Google's AI Mode Recommends Products That Cost 21% More. And Perplexity's Sources Are Fake.

When an AI search tool hands you a "best pick" or a cited source, the question isn't whether it might be wrong. It's whether it's being fed by the same incentives that already distorted normal search — someone paying to rank higher, or someone building a page specifically to get quoted by the AI that's answering you.

Two studies published the same week point at exactly that failure. A study of more than two million product listings, circulated on Hacker News September 4 (396 points), found that Google's AI Mode recommends products averaging 21.6% more expensive than the same search run through traditional Google results — the AI layer is steering shoppers toward pricier options than the ranked list underneath it would. Two days earlier, Trellner Research reported (513 HN points) that Perplexity extensively cites 215,128 "best software" pages that are themselves machine-generated — content built for the specific purpose of getting picked up as an AI citation, not to inform a human reader. Perplexity's answers look rigorous: numbered sources, links, the appearance of research already done. A large share of what it's citing was written to be cited, not to be right.

The skill implication isn't "stop using AI search." It's that a recommendation or citation from an AI answer engine is a starting point, not a verdict — the same discipline you'd apply to one search result, not a synthesized answer that looks like it already did the checking for you. Two questions before you act on anything an AI tool surfaces as a top pick or a source: what is this ranking optimizing for, and would this page exist if no AI were ever going to read it?

Picture an operations lead at a 30-person logistics company shortlisting a new routing tool. She asks Google's AI Mode to compare three options and gets a clear top recommendation. If that recommendation engine is drawing from the same pool the price study measured, the "best" option may just be the one with the highest-margin listing or the most AI-optimized product page — not the cheapest or the best fit for a 30-person team. The check costs five minutes: run the same comparison through plain search results, not the AI summary, and see whether the top pick still holds once it's sitting in an unranked, unsummarized list next to its competitors.

Or picture a solo consultant who asks Perplexity for the best CRM for a five-person agency and gets an answer with eight citations. Some of those may be exactly the kind of page Trellner Research flagged — content produced at scale, formatted to answer that precise question, built to be machine-readable rather than trustworthy. The citations make the answer look more rigorous than a plain chatbot response would, and that's the trap: a numbered source list reads as evidence. It isn't evidence until you've opened at least one of those sources and confirmed a person wrote it for a person, not an algorithm.

Neither study says AI search is broken. Both say the "this has already been checked for you" feeling it produces is exactly the thing to be suspicious of: verify the pick, not just the polish.