Methodology & measurement rules

A percentage is only information when its methodology is explicit and comparable. This page lays out our sampling, scoring formula, non-computable rules and known limits — not to look rigorous, but so you can verify it or reject it: if you think a rule is wrong, email us and say so.

Methodology version: vs2 (source constant VISIBILITY_SCORE_CALIBER). Reports across versions are not directly comparable.

1. Where the data comes from (sampling)

Granularity: brand × AI engine × date. Sampled once per day (cron: UTC 01:00 = 09:00 Beijing time).
Engine slots: ChatGPT, Perplexity, Gemini, Claude (internal keys chat_gpt / perplexity / gemini / claude). Unsupported slots error out — we never substitute ordinary search results for an AI model's answer.
Sourcing discipline: engine slots are fed only by compliant automated channels that permit resale.
Time lock: what an AI engine said today cannot be retrieved tomorrow. If you don't sample today, that day is permanently empty and cannot be backfilled.
Retention: raw payloads are deleted after 7 days; derived metrics are retained permanently.

2. How the score is computed (bp & rank weights)

Integer basis points (bp): a 0–100 score is stored internally as 0–10000 bp to avoid floating-point accumulation error; divided by 100 for display.
Rank weights: mention at rank ≤ 3 scores 1, rank ≤ 5 scores 0.8, anything else 0.5; not mentioned scores 0.
Visibility score: visibility_score_bp = weighted mention sum / valid observations × 10000.
Valid observation: only samples with answer_source !== 'none' — a prompt where the engine produced no AI overview is not a valid visibility measurement.
Mention rate: mention_rate_bp is a different methodology, still counted over all prompts. It must not be mixed with the visibility score.

3. Methodology version: vs2

vs1: denominator = all samples (including answer_source='none') → prompts with no AI overview counted as "brand invisible" zeros, systematically understating visibility.
vs2 (current): denominator = only samples that did produce an AI overview → no-overview prompts stay out of the denominator.
Changing methodology = changing the source constant VISIBILITY_SCORE_CALIBER + updating tests + bumping the version. A score without a version cannot be compared across periods or recomputed.

4. null and 0 are different things

null = not computable: valid observations scored_samples = 0, or no mentions anywhere in the comparison set. Displayed as N/A.
0 = a real finding: there were valid observations, and the brand genuinely was not mentioned once.
We never display 0 in place of null: rendering "not measured" as 0 turns an engine / prompt / market configuration problem into "the brand is completely invisible in AI".
An empty sampling day is not 0: a day with no sampling is null, never padded with 0 (0 would be misread as "we dropped").

5. The SoV (share of voice) denominator

The denominator includes the brand itself: share_of_voice_bp = own_mentions / total_mentions × 10000, where total_mentions counts mentions of all brands in the comparison set, including itself.
Excluding self from the denominator produces values like 200% — mathematically self-contradictory, yet routinely written into reports as "we're strong in this set".
When the denominator is 0 we return null (not computable) — we never substitute 1 for 0. That would hand out the most optimistic number exactly when restraint is needed most.
The breakdown is returned with the result: sov.own_mentions / sov.total_mentions / sov.brands / sov.comparable / sov.issues / sov.sample_note.

6. The comparison set must be declared explicitly

peerIds is required — there is no "default: everything in the database" fallback. A default would make a report's denominator drift with whatever happens to be in the database.
The comparison set includes the brand itself and must be given explicitly by the caller (market, prompt version, authorized brands).
Competitor citation gaps (citation_gap) are computed within the same comparison set — passing the whole database would mix "who cites other customers" into your gap list.

7. Incomparability is disclosed, not hidden

Comparison sets should only be compared within the same market + the same prompt version. Cross-market or cross-version sets get sov.comparable = false, with each reason listed in issues (e.g. cross-market mixing, cross-version mixing).
When comparable = false, that number must not be used as a client conclusion — internal reference only.
Multi-engine is by design (engine_divergence looks specifically at engine differences), but reports must disclose which engines were covered.

8. Uneven sample sizes must be visible

When brands have different sample counts we return sample_note and never pretend sample sizes are equal. Unevenness is not judged incomparable, but it must be visible — share of voice counts mentions, so the less-sampled side is systematically understated.

9. Trend and engine-divergence rules

Trend (trend_delta): bp difference between the recent window and the preceding one. If either window has no valid observations → null; we never fabricate a delta from null − number.
Engine divergence (engine_divergence): the spread is computed only when at least two engines have valid observations; with a single engine, spread_bp = null — with no second engine to differ from, a "spread" is meaningless.

10. Known limits (disclosed as-is)

History cannot be backfilled: unsampled dates are permanently missing; starting one day late means losing that day forever.
We measure; we don't promise lift: whether mention rates improve depends on the client's own content, backlinks and PR.
Generable ≠ deliverable: when valid samples in a window are insufficient, we return covered and empty day counts and the trend field is N/A rather than an invented number.
Sample data is not delivered: rows with is_sample count but are not delivered as conclusions; sample-watermarked reports must not be handed to clients.
Aggregation boundaries: public indices and benchmarks require k-anonymity (k≥5) and the confidentiality boundary (client brands never enter cross-client benchmarks); no single client's data appears outside public aggregates.
Invite-only: currently invite-only; self-serve signup and online payment are opening progressively.

11. Change discipline

Changing methodology = changing the source constant + updating tests + bumping the version, and updating this page. No silent methodology changes.
This page, /llms-full.txt and /docs must agree; where they conflict, the newer statement wins, and you can verify by emailing service@saler.ai.
Field names referenced on this page are pinned by tests (byte-compared against src/metrics/visibility.ts): renaming a field without updating the docs turns tests red.