How AI engines decide which companies to cite
A generative engine picks the companies it can retrieve, read, identify and corroborate. Retrieval gets your page into the candidate set; extractability lets the model lift a usable passage; entity clarity tells it who you are; corroboration from independent sources gives it the confidence to name you. Weakness in any one of the four is enough to leave you out of the answer.
Key takeaways
- Four gates: retrieval, extractability, entity clarity, corroboration.
- PDF-only and image-only specifications fail the extractability gate.
- Independent sources that agree with you are worth more than more of your own pages.
- Being unnamed is silent — nothing in your analytics reports it.
Gate 1 — retrieval
Before anything else, the engine must have your page in reach: indexed by the search backend it queries, fetchable by its crawler, and rendered without JavaScript gymnastics. Blocking AI user agents in robots.txt, gating specifications behind forms, or shipping a client-only shell all fail here — quietly.
Gate 2 — extractability
Models quote passages. A passage is extractable when it states one fact plainly and completely: the grade, the value, the unit, the condition, the standard. "Excellent thermal stability" is not extractable. "Retains 92% tensile strength after 1,000 hours at 150 °C per ISO 188" is.
This is where most industrial sites lose the answer. The number the buyer asked for exists — inside a PDF, or inside an image of a table.
Gate 3 — entity clarity
The engine needs to know that the page belongs to a company, that the company makes the product, and that this company is the same one mentioned in three other sources. Consistent naming, an Organization graph with a stable identifier, Product and TechArticle markup, and matching details across your own site, LinkedIn and directories all feed this.
Gate 4 — corroboration
Models hedge on claims only the claimant makes. When a standards body, an industry publication, a customer case study or a named expert says the same thing, the model's confidence rises and your name survives into the final answer. This is the slowest of the four gates and the one that most durably separates leaders from everyone else.
What this means for the fix queue
- Fix retrieval first — it is cheap and it unblocks everything else.
- Then extractability: get numbers out of PDFs and into HTML with units and conditions.
- Then entity markup, which is a one-time structural job.
- Then corroboration, which is an ongoing publishing and PR programme.
Frequently asked questions
Do backlinks still matter?
Indirectly and significantly. Links correlate with being indexed, being retrieved, and being corroborated. But a mention without a link from a source the engine trusts can be just as valuable, because retrieval is text-based.
Can I pay to appear in AI answers?
Not in the organic answer itself. Engines are adding advertising surfaces around answers, but the cited companies inside the answer are earned.
Why is a competitor cited when my page is better?
Usually extractability or corroboration. Their weaker page states the number in plain HTML, or three independent sources repeat their claim. Better content that cannot be lifted loses to worse content that can.
See where you stand in AI answers
Run a free GEO scan: a 0–100 score across five dimensions, page-level evidence and a prioritized fix queue. No card required.
Keep reading
Last updated 8 September 2026 · Written by the EGNITE GEO team.