There's no public ranking algorithm for AI citations. No list of factors with assigned weights. But the patterns are consistent enough across models that you can reason about what drives mention decisions — and more usefully, what causes content to get skipped.
The basic retrieval loop
For models with web access, the process runs roughly like this: receive the query, determine what kind of answer is needed, retrieve a set of pages from the ranked web, extract the relevant portions, synthesise a response.
Each stage filters. The query determines whether retrieval happens at all — factual lookups the model already knows don't trigger it. The retrieval returns pages that rank for related terms. Extraction pulls the most prominent, legible content from those pages. Synthesis picks what to include in the final answer.
By the time you get to the mention decision, most content has already been filtered out.
What gets extracted
Models extract from the top of the page first. The first 200 words carry disproportionate weight. If the clearest answer to the question is three scrolls down, it may never be reached. If the opening is vague brand copy, that's what the model has to work with.
Structured content — clear headings, tables, numbered lists — is easier to extract than dense prose. Not because models can't read prose, but because structure signals where the information is.
What gets cited
Extracted content gets cited when it's corroborated. A single source making a claim is lower confidence than three independent sources making the same claim. Models are calibrated toward convergence — they cite what multiple credible sources agree on.
This is why strong content on a weak domain underperforms. And why a brand mentioned consistently across independent sources gets surfaced even when its own site is mediocre.
The most common failure modes
The answer is buried. Move it to the top. This is the single highest-return structural change for most pages.
The claim is unsupported. Specific, factual claims with evidence behind them get cited. Superlatives without backing don't.
The page is hard to parse. Test your key pages in reader mode. If the resulting text is confusing or incomplete, a model will have similar trouble.
The content exists in isolation. If nothing else on the web corroborates what you're saying, the model has no basis for confidence.
None of this requires exotic optimisation. It requires being clear, specific, and present in the right places.
Want to know whether your content is making it through these filters? The V-Score gives you a direct read — run your free AI visibility audit here.