How we measure — and when we admit a number is worthless.
A bit of honesty before we start: nobody can hand you a perfect number for your AI visibility. An AI’s answer shifts depending on who’s asking, when, and what mood the model is in today. Anyone selling you visibility to two decimal places either missed that — or is hoping you won’t ask. We do it the other way round: we show you how sure a reading is, and we tell you when not to trust it.
We ask many times, not once.
Asking ChatGPT once is a snapshot, not a measurement. So we put the same question to each engine repeatedly and watch how much the answer moves. That becomes a confidence score — a range, not a point. With us, “22%” means “22% ± 4 points across the observation window.” Less marketing-friendly. More true.
For every engine we run 49 buyer-intent prompts, three times each. Your share of answer is the average across all runs — and the spread between them becomes the confidence score in the next section.
One run said #2. Another said #5. Only the spread tells you which number you can take to your CMO.
One prompt → 9 engines × 3 runs each — the spread across runs is the Determinism Index.
Share of Answer, counted out of 100.
Once the runs settle, we count how often a brand is named across every answer. Share of Answer is that count out of 100 — like market share, but for AI answers instead of search clicks.
A brand named in 31 of 100 answers holds 31% Share of Answer. The other 69 go to competitors or no one — the conversation you never see in analytics.
A confidence score on every single number.
The Confidence Score measures how much an answer moved across its three runs. 1.00 means the engine said the same thing every time; lower means it’s guessing. Every score, chart and chat answer in Quolens carries its Confidence Score — this is the instrument no other GEO tool has.
Toggle the tiers above to see the readout move. This trust range is the only place a judgment colour ever touches a number — everywhere else, colour is identity, not verdict.
±0.05 across 3 runs · salesforce.com
From memory or looked up live — and the honest catch.
We separate two things almost everyone mashes into one number: what the AI “knows” about you from training (memory), and what it finds live on the web right now. Both weak means nobody knows you. Only the live picture weak means your content is stale. The honest catch nobody else mentions: this split only works where the engine actually searches the live web. If a model answers purely from memory, there’s no live picture to split off — and we’ll tell you exactly that, instead of inventing one.
Salesforce is named far more often when the engine searches the web than from training memory alone. That 24-point gap is the GEO roadmap — get cited on the pages engines reach for, and the model number climbs to meet the live one.
Concretely, as it stands today — where we can split, we split; where we can’t, we say so:
Here you get the full split.
No memory picture to compare against — the answer is always live.
The model answers from training; there’s no live picture.
When a reading wobbles, we say so.
If a number only moves inside that range, it’s neither progress nor a crash — it’s noise. We flag it, and we don’t raise an alert on it. You should act when something really moves, not every time the model slept badly. A reading we don’t trust ourselves gets a red mark from us, not a nicely rounded figure.
We measure what AI tells everyone — not your private chat.
AI answers are partly personal. ChatGPT answers you differently than it answers your customer, because it knows your histories. Nobody can see into those private chats — us included. So we measure the generic consensus answer: what AI says when it doesn’t know you. That’s the version your new buyers see first, and the one that shapes their shortlist. We don’t pretend to know every individual case. We show you the picture most people get.
What we don’t promise.
We can’t steer the AI, and we won’t guarantee you a placement. Anyone who does is selling snake oil. What we can do: show you honestly where you stand, how reliable that number is, and the specific move that improves your odds of being named more often. What the AI does with that is the AI’s call. We’re the instrument, not the lever on the machine.
What’s in every reading
What we stand on.
That a single AI number misleads isn’t a claim from our marketing team — the research says it. The field was formalized in 2024 (the GEO paper, Aggarwal et al., KDD 2024). A recent statistical study (Sielinski 2026, “Quantifying Uncertainty in AI Visibility,” arXiv:2603.08924) uses repeated sampling across Perplexity, SearchGPT and Gemini to show that single-run metrics look misleadingly precise — AI visibility belongs reported with uncertainty, as a confidence interval, not a point value. That’s exactly how we build it.
These are our cited sources, not endorsements — Quolens is pre-launch and we won’t imply partnerships we don’t have. We name our influences plainly because honest measurement starts with an honest bibliography.
What this data can — and can’t — tell you.
See your first reading in 60 seconds — no login.
Watch the repeated runs, the Confidence Score settle, and the grounding split — on your own brand, free.