Ask an AI twice, you get two answers.
We measure it anyway.
Findrix tells you what AI engines say about your brand — honestly enough to act on. This page shows what we guarantee about every number: how we ask, what we count, and what we refuse to claim.
AI visibility is easy to fake and hard to measure.
Three facts break every naive tracker. Everything on this page exists to survive them.
Same question, different answer
Ask an engine the same question minutes apart and the brands, the sources, and the order can all change. One run is a screenshot, not a measurement.
There is no search volume
Nobody can tell you how often people ask ChatGPT about your category. That number does not exist anywhere — a tool that sells it is selling a guess.
Averages that describe nobody
ChatGPT and Gemini behave like different species. Pour them into one blended score and you get a number that describes neither — and hides which engine actually moved.
Your market, drawn as the questions buyers ask.
Findrix doesn't track keywords. It builds a map of your market: the questions real buyers ask on the way to a purchase — from someone who just found the problem, to someone comparing options, to someone double-checking a name they already trust. A question earns its place by covering ground the others don't; twenty rephrasings of the same question teach you nothing new, so they don't get in.
You see every question before anything runs, and nothing enters measurement without your approval. You know your buyers better than any tool does.
Then the map freezes. The same fixed set runs week after week, so this Monday compares cleanly to last Monday. If the set ever changes, the trend is cut and restarted — a new question set is never allowed to masquerade as movement.
Asked where your buyers actually ask.
We measure up to seven surfaces: ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode, with Claude and Grok as add-ons. Google alone counts as three — your buyer meets a different answer in AI Overviews, in AI Mode, and in Gemini.
Core engines are read from the consumer answer page — the one your buyer sees — in controlled sessions with no history, no cache, and nothing carried over. When a surface won't answer, we fall back to the engine's official interface rather than losing the week. Either way, every answer is stored with the engine and the model that produced it.
Each engine lives in its own lane, measured separately and reported separately. The only cross-engine figure you'll ever see is a clearly labeled average of those separate reads — never a pooled blend that hides which engine moved.
First, we learn how noisy your market is.
Before your first report, Findrix runs a calibration pass: every question in your set is asked again and again, per engine, to measure how much the answers move on their own — your market's noise, measured before we dare to measure you against it.
That noise fingerprint becomes the yardstick. Every later week is judged against it, which is what keeps a weekly read honest. You never see week-one jitter dressed up as a trend.
Every number ships wearing its uncertainty.
No metric stands alone on the dashboard: under the number sits its range — a confidence interval, clustered the way working statisticians do it, so that asking the same question again never counts as new evidence.
The dashboard tells you how to read every band in plain words. A confident read is tight enough to act on. A directional read means trust the direction, not the decimal. And where an engine wasn't measured, you see ∅ — never a decorative zero.
A number without its range is a guess wearing a suit. We don't sell suits.
A real move is a verdict, not a vibe.
Week to week, every metric carries its band. Findrix declares a real move only when this week's band and last week's band don't touch — when the noise explanation is dead, not merely unlikely. Anything less prints as stable, and we say so out loud.
That standard costs us drama. Dashboards that celebrate every wiggle look more alive than ours — and are wrong more often, in both directions.
Your move, minus the market's.
Engines re-rank whole markets at once — a model update can lift or sink every brand in your niche in the same week. A tracker that watches only you will sell you the tide as your own swimming.
So Findrix watches your niche alongside you: the same questions, the whole competitive field. When the tide moved everyone, your report says so. When you moved and the market didn't — that one is yours, and only then do we say it. And when the comparison can't clear its own checks that week, we credit nothing.
Seven things, counted per engine.
Per engine, per week, each with its range — and ∅ where we didn't measure.
Measure. Fix. Measure again.
A visibility gap on its own is trivia. Findrix turns it into a concrete change — the exact schema block, the exact page edit — backed by a 29-point technical audit and the citation data from your own market. You review it. You approve it. It ships. It can be rolled back.
Then the same fixed set runs again, on the same engines, the following week. The fix is judged on the exact questions it was meant to move — against the calibrated noise yardstick, under the real-move standard, with the market's weather accounted for. If it worked, you'll know. If it didn't, you'll know that too.
The vows.
- Never call noise growth.A wiggle inside the band prints as stable — however good it would look on a chart.
- Never pool engines.Measurement lives per engine. Only labeled averages of separate reads exist, and they say so.
- Never invent search volume.If a number is unknowable, we don't sell it.
- Never ship a bare number.Every metric carries its range, spelled out in plain words.
- Never swap the question set under a trend.A set change cuts the chart and says so. Trends never span two different maps.
- Never dress a gap as a zero.Unmeasured is ∅, and ∅ is what you'll see.
- Never take credit for the tide.Market-wide moves are named as market-wide — even when the other story would flatter us.
An earlier version of this page printed equations. We've stopped: the statistical design is the instrument, and the instrument is the product. What we publish instead is the part you can hold us to — the vows above, each one checkable from your dashboard. If your data team is doing diligence and wants the internals, we'll walk them through the full design under NDA: team@findrix.ai.
Asked by every skeptic we like.
Why does my own ChatGPT answer differently?
Your ChatGPT knows you — your history, your location, your custom instructions. And even a clean session answers differently run to run. Your chat is one draw; Findrix reports the pattern across a controlled sample, with the range printed under it.
Other tools show one confident number. Why do you show a range?
The one confident number has a range too — it just isn't printed. Ours is, and it's sometimes wider, because we refuse to count the same question asked twice as twice the evidence. An honest wide band beats a flattering tight one: the first tells you what you know, the second tells you what you want.
My mention rate came back near zero. Is that a bad grade?
It's a starting line, not a grade. Plenty of strong brands start near zero because the engines haven't met them yet. Moving that number — and proving the move was real — is what the loop is for.
Can I game the score without actually fixing anything?
Briefly, maybe. The metrics track what engines actually do with your brand, so a cosmetic bump won't survive the next measurement. The score is a thermometer. Treat the fever, not the reading.
Where did the formulas go?
Into the product. We publish the guarantees and keep the machinery private — promises you can verify beat equations you can't audit anyway. If you're evaluating us seriously, we'll go as deep as you like under NDA.
Check it on your brand.
You've read what we promise. Now watch it run: start a project, then open ChatGPT yourself and ask about your category. The dashboard and your own eyes should tell the same story.