Search used to return a list and let you decide. Increasingly it returns a decision. That single change rewrites what “being visible” means — and most measurement stacks haven’t caught up.
Ranked lists versus synthesised answers
A classic search engine retrieves documents, orders them by relevance and authority, and hands you ten links. Your job as a marketer was to occupy one of those slots. The user still chose.
A generative engine does something different. It retrieves candidate sources, reads them, reconciles conflicts between them, and writes one answer. The user usually accepts that answer. There is no slot to occupy — there is only whether your brand appears inside the sentence the model produces.
This is why a page can rank first and never be mentioned. Ranking measures retrievability. Citation measures whether a model decided your content was worth repeating.
What actually drives citation
Across the audits we run, four properties separate cited sources from ignored ones:
- Extractable claims. Specific, checkable statements — numbers, dates, defined terms — rather than adjectives. A model can quote “reduces onboarding time from 14 days to 3”; it cannot usefully quote “dramatically faster onboarding”.
- Structural clarity. Content where the answer is near the top and the supporting evidence follows. Models reward documents that are easy to excerpt without losing meaning.
- Corroboration. Facts that agree across your site, your directory listings, review platforms and press coverage. Contradiction is the fastest way to get filtered out, because the model cannot tell which version is true.
- Entity resolution. An unambiguous answer to “what is this company and what does it sell”. Brands with generic names or unclear categories are routinely merged with other entities.
Why keyword density does nothing
Keyword targeting worked because lexical matching was part of retrieval. Language models operate on semantic representations — repeating a phrase adds no signal, and stuffed pages read as low quality, which suppresses citation.
The replacement discipline is closer to technical writing than to copywriting: say the thing precisely, support it with evidence, structure it so a machine can lift it cleanly.
How to measure it
Rank tracking cannot tell you any of this. You need prompt-level measurement: a fixed set of buying-intent questions, run repeatedly across each engine, scored for whether you were mentioned, cited, or displaced by a competitor.
Run it weekly. These systems change without announcements, and the only way to notice a drop is to have been watching the same prompts before it happened.
If you cannot show the answer you got last month, you cannot prove anything changed this month. Baseline first.