Two audits show AI shopping answers are gamed and wrong

Two research teams published audits of AI shopping recommendations on the same day, September 2, without citing each other, and they land on opposite ends of the same problem.

Trellner's report traces 215,128 templated "best software" pages across three interlinked domains that answer engines pull from as sources. Almost 60% of the grounded citations in those recommendations come from sites outside the top 100,000 by traffic, meaning the pages an AI cites as authorities are often obscure template farms built to be cited, not read.

Haus Research looked at the other side of the same pipe: whether the answers those citations produce are even accurate. Over a third of citations that included a specific figure, 34.7%, pointed at a source that didn't actually support the number. Questions about entry pricing, the kind a buyer asks before signing up for anything, only got a correct answer 62.6% of the time.

Put together, the picture is a discovery channel that's both manufacturable and unreliable at once. Someone can build 215,128 pages designed to get cited, and even when a real source gets cited, there's a one-in-three chance the number attached to it is wrong.

For anyone selling software, that means the page an AI decides to quote for your price isn't necessarily yours. If your own pricing page isn't structured so a model can parse it without ambiguity, a templated third-party page is sitting there ready to get cited in its place, correctly or not.

Neither report says whether the platforms doing the citing have fixed anything since these numbers were pulled, which is the open question worth checking next.