Cited, Not Chosen
The go-to advice for winning AI search is to publish your own comparison page. We analyzed 1.07M AI answers across 180 brands to test whether it actually works, because getting cited by AI is not the same as getting chosen.
AI can quote your page and still recommend your rival.
A citation means the AI read your page. A recommendation means it picked you. We separated the two across 1.07M AI answers, then asked the harder question: when you publish your own comparison page, does your chance of being recommended actually go up?
Short answer: yes, but more modestly than the hype suggests, and only where you would expect. A newly cited page lifts recommendations by a couple of points overall, noticeably more for challengers, and the entire effect lives in the engines that can read the live web. Below, the gap, the proof, and the catch most brands miss.
Cited A Lot, Chosen Rarely
Start with the problem a comparison page is meant to fix. Plenty of brands get quoted as a source and still miss the shortlist, challengers most of all. Each brand's overall visibility sits next to the share of its cited prompts where it was a source but never made the top three picks.
Known public benchmark rows are named; smaller or unknown rows are anonymized. Visibility is the brand's overall AI visibility score. Cited, not chosen is the share of prompts where the brand's own page was cited as a source yet the brand did not land in the top three picks.
Do Comparison Pages Work?
So you publish the page. Does it actually move anything? We found 1,770 moments where a brand's own page was first cited, then compared each brand against never-cited controls in a staggered difference-in-differences design. The answer is yes, but modestly, and the gains land most with the brands that had the most room to climb.
Each point is the average treatment effect (ATT): how much a brand's weekly recommendation rate moved, measured against never-cited control brands, relative to the week its page was first cited (week 0). The flat pre-period is the parallel-trends check; the step at week 0 is the page doing the work. Shaded area marks the weeks after the page appears.
How We Know It's Real
A jump right after citation could just be coincidence, so the data carries its own control. Some engines fetch the live web to answer; others reply from memory and cannot see a brand-new page. If the page is really doing the work, only the engines that can read it should move. That is exactly what happens.
Web-grounded
+5.03ppFrom memory (cannot read the page)
+0.19ppYour 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. This is the share of each brand's cited pages that give the AI at least one rival to suggest.
What To Do
It works, just modestly, and most for challengers with room to climb. Expect to be cited a while before you are chosen.
The entire lift comes from engines that fetch the live page. If your page is not easy to retrieve, it does nothing.
Every competitor you list becomes the evidence AI uses to recommend them. Make the case for you, not a directory of everyone.
A citation is a foot in the door, not a win. Measure whether you are the pick, not just whether you were quoted.
Methodology
We built a weekly panel of recommendation rates from 1,073,831 AI answers across 180 brands, then dated each brand's owned pages to the week they were first cited. Using a staggered difference-in-differences design with never-treated controls and 8-week windows on either side, we estimated the change in recommendation rate, with brand-clustered bootstrap confidence intervals and a Callaway-Sant'Anna event study as a cross-check.
- ·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.
See how your brand performs in AI search