How a score is worked out

In plain language, with the actual arithmetic shown. If you disagree with any step you can see exactly where to disagree, which is the point.

  1. 1

    We collect the evidence

    Owner reviews, expert reviews, forum threads, video transcripts and official records, read at a polite rate, cached, and linked back to. We store references and structure, not copies of anyone's writing.

  2. 2

    We match every review to one exact variant

    A petrol listing is never matched to a diesel variant, whatever the text similarity says. Specifications act as hard constraints before any learned matching runs. Anything ambiguous waits for a person instead of being guessed at.

  3. 3

    We split each review into nine topics

    A review is rarely about one thing. One paragraph can praise the ride and condemn the service centre, so we score those separately rather than averaging them into a single star rating.

  4. 4

    We work out how much each review should count

    This is the part that makes Revix different, and it is spelled out below.

  5. 5

    We combine, and we say how sure we are

    A weighted average, plus a range that widens when the evidence is thin or lopsided.

What decides a review’s weight

Every review gets a weight between 0 and 1 for each topic. It is a product of six terms, so any single bad signal pulls the whole weight down.

weight = source_prior
       × (1 − spam_probability)
       × reliability
       × aspect_fit          ← the interesting one
       × recency
       × launch_window

Five of those are ordinary. aspect_fit is the one that makes the argument, and it exists only because this is the automobile domain.

Why the same person is a good witness to one thing and a bad witness to another

An owner who has covered 500 km can tell you a great deal about the showroom, the delivery experience and the first service appointment. They can tell you nothing useful about whether the clutch survives 60,000 km. An owner at 60,000 km is the reverse.

Indian review platforms record ownership duration and kilometres driven, and then throw that information away by averaging every review equally. We use it. Credibility is therefore not one number per review; it is a short vector across topic groups.

{ "base": 0.71,
  "by_aspect_group": {
      "durability":  0.88,   ← trusted on long-term reliability
      "immediate":   0.42,   ← less so on first impressions
      "service":     0.83,
      "efficiency":  0.79
  } }

Why the range is not decoration

When reviews carry different weights, the number of reviews stops being a useful measure of how much you know. Two hundred low-weight reviews can carry less real information than thirty high-weight ones. The standard way to express that is the effective sample size:

n_eff = (Σ w)² / Σ w²

This is why switching to trust weighting makes the range wider, not narrower. Being more careful about which evidence counts means admitting you have less of it. You can watch that happen on any verdict page: flip the switch and the effective sample drops while the interval opens up.

Why we show disagreement instead of hiding it

When owners split on a topic we do not just report the split, we try to explain it. For each topic we test which characteristic best accounts for the disagreement: where the review came from, whether the owner is verified, how long they have owned it, how far they have driven it. That produces statements like:

54% of the disagreement about build quality is explained by how far people have driven. Under 10,000 km they rate it 5.1. Over 40,000 km they rate it 9.1.

That is a statistical decomposition of the spread, not a language model’s opinion. It is cheaper, more reliable, and far more useful to somebody choosing between two versions of the same car.

What we do not claim

We do not rank by who pays us, because nobody pays us. Where we do not have enough evidence, we show no score at all. See how accurate we are.