Your buyers are asking AI who to buy, long before they ever request a demo.See how we measure it →
Pick44Pick44
Sign in
Methodology

How we measure AI visibility — in the open.

No black box. Every score traces back to a real answer from a real model, weighted by a published formula and labeled with a confidence tier. We measure how AI already talks about you — we never manipulate it.

OBSERVING RECOMMENDATION SHARE ACROSS
OpenAIGPTClaudeClaudeGeminiGeminiPerplexityPerplexityMetaLlamaMistralMistralNvidiaNemotronDeepSeekDeepSeek
8 models · last tested 2h ago
01
Define the query set
We compile real buying questions for your category — the high-intent prompts where a recommendation decides a deal.
02
Observe every model
Each query is run across ChatGPT, Claude, Gemini and Perplexity. We record the full answer, not a guess.
03
Extract rank & mention
A parser reads each answer for whether you appear, your position, how you are framed, and which sources are cited.
04
Score & weight
Signals are combined into one weighted AI Visibility Score, normalized across models and labeled with confidence.
05
Monitor & alert
We rescore weekly and alert you to rank drops, factual errors, and new competitors entering the answer.
The scoring formula
AI Visibility =
40%Mention Rate
25%Average Rank
25%Recommendation Strength
10%Source Coverage
Proof tiers

We label how strong the evidence is.

1
Observed in sample
We measured it in this run of queries.
2
Moved after intervention
The score changed after a truthful fix shipped.
3
Persists over time
The improvement held across repeated retests.
4
Correlates with traffic
Visibility gains track real referral traffic.
5
Correlates with conversion
Visibility gains track pipeline and revenue.
High confidence
Consistent across models and reruns
Medium confidence
Some variance between models
Low confidence
Sparse or conflicting signal
Insufficient data
Too few observations to score responsibly — we say so rather than guess.
What we measure
  • ✓Whether AI mentions you for real buying questions
  • ✓Your rank/position when you are mentioned
  • ✓How positively you are framed (recommendation strength)
  • ✓Which sources AI cites about you
  • ✓Factual errors, stale claims, and missing facts
  • ✓How all of this changes over time
What we do not claim
  • ✕That we can guarantee an AI ranking
  • ✕That we manipulate or game model outputs
  • ✕That rank can be purchased — it cannot
  • ✕That a single run is universal truth
  • ✕That model outputs are always accurate
  • ✕Any legal, financial, or compliance advice
The trust contract

Numbers that survive you checking them.

Every score expands to the exact prompts, models, and raw answers that produced it — re-run any prompt in the actual model and you should see what we saw. These are the rules that make that possible, stated precisely enough that you could catch us breaking them.

Median, never best-case
Your headline rank is the median across models, with "not mentioned" counted as worse than any rank. One friendly model can never carry your score, and we never quote the best model as the number.
Deltas only between same-model runs
Every run is stamped with its exact model set. A trend or "+N vs baseline" is only ever computed between runs on the same models — switching model stacks resets your baseline instead of manufacturing an improvement.
How we ask: API calls, disclosed
We measure via the models’ APIs with a neutral prompt and (when available) live web grounding — not a logged-in browser session. That differs from what a specific signed-in user might see (memory, custom instructions, A/B variants), so every report is labeled live-web grounded vs training-recall, and every prompt is shown so you can replay it in the consumer app yourself.
Owned vs third-party sources
When we count citations, your own domains are labeled "owned" and excluded from third-party coverage — your blog citing you is not independent proof, and we never let it read as such.
What we will not do
  • ✕Show a relative gain without its absolute score ("+200%" of a tiny number is still tiny — you’ll always see both)
  • ✕Call a visibility gain "searched X% more" — we do not measure search volume and will not imply we do
  • ✕Fabricate or seed a baseline — a baseline is a real run that happened, and the first run says "baseline established," never a fake delta
  • ✕Replace a failed model call with generated filler — it reads "no live answer" and is excluded, never counted as 0 or faked
  • ✕Present a noisy delta as a clean win — small samples, wide bands, and backup-model substitutions are flagged on the number itself
  • ✕Hide the prompts — every score expands to the exact queries and raw answers so you can re-run them and check us
Query taxonomy

The buying questions we run.

CategoryAlternativeComparisonUse-caseProblemIntegrationCompliancePricingControl
Model coverage & repeated runs
Every query runs across GPT-5 mini, Claude Haiku 4.5, Gemini 2.5 Flash, and Perplexity Sonar. Paid tiers repeat runs to raise confidence and average out variance; each report shows the per-model score range.
Live-web grounding
When available we ground each answer in a current web search — the same up-to-date context a search-enabled assistant uses — and label every report live-web grounded vs. training-recall, so you know which you are reading.
Source-trust & corrections
We score which sources actually influence each answer (cited links + grounding results). Vendors can request a correction with evidence; rank still cannot be purchased.