Comparison Pages
The go-to advice for winning AI search is to write your own comparison page. But being cited by AI is not the same as being recommended: a model can quote your page as a source and still pick a rival. Does the page actually close that gap? This study separates the descriptive result from an exploratory estimate that does not pass every robustness check.
Cited a lot, chosen rarely
Start with the problem a comparison page is meant to fix. These brands are quoted as a source but left out of AI’s top picks. The less visible a brand, the more its citations leak: cited for evidence, skipped for the recommendation.
What changed around first citation?
An exploratory staggered difference-in-differences estimate compares recommendation rates around the first observed citation of an owned page with brands not yet cited. First citation is not publication, and the design did not pass every robustness check.
Exploratory percentage-point estimate by week relative to the first observed citation. It is not a causal effect: publication timing was unavailable, placebo estimates were positive, and a publish-anchored estimate crossed zero.
The engine split is suggestive, not proof
The exploratory estimate was larger for engines that retrieve live pages than for engines answering without live retrieval. This pattern is consistent with a page contribution, but a borderline pre-trend and other failed checks prevent a causal conclusion.
The lift appears only in engines that can read the page. Engines answering from memory show nothing, which is what you would expect if the page itself is doing the work. Bars are the difference-in-differences estimate; whiskers are the 95% brand-clustered confidence interval.
Your own page sells your rivals
Here is where brands hand the gains back. The same comparison page that helps you also names your competitors, and AI reads those names as a vetted list of who else to recommend. Share of each brand’s cited pages that give AI at least one rival to suggest.
A weekly recommendation panel for 180 brands, built from 1.1M AI answers, with a staggered difference-in-differences and a Callaway-Sant’Anna event study around the first citation of an owned page. Controls are brands not yet cited; windows run eight weeks either side; confidence intervals are brand-clustered bootstraps. The study is not randomized, first citation is not publication, the publish-anchored estimate crosses zero, placebo estimates are positive, and the grounded-engine pre-trend is borderline.
- Quasi-experimental, not a randomized trial.
- Publish-date anchoring could not be completed (archival publish dates were unavailable), so we date the effect to first citation, not first publication.
- A mild pre-trend exists in grounded engines (joint pre-trend test p=0.05), consistent with the citation timestamp lagging true onset.
Common questions
Is being cited by AI the same as being recommended?
No. A model can quote your page as a source and still recommend a competitor as its top pick. In this study, owned pages are a median 4.6% of what AI cites about a brand, and nearly a third of cited pages name a rival. Citation is evidence; recommendation is the verdict.
Why does AI cite my page but recommend a competitor?
Comparison and "best of" pages are the pages AI reaches for, and those pages name rivals by design. So the page you wrote to win becomes the evidence the model uses to list everyone else. The less visible a brand already is, the more its citations leak into a competitor recommendation.
Do comparison pages actually help you get recommended?
The study found a 2.7-point exploratory estimate around the first observed citation of an owned page. It is not treated as causal: publication timing was unavailable, a publish-anchored estimate crossed zero, placebo estimates were positive, and the grounded-engine pre-trend was borderline.
What did the engine split show?
The exploratory estimate was larger for web-grounded engines than for engines answering without live retrieval. That pattern is suggestive, but it does not establish that the page caused the recommendation change.