How we check our own work.
A program that reads dealer websites is only worth paying for if it's right. So before anything it reads reaches you, it's tested, and the test never stops.
The proof
700+
dealer websites read, in 47 states, and counting
3,000+
cars checked one at a time against their own pages
99%+
of printed findings right in our own check, all graded test rounds combined
The test runs every night on about 40 dealer websites it has never read. We update the counts on this page by hand when we rebuild the site, so every count below carries its date.
478 of 481
findings right in our last five test rounds, as of Oct 25
99.4%, each checked against the car's own page by a second, separate read
98 in 100
the bar a change to the reader must clear before it goes live
tested on dealer websites it has never read
Every audit
read by a person before it goes out
and you approve every change
Before any change to the program that reads your listings goes live, we run it on dealer websites it has never seen, and a second, separate read of each car's own page checks what it printed. If it misses the bar, it doesn't ship. That test now runs every night, on about forty websites it has never read.
The test, in plain words.
Fresh websites every night. Every night the reader runs on about forty dealer websites it has never seen, picked from across the country, the way a real audit runs. None of them is told; nobody is contacted; every website's rules for automated reading are honored.
A second, separate read. For each of those lots, a separate AI read opens the cars' own pages and writes down what they show: the price, the miles, the photos, the VIN, the words, the sold mark. That's the answer key. The reader's audit is then graded against it, finding by finding.
The gate on every audit. Before any audit goes out, two of the dealer's cars are checked against their own pages. A field found wrong is dropped for the whole lot: no score from it, no finding from it. A wrong price holds the audit for a person. If we can't confirm it, we don't print it.
The bar. At least 98 of every 100 findings printed on the test cars must be right. A wrong price on an audit that would go out: no more than 1 audit in 100. Audits held for a person: 1 in 10 or fewer. And a lot read short of what the website says it has must say so in the audit. A round that misses the bar gets a fix, a re-check on every car in the answer key (3,756 as of Oct 19, more every night), and a new locked version of the reader. Until then, the old one runs.
The result so far. In the last five test rounds, as of Oct 25, 478 of 481 findings printed on the test cars were right (99.4%). Across every graded round since Oct 19, 908 of 913 (99.5%). Every miss goes on the fix list.
And a person. A used-car manager reads every audit before it goes out, and you approve every change. The program is what makes it affordable. The reading is what makes it right.
What the counts are, and aren't.
Every number we print is a count of what a listing shows, checked against the page. We never say what shoppers do, think or skip, because we see listings, not shoppers. We never state what a car is worth. We never name another dealer, and the Listing Check we send you is never shown to another dealer. The live check on our homepage is the one thing that is not private: it is built only from pages on a website that anyone can open, so anyone who enters that website sees its check.
716 dealer websites, 2,966 cars, checked one at a time, October 2026
The number one gap isn't photos. It's words.
53%
of the cars had under 25 words about the car.
42%
had no description at all.
22%
met both of our standards: 20 photos and 60 words.