2 million listings: Google AI Mode shows the same product 21.6% more expensive, and 98.7% of your search-ranked products never appear

Somebody finally measured what these things do to buyers instead of to developers.

ProductRise analyzed more than 2 million product listings across over 100,000 SERPs and AI Mode responses in the US and UK between August 9 and 31, matching on Google's stable product identifier so the same query on the same calendar day gets compared like for like (ProductRise).

On matched products, AI Mode averaged 21.6% more expensive. Median price $149 against $100. The two disagreed on price for 38.1% of matched pairs, and AI Mode was the pricier side in 68.4% of those disagreements.

The number that should worry anyone with a storefront is the overlap. Only 1.28% of products ranking in traditional search appeared in AI Mode at all. AI Mode surfaces 3.9 products per query against traditional search's 27.8. And 49.6% of matched products came from a different main seller entirely.

That last figure kills the comfortable reading. If AI Mode were just a shorter list drawn from the same ranking, you'd expect the same sellers at the top. Half the time it's someone else. Whatever is selecting products in AI Mode is not the ranking you've been optimizing against for the last decade, and the price skew suggests it's weighting something correlated with higher prices. Structured data quality, retailer feed completeness, brand-name matching, merchant program participation. The study measures the effect, not the mechanism, and I haven't seen anyone credibly explain the mechanism yet.

The method holds up better than most of what gets published in this space. Same product identifier, same query, same calendar day, three weeks of collection. It's a vendor blog and the vendor sells a product-visibility tool, so read the framing with that in mind, but the matching design is the right one and the sample is large enough that a 21.6% gap isn't noise.

If you sell anything online, pull twenty of your highest-value queries and compare what AI Mode returns against your search rankings this week. Not to optimize for it yet, because nobody knows how, but to find out whether you're in the 1.28%. And if you're on the buying side, the practical read is that asking an assistant for a product recommendation costs you money in a way that asking Google for a product did not.

This connects to something Sylvain Kalache wrote on September 4 about incidents, and the connection is a general one about automation eating judgment. His argument is that as AI resolves routine incidents, engineers lose the repetitions that build judgment for the incidents automation can't handle. He grounds it in Bainbridge's 1983 "Ironies of Automation" and in aviation, where turbine engines see fewer than one in-flight shutdown per 100,000 engine flight hours and the FAA still mandates recurrent emergency training every six months. He cites TransAsia 235, where the crew misidentified which propeller had failed and stalled 117 seconds after the first warning (Sylvain Kalache). His prediction is falsifiable and I'd bet on it: average MTTR improves, resolution time for complex incidents gets worse. Same shape as AI Mode. The routine case gets easier and the tail gets harder, and nobody is measuring the tail.