AI answers are assembled from web pages. We ran several related analyses to see which pages got cited and what tracked with it. This page reports observations, not a proven mechanism: everything below is a dated association from our own runs in one service vertical, on a proxy AI-search setup, not a causal claim about how any model decides.
The outcome we measured is narrow and specific: whether a business's own domain appeared as a cited source in an answer. That is not the same as being "recommended," being named in a particular way, or an AI repeating a specific fact from that page. We did not trace individual facts back to their source pages.
What we observed
All figures are first-party, dated, and reported with their denominators and definitions. The runs used an AI search model with live web access as a proxy for AI search, not the default ChatGPT app.
1. In these queries, the cited sources were mostly business websites. In a 150-query Google AI Overviews sweep (2026-09-03), 183 of 198 saved source items (92%) were classified as clinic or chain websites by title and description, not a full URL-by-URL check. In the recommendation runs, social-platform domains were a small share (6 of 2,199 domain items, 0.27%) on a different engine and denominator. Read this as "in these specific queries, saved sources skewed to business sites," not as a universal rule that AI ignores social, and a cited domain does not prove the business first authored the fact.
2. Pages that stated their facts were cited somewhat more often, in one slice. Defining a stable outcome as a domain cited in at least two of three passes: among domains whose scored pages contained 0 to 2 facts, 22 of 108 (20.4%) were cited stably; for the 3 to 4-fact group the figure was 25 of 68 (36.8%), about 1.8 times as high. On a looser "cited in any pass" definition the ratio is about 1.4, and the top fact bucket is a single page. So this is a specific slice, not a "2× at every threshold" effect and not a probability for any given query.
3. Page-fact richness was associated with citation in this sample. We scored reviews for 174 clinics that met our threshold (48 had a stable citation) and ran a regression. The richness of a clinic's own page facts was associated with getting cited (odds ratio about 2.0, 95% CI 1.4 to 3.0, p under 0.001), while review count (OR 1.32, not significant) and review content (wide, non-significant interval) were not. Our content classifier also had modest recall (~63%), so the content measure is imprecise. We did not detect a statistically significant association for the review metrics; this does not establish that their effects are absent. It is one vertical and an association, and we did not measure site age, so we make no "newer beats older" claim.
4. Cost questions tended to trigger an AI answer, in the queries we saw. Of 150 AI Overview queries, 81 returned an answer in our capture; among those, the 50 cost/price queries all showed an AI Overview. We report that as "cost queries triggered an AI Overview in the 50 we captured," bounded by that set and date, not as a universal 100%. Of the pages cited for cost, most stated a price, though some quoted a market average from a third-party site rather than the clinic's own price.
What this suggests for your site
- Stating your decision-facts plainly, in crawlable text, is something you control. In our sample it tracked with getting cited; it is not a guaranteed outcome, and a model can still pull a fact about you from another source.
- Most competitors leave these facts out of the opening (see the 100-Word method), so stating yours is a low-cost thing to get right.
- This is the content layer we build in by default.
Method and limits. Several related analyses of overlapping datasets in one service vertical, run 2026-08 to 2026-09. The clinics and one outcome overlap across analyses, so these are not four independent confirmations, and we do not sum them into a single "data points" figure. Outcome definition: a business domain cited in an answer to any of the city's queries in a pass; "stable" = cited in at least two of three passes — domain citation, not a fact-level match. The recommendation engine was an AI search model with live web access over three passes, a proxy for AI search, not the default ChatGPT app. Review sample flow: 8,320 raw reviews, 7,484 with text, two clinics excluded, 7,389 scored, resolving to 174 clinics at our threshold; the analytic unit is the clinic, not the individual review. Runtime and model metadata were not fully saved, so exact engine and dates are reported values, not independently reproduced. Association, not causation, throughout. No promise of a position.