Productrise ran the same shopping queries through Google AI Mode and traditional search on the same day, every day for three weeks in August, and found identical products priced 21.6% higher in AI Mode. That is the number every write-up led with, and it is the least interesting figure in the dataset. Further down sits one that should concern anyone selling products online a great deal more: only 1.28% of the products appearing in traditional search also appeared in AI Mode, for the same query, on the same day. These are not two rankings of one shelf. They are, for practical purposes, two different shops.
TL;DR
- Productrise tracked more than 2 million product listings and 100,000 search responses across the US and UK between 9 and 31 August 2026, comparing Google AI Mode against traditional search on identical same-day queries.
- Matched products were 21.6% dearer in AI Mode on the lead offer. Prices disagreed on 38.1% of matched pairs, and AI Mode was the higher of the two in 68.4% of those disagreements.
- The more consequential numbers are structural: 1.28% product overlap between the two surfaces, and 3.9 products shown per AI Mode response against 27.8 in traditional search.
- The lead seller differed on 49.6% of matched products, which says price has been demoted as the tiebreak and something else is deciding placement.
- EU law already requires published ranking parameters (P2B Regulation Article 5) and a non-profiling recommender option from designated search engines (DSA Article 38). Neither has yet been tested against AI Mode.
What was actually measured
The study covers 23 days, 9 to 31 August 2026, in the US and UK, restricted to product-centric shopping queries. Products were matched across the two surfaces using Google’s stable product identifier where one was present, for the same query on the same calendar day, and the comparison was between lead offers: the first listing shown on each side, not the cheapest available.
Two caveats matter, and they are the study’s own. The traditional-search side was limited to the popular_products carousel, so this is carousel versus AI Mode rather than the whole of Google versus AI Mode. And the lead offer is a display position, not a market price. A merchant reading this as “Google has put its prices up” has misread it: Productrise attributes the gap to a change in ranking weighting, not to inflated prices. Google has not commented.
The shelf did not get dearer. It got shorter.
An AI Mode response carries 3.9 products on average. The traditional carousel carries 27.8. That is roughly a sevenfold reduction in the number of merchants who get to exist in a given answer, and it is the mechanism behind most of the price finding. When a ranking system that showed twenty-eight results shows four, the cheapest offer is no longer guaranteed a seat by arithmetic alone. It has to be one of the four best on whatever the new criteria are.
This is why the full-catalogue comparison is even starker than the matched one. Across every product listed on both sides rather than just the matched pairs, the median AI Mode price is $149 against $100 in traditional search, roughly 49% higher. That is not the same product costing more. That is a different, dearer selection of products being put in front of the buyer.
The 1.28% is a measurement problem before it is a commercial one
Most e-commerce teams monitor their position through rank tracking, Shopping campaign reporting and impression share. All of it describes the traditional surface. If only 1.28% of products cross over, then that entire apparatus is reporting on an index that is very nearly disjoint from the one AI Mode draws from.
The practical consequence is worse than a blind spot, because a blind spot at least announces itself. Absence from AI Mode currently looks identical to not having measured AI Mode. A merchant whose rankings are healthy, whose feed is approved, and whose Shopping metrics are stable can be entirely missing from the surface Google is promoting hardest, and no dashboard in the stack will raise a flag. The first signal is a revenue line that softens for no reason anybody can name.
Price stopped being the tiebreak
The single most useful number in the study is the one attracting the least attention. On 49.6% of matched products, the main seller was different between the two surfaces. Same product, same identifier, same day, and half the time a different merchant won the lead position.
If price were the dominant input on both sides, that figure would be far lower, because price is a fact about an offer that both systems can see equally well. Half the leads changing hands means the deciding inputs are ones the two surfaces weight very differently. The evidence points at everything wrapped around the offer rather than the offer itself: the completeness and accuracy of the product feed, whether attributes such as material, dimensions, compatibility and model number are populated as structured fields rather than buried in a description, whether the product is resolvable as an entity with consistent identifiers across the catalogue, and how legibly shipping, returns and availability are expressed.
That is a genuine opening for anyone who has been losing on price, and the practitioners reading this optimistically are not wrong to. A merchant with better data about its own products can now beat a cheaper merchant with worse data. The uncomfortable corollary is that the work has moved from a commercial lever any founder can pull in an afternoon to an engineering one that requires the catalogue, the feed pipeline and the site’s structured data to actually agree with each other. Most catalogues we are handed do not.
It rhymes with what we found measuring which tools coding agents choose: capability did not predict selection, familiarity and output shape did. Being the best offer does not get you selected. Being the most legible one does, and legibility is an engineering property, which is more than can be said for a race to the bottom on margin.
The transparency obligation nobody has invoked
The regulatory question under this is not a new law waiting to arrive. It is two that already apply.
Article 5 of the P2B Regulation (EU) 2019/1150 requires providers of online search engines to publish a description of the main parameters determining ranking, and their relative importance, in plain and intelligible language, and to keep it up to date. The article explicitly does not require disclosure of algorithms, so this is not a demand to open the black box. It is a demand to say what the main parameters are. If lowest price has been materially demoted as a ranking input on a surface that shows shopping results, that is a change to a main parameter, and business users are entitled to a description reflecting it.
Separately, Google Search is a designated very large online search engine under the Digital Services Act. Article 38 requires designated search engines using recommender systems to offer at least one option not based on profiling, and Articles 34 and 35 oblige them to assess and mitigate systemic risks arising from their ranking systems. A conversational surface that curates four offers from a possible twenty-eight is a recommender system by any reasonable reading.
None of this makes the finding an infringement, and we are not claiming it is. It does give business users a lever most do not know they hold. In Ireland, Coimisiún na Meán is the Digital Services Coordinator; P2B ranking descriptions are pursued through the courts and representative bodies rather than one regulator. Merchants trading into the EU are better positioned here than they think.
What to do this quarter
- Measure the surface directly. Stand up a harness that runs your top revenue-driving queries through AI Mode and traditional search on the same day and logs both result sets. Track AI Mode presence as its own metric with its own owner.
- Audit the feed as data, not as marketing. Count the products in your catalogue with populated structured attributes, not just approved status. Approved means valid. It does not mean complete.
- Reconcile identifiers across systems. GTINs, MPNs and brand values that disagree between the feed, the product page markup and the ERP are the cheapest thing here to fix and the likeliest to be silently costing you matches.
- Put a number on the exposure. Estimate what share of your discovery-driven revenue depends on a surface you currently cannot see into, and say it out loud to whoever owns the forecast.
REPTILEHAUS builds and instruments e-commerce platforms, product data pipelines and the structured data layer underneath them, including the AI visibility measurement described above, for clients who would rather find out from a dashboard than from a quarterly. If your catalogue is now judged on data quality instead of price, it is worth knowing what your data says. Get in touch.
📷 Photo by Clark Street Mercantile on Unsplash
