AI visibility, with the sample size attached.
The AnswerRank Score compresses five measured components into a single 0–100 figure — and then shows you every component, every weight and every check that produced it. A score you cannot take apart is a score you cannot argue with.
- Composite range
- 0–100, computed from valid checks only.
- Components
- Visibility, recommendation, position, citation, sentiment — each weighted and published.
- Never shown bare
- Sample size and confidence level travel with the number wherever it appears.
What the AnswerRank Score is made of.
These are the default weights the product ships with. They live in a single configuration module, are unit-tested to sum to exactly 1, and are published here and on the methodology page rather than kept behind a marketing curtain.
| Component | Weight | Definition |
|---|---|---|
| Visibility | 45% | Intent-weighted share of valid checks where your brand is meaningfully mentioned. |
| Recommendation | 20% | Share of valid checks where your brand is actively recommended, not merely referenced. |
| Position | 15% | Where you place when an answer contains an objectively ordered list. Prose answers have no rank and are excluded. |
| Citation | 15% | Presence of your own pages and relevant earned sources among the citations behind answers. |
| Sentiment | 5% | How favourably your brand is characterised where it appears. |
- Visibility45%
- Recommendation20%
- Position15%
- Citation15%
- Sentiment5%
Visibility dominates because presence is the precondition for everything else. Recommendation is weighted separately because being named in passing and being actively recommended are different commercial outcomes, and collapsing them would hide the gap that matters most.
Prose answers have no rank, so we do not invent one.
Position is scored only where the answer actually contained an objectively ordered list and your brand appeared inside it. Most AI answers are prose.
- Where no ordered list exists, the position component is marked inapplicable for that check.
- Its weight is redistributed proportionally across the components that could be measured, so the result stays on a 0–100 scale.
- Scoring a missing rank as zero would manufacture a penalty out of a formatting choice the engine made.
- Each component reports its own sample size, so you can see how much of the score rests on how much data.
- 01Discard invalid executionsOnly success and no_answer survive. Everything else never reaches the arithmetic.
- 02Score each component on its own sampleA component with no measurable observations is marked inapplicable rather than scored as zero.
- 03Redistribute inapplicable weightRemaining weights are scaled proportionally so they still sum to 1.
- 04Report with n and confidenceThe composite is published alongside the sample size and the confidence assessment that qualifies it.
A loss on “best tool to buy” is not the same as a loss on “what is”.
Every prompt is classified by buying intent, and that classification changes how much a win or a loss moves the score. Vanity queries cannot inflate a number that is meant to describe commercial reality.
| Intent | Weight | Looks like | Why it is weighted this way |
|---|---|---|---|
| Transactional | ×1.5 | “sign up for X”, “X pricing”, “buy X” | The buyer is at the point of decision. Absence here is the most expensive kind. |
| Comparison | ×1.3 | “Acme vs Northwind”, “alternatives to Initech” | The shortlist is being formed. Not appearing means not being considered. |
| Commercial | ×1.15 | “best tool for a mid-market team” | Evaluation intent with real purchase pressure behind it. |
| Informational | ×1 | “what is answer engine optimisation” | Useful for authority, but a loss here costs less than a loss at the decision point. |
The ordering — transactional above comparison above commercial above informational — is asserted in tests, so it cannot be silently inverted by a future edit.
AI answers are stochastic. One run is a sample, not a fact.
Re-asking the same question can produce a different answer, so a precise-looking number built on a handful of checks is a misleading number. Confidence rises with three independent things, and any one of them being thin caps the result.
| Rule | Threshold |
|---|---|
| Minimum valid checks before a figure is reported without a small-sample warning | 10 |
| Valid checks required for medium confidence | 30 |
| Valid checks required for high confidence | 80 |
| Distinct engines required for high confidence | 3 |
| Distinct scan runs required for high confidence | 3 |
High confidence requires all three at once: enough checks, enough engine diversity, and enough time spread. A single lucky snapshot on one provider never reaches it, however large the sample.
Stability is reported over the trailing 10 valid checks — “mentioned in 8/10 recent valid checks” — because a brand that appears intermittently is a different situation from one that appears reliably at the same average rate.
AnswerRank · Acme
n = 240 valid checks · 5 engines · 4 scan runs · high confidence
- Visibility
- 44%
- Recommendation
- 21%
- Position
- n/a
- Citation
- 33%
- Sentiment
- 62%
- Stability
- 8/10
Position reads n/a here because the sampled answers were prose. Its weight was redistributed across the components that could be measured.
Why provider failures are excluded from the denominator.
A check resolves to exactly one status. Only two of them are valid, and the distinction is the difference between a metric that describes the world and a metric that describes your infrastructure.
success · no_answer
The engine responded. If your brand was not in the answer, that is a genuine absence and it counts against you — no_answer is deliberately valid, because pretending an honest miss never happened would flatter the score.
provider_error · timeout · rate_limited · unsupported · not_configured
We failed to observe. That is not evidence of absence, so the check is dropped from both the numerator and the denominator rather than counted as a miss. The status counts stay visible, so you can always see what was excluded and why.
The rule exists for one concrete reason: a provider outage on a Tuesday must never render as a visibility collapse on your chart. An engine without credentials renders as unavailable and awaiting configuration — never as a connected engine reporting zero.
One number for the category, five stories underneath it.
Engines disagree. They ground on different indexes, cite differently, and refresh on different cycles — so a composite that hides the split hides the actionable part.
- OpenAI41%n = 58
- Perplexity36%n = 54
- Gemini28%n = 51
- Claude22%n = 49
- Google AI Overviews18%n = 28
What each engine figure actually reflects
- OpenAI web-grounded API
Reflects the web-grounded API surface. It is not a capture of a signed-in ChatGPT session, which is personalised by account history and settings we cannot observe.
- Claude web-search API
Reflects the web-search API surface. It is not a capture of a signed-in claude.ai conversation.
- Gemini + Google Search grounding
Grounding sources are those returned in grounding metadata. Consumer Gemini app results may differ.
- Perplexity web-grounded API
Reflects the API surface. Perplexity’s consumer product applies additional personalisation and model routing.
- Observed Google AI Overview SERP data
AI Overviews do not appear for every query, and presence varies by location, device and time. Absence in a sample is not proof of permanent absence.
What this figure does not claim
A visibility percentage describes the sampled prompt set on the surfaces listed above, over the period shown. It is not a census of every question ever asked, and it is not a capture of a signed-in consumer chat session.
Movement in the score is associated with the changes recorded around it. We phrase findings as “evidence suggests” and “likely contributing factor”, because the mechanism inside the model is not observable from where we stand.
Get your AnswerRank Score.
One free visibility check, reported with its sample size and confidence.
No signup required for your first visibility check.
Every figure AnswerRank reports carries its sample size and how it was collected.
Read the methodology