Accuracy
A false alarm costs more than a missed one.
If a tool tells you to fix something that was never broken, you break a working setting and you stop believing the tool. Everything here exists because we caught ourselves doing exactly that.
Three answers, not two
Every comparison can end in three ways: the two sides agree, the two sides disagree, or we could not tell. Most tools have only two buckets, so the third quietly becomes the second.
We keep the third. A product where a check could not be completed is not counted as healthy and not counted as broken — and the health figure does not quietly improve by dropping it.
False alarms we found and removed
Each of these was producing a critical, confident, wrong finding on real stores.
| What happened | What we did |
|---|---|
| A page published no price at all, and the tool reported it as a mismatch | Missing data is now reported as unknown, not as disagreement |
1.500 was read as one and a half, so an identical price looked like a mismatch | Ambiguous number formats are refused, not guessed — some currencies really do use three decimals |
A page said oos, the feed said out_of_stock — the same state, reported as a conflict | Unrecognised values are reported as unknown instead of assumed to be different |
| A page returned a rate limit, and the tool said the product page was unreachable | A server refusing us is no longer treated as evidence about your page |
| Two currencies were compared without knowing which country the feed targets | A currency difference alone is no longer called a mismatch |
Why we are telling you this
Publishing your own false alarms is unusual for a product page. But the only thing a diagnostic really sells is whether you can believe it, and a list of removed mistakes says more about that than an accuracy percentage we would have to invent.