Methodology

1. What we measure, and what we do not

AnswerRank measures how brands appear in web-grounded AI answer surfaces that expose either a documented API or a licensed structured search-results feed. Those two surfaces are the entire scope of our data.

2. Collection modes

Every stored data point carries the mode it was collected in. The interface never aggregates across modes without labelling them, and a workspace running on demo data shows a demo badge on every metric surface.

3. Per-engine collection

These labels are rendered from a single map in the codebase, so the description on this page, the tooltip in the product and the value stored on the row are the same string.

• OpenAI

  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

  Reflects the web-search API surface. It is not a capture of a signed-in claude.ai conversation.

• Gemini

  Grounding sources are those returned in grounding metadata. Consumer Gemini app results may differ.

• Perplexity

  Reflects the API surface. Perplexity’s consumer product applies additional personalisation and model routing.

• Google AI Overviews

  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.

4. Sampling and confidence

AI answers are non-deterministic. The same question asked twice can produce different sources and a different recommendation, so a single execution is a sample rather than a ranking fact. Every metric is reported with the number of valid checks behind it.

Stability

Stability is appearances divided by valid checks over the trailing window of 10 checks, and is written the way it is measured: “mentioned in 8 of 10 recent valid checks”.

Confidence thresholds

Confidence rises with three independent things: sample size, engine diversity, and spread across scan runs. High confidence requires all three — one thin dimension caps the result.

5. Execution status and validity

Every check resolves to exactly one status. Only two of them enter a metric. The rest are excluded from both the numerator and the denominator, because failing to observe something is not evidence that it is absent.

• success
  Counts in metrics

  The engine answered and the answer was captured.

• no_answer
  Counts in metrics

  The engine answered and your brand genuinely was not in it. A real absence, not a failure.

• provider_error
  Excluded from metrics

  The provider returned an error. We observed nothing.

• timeout
  Excluded from metrics

  The request exceeded its time budget. We observed nothing.

• rate_limited
  Excluded from metrics

  The provider throttled the request. We observed nothing.

• unsupported
  Excluded from metrics

  The engine cannot serve this kind of check.

• not_configured
  Excluded from metrics

  No credentials are configured for this engine in this deployment, so it never ran.

6. The AnswerRank Score

A 0–100 composite over valid checks only. The default weights below are the exact values used by the scoring code, and they are unit-tested to sum to 100%.

Position is redistributed, never zeroed

Prose answers do not contain ranks. Scoring a missing rank as zero would invent a penalty that the data does not support, so when no objectively ordered list exists the Position weight is redistributed proportionally across the components we could actually measure. The same rule applies to any component with nothing to measure. If no component is applicable, the product reports “no data” rather than a score of zero.

Intent weighting

Not every question is worth the same. A loss on a buying question costs more than a loss on a definition, so presence is weighted by prompt intent before it is averaged.

7. Citations and source influence

Citation URLs are canonicalised before they are counted: host lowercased, www. stripped, tracking parameters removed, fragment dropped, trailing slash normalised. Without that step the same page counts several times and inflates its apparent influence.

Each source is classified as owned, competitor-owned, review platform, community, publication, third party, or unknown. A low-confidence classification resolves to unknown rather than a guess, because a wrong label is worse than an honest gap.

8. How we talk about cause

Our evidence supports association, not causation. We can show that a competitor who wins a question is also cited by three review platforms you are absent from. We cannot show that the citations caused the win. Our explanation templates are linted against this rule.

9. Experiments

An experiment compares the same prompt × engine × market set before and after a change. We report the delta together with its sample size and the evaluation window, and we refuse to declare a verdict below 20 valid checks on either side of the comparison.

External factors — competitor changes, model updates, index refreshes — are uncontrolled and we say so on the result. An experiment is evidence, not proof.

10. llms.txt

The site audit detects and reports llms.txt, and labels it as an emerging convention with unproven impact. We report its presence because customers ask. We do not market it as a ranking factor, because the evidence for that does not exist yet.