What we learned auditing 262 live listings, and the three that were not the product at all.
In August 2026 we ran eighteen ordinary product searches through our production engine and read every listing that came back. Two hundred and sixty-two titles in total, across headphones, robot vacuums, laptops, televisions, kitchen appliances, power tools and footwear.
We were not hunting counterfeits. We were checking a narrower question: is this listing the product it appears to be?
A MacBook Air search returned an empty retail box at 36.95 dollars, ranked first, against a median of 947 dollars for that search. A Roomba search returned a side brush motor at 14.95 dollars, and a second listing described as a barebones unit sold bricked for board repair.
Each of the three carried the complete model designation of the product it belonged to. That is the mechanism worth understanding: a component or an accessory legitimately contains the full model number, so text matching alone cannot separate it from the product, and price ordering actively promotes it to the top.
Three out of 262 is a low rate. It is also not the right way to think about it, because all three appeared in the first position of their search. The rate that matters is not how many bad listings exist but how often one is the first thing you see.
A list ordered by price puts the least product-like listing at the top. The empty box was not hidden or disguised. It was the first result, exactly where a shopper looks first, and its title was accurate.
When a result sits dramatically below its neighbours, the useful reflex is to read the title word by word rather than to assume you have found something the market missed. Ask what noun the title actually names. Box, motor, module, brush, assembly, housing and unit for parts are all nouns that describe something other than the product.
We extended the vocabulary that recognises parts and packaging, then verified the change against all 262 titles before shipping it. The additions flag those three listings and nothing else.
Narrowness was the hard part. A rule matching the bare word box would have hidden every legitimate open-box listing, which is a real and often good purchase. A rule matching the bare word motor would have hidden real vacuums, because a Dyson title advertises a digital motor. Precision here is not fussiness; a careless rule removes more good listings than bad ones.
The rule also does not fire when your own search asks for the part. Search for a Roomba side brush and you will get side brushes, because the exclusion only applies when the query did not ask for a component.
The opposite failure is worth knowing too. In the same period we found a parts supplier listing a television at 99,999 dollars. Nobody buys it, but a listing like that distorts any calculation that uses the range of prices in a result set to judge what is normal.
Extreme prices in either direction are usually a signal that the listing is not the product being searched for, rather than a signal about the market.
Does the title describe the whole product or a component of it? Does the price make sense against the other results, in both directions? Is the seller the retailer itself or someone on its marketplace? Is the condition stated explicitly rather than implied? Are returns accepted, by whom, and for how long?
Do the photographs show the product or a box? Does the quantity say one? Does the listing mention compatibility with a list of models, which is how parts are usually described?
None of these require expertise. They require reading the listing rather than the price.