Commerce metrics: answers, recommendation units, and shares
Use exact numerators and denominators for recommendation frequency, relevant rates, brand and retailer shares, top-three placement, and catalogue coverage.
On this page
Answer versus recommendation unitMetric definitionsRelevant does not mean only the positive answersA worked example with several denominatorsWhy a competitor can have several percentagesMissing values and small samplesEvidence strength is separate from the rateUse rates to choose an investigationAnswer versus recommendation unit
An answer is one valid AI response in the selected measurement basis. A recommendation unit is one product, brand, or retailer recommended in one answer. Several units can occur in the same answer. Counting units answers 'how were recommendations distributed?'; counting answers answers 'how often did this happen in responses?'. Neither is a share of customer purchases.
Metric definitions
The shared metric layer names the unit and denominator explicitly. Use the supporting counts below the percentage rather than inferring the base from the card title.
| Metric | Numerator | Denominator |
|---|---|---|
| Global recommendation frequency | Valid answers explicitly recommending the entity | All valid answers in the chosen measurement basis |
| Relevant recommendation rate | Answers explicitly recommending the product | Valid answers to questions in which that product was mentioned |
| Top-three rate | Answers explicitly recommending the entity at supported rank 1–3 | All valid answers in the chosen measurement basis |
| Brand share | Recommendation units belonging to the brand | All brand recommendation units in the period |
| Retailer share | Explicit retailer recommendation units for that shop | All retailer recommendation units; citations alone do not count |
| Catalogue coverage | Product recommendation units with confirmed catalogue matches | All product recommendation units, deduplicated by answer and product |
| Classification coverage | Clearly classified answers | All assessable answers; a quality measure rather than performance |
Relevant does not mean only the positive answers
A product's relevant-question population includes the valid answers to the selected questions where the product appeared, not just the answers that already recommended it. Other providers or repetitions of those questions can provide negative evidence. This prevents a relevant rate from becoming 100% merely because the system kept only positive responses. Some specialised views can expose a versioned, narrower scenario basis; read its own supporting counts before comparing it with the overview.
A worked example with several denominators
Suppose twenty valid answers contain ten recommendations of Product A. The selected relevant-question population contains twelve valid answers, and six of A's positive answers place it in the top three. The resulting percentages are all legitimate and describe different questions.
Synthetic example
Global frequency: 10 / 20 = 50%
Relevant rate: 10 / 12 = 83.3%
Top-three rate: 6 / 20 = 30%
If the same panel has 40 product recommendation units
and 14 have confirmed catalogue identities:
Catalogue coverage: 14 / 40 = 35%
Forty product units can come from twenty answers.Why a competitor can have several percentages
A retailer's share among retailer recommendations, its answer-level recommendation rate, and its literal mention rate are not interchangeable. AEO can additionally use another question portfolio or weighting scheme. Before declaring a disagreement, compare entity identity, time range, model/method, filters, unit, and eligible denominator. If the same metric and cohort disagree across pages, that is a consistency problem to report.
Missing values and small samples
A zero numerator over an eligible positive denominator is a recorded zero. A zero denominator produces an unavailable value, not 0%. A 100% result from one answer is one observation, not strong market evidence. Preserve the answer, question, provider, and run counts. Do not add independent category percentages whose populations overlap.
Evidence strength is separate from the rate
The reviewed evidence label limits repeat inflation using distinct question–provider pairs, capped by the available answer count. Moderate support requires at least five independent pairs and two questions. Stronger support requires at least twelve pairs, five questions, two providers, and either three runs or a qualifying repeat-consistency pattern. That repeat pattern requires at least three repeated cells with at least 80% consistency. These editorial evidence labels are not a probability that buying the product will succeed.
Use rates to choose an investigation
Start with the observed rate and exact base, then inspect source evidence and catalogue identity. Compare changes in percentage points only for compatible panels. A large share can reflect a narrow panel, and a low rate can reflect a broad panel. Neither automatically establishes commercial potential or the correct next purchase for your business.
Still need help?
Tell us what you expected and what happened.