you into a category you did not choose. The brands that do well here have usually done the harder work of deciding what they are actually the best answer to, and saying it plainly enough to be repeated.