Skip to content
Relevant.aiRelevant.ai

Measurement, not opinion.

Answer engines publish no ground truth: no official query volumes, no impressions, no stable answers. So we estimate what matters, then measure it like an experiment. This page explains how, and why the difference matters.

The measurement problem

Most tools ask you which prompts to track, read each one once, and report the result as a score. A prompt list you wrote yourself is not a measurement. It is an opinion with a dashboard attached.

Relevant treats visibility the way good statistics treats anything you cannot see directly. We work out which questions actually matter, weighted by real prompt volumes, ask each one many times (that is sampling), and report the answer with a clear sense of how sure we are.

Same brand · same question · six readings · illustrative
  • "Best running shoes for flat feet", Bengaluru IP, default model#2
  • Same words, Delhi IP#5
  • Same question, Hinglish phrasing#7
  • Reworded: "which running shoes suit flat feet"Absent
  • Same words, reasoning tier instead of default#1
  • Asked as turn 3 of a shoe-shopping chatAbsent
Where a weekly “drop” actually came from · illustrative breakdown
Phrasing34%
Geography27%
Model22%
Real change17%

Only the last slice is genuine change. Most of a weekly drop is not a drop. We break it down before we report it.

The wobble is the signal

The same question, asked slightly differently, returns a different answer. Statisticians call this variance. Most tools average it away; we design experiments around it, because knowing where your visibility swings around is what makes it fixable.

Four principles behind every number

Break it down, don't average
Every score is reported by geography, language, persona and model tier. We never collapse it into one number that hides where you are strong and where you are invisible.
Explain every swing
When your score moves, we work out how much came from phrasing, geography and the model, and how much is genuine change, before anyone panics.
Always say how sure we are
Every number ships with a confidence interval: a range that shows how much it could move by chance alone. If we have not asked enough times to call a move real, we say so.
Blind spots are a to-do list
"Strong in Bengaluru English, invisible in Hinglish" is not noise. It is something your team can act on this week.

Recommendations are experiments

Every fix we recommend is tracked closely. We already have the before, and we measure the after the same way every time, so the improvement you see is really from the fix, not from a different model or the time of year. Fixes that work rise. Fixes that don't, fall. The advice gets sharper every month.

Built to survive model upgrades

When a new model ships, scores reset for everyone. What survives is what we keep on record: what the engines used to say, which questions matter, and evidence of which fixes moved the answer. A model upgrade resets the score. It doesn't reset what we've learned.

Be statistically Relevant.

AI's answer will keep changing. Your numbers shouldn't.

Book a demoExplore the platform