We ask you to trust our numbers, so here is how well they hold up when we test them. These figures are measured, not claimed. When one of them gets worse, it gets worse on this page too.
The measurement harness is built but has not been run against a hand-labelled set yet, so we are not going to print numbers here and imply they mean something. What follows is exactly what we will measure and how, published now so the method is fixed before we see the results rather than chosen afterwards to flatter them.
We take the people best placed to judge a car, owners who have kept it over a year and driven it more than 10,000 km, and treat what they think as the answer. Then we throw their reviews away and see how close we can get using only the ordinary reviews that are left.
This is not circular, because the target is defined by information the estimate never sees. If weighting by trust did not help, counting every review the same would do just as well, and the whole premise of this site would be wrong.
The baseline every review site uses today.
Fixed priors per source.
The default here.
| What we measure | In plain terms | Status |
|---|---|---|
| Right car, right variant | How often a review lands on the exact variant it describes. | Harness built, gold set pending |
| Right topic | How often a sentence is filed where a person would file it. | Harness built, gold set pending |
| Fake reviews caught | How much of a known-fake set gets pushed down before it can move a score. | Classifier not yet trained |
| Honest ranges | Whether our 80% range contains the answer 80% of the time. | Harness built, run pending |
| Serving speed | How long a verdict page takes to answer. | Measured: 29 ms at the 95th percentile |
Plenty of Indian reviews are written in a mix of Hindi and English, often typed in the Latin alphabet. Our current topic extraction is a keyword and rule system built around English words, so it understands those noticeably less well. A topic score built mostly from them is weaker than one built from English reviews. We would rather tell you that than average the problem out of sight, and it is the first thing a trained classifier has to fix.
None of this means our verdict is the truth. There is no such thing as the objectively correct opinion of a car. What we can show you is our working, so you can decide how much of it to believe. See how a score is worked out.