Commercially focused question sampling
Queries are scored for commercial intent, relevance, strategic value, funnel stage, differentiation potential, testability, and likely business value.
Methodology
A transparent benchmark, not a universal ranking score.
AI answers are probabilistic. The goal is not to pretend otherwise; it is to measure a defined market question set consistently enough to make better decisions.
Controlled provider conditions
Every question receives a controlled first pass. Paid audits repeat only the five highest-value buying questions and provider disagreements, confidence risks, accuracy risks, or citation conflicts, capped at ten repeated questions.
Evidence before inference
Mentions, positions, citations, page evidence, and competitor sources are shown separately from the system’s interpretation of the gap.
Confidence governs priority
Provider agreement, adaptive repetitions, entity matching, extraction quality, and result consistency determine confidence. Low-confidence findings cannot become high-priority actions automatically.
Read the worked metric calculations to see how excluded answers affect a denominator. Then start with a free five-question baseline, or inspect the human-reviewed audit scope for a wider diagnosis.
Three named metrics, each defined once for every source.
RecoProof reports Shortlist Share, Recommendation Gap and Citation Share, supported by 6 secondary measures. Each is defined once on this page. The AI Recommendation Index, the paid audit and the free check each measure their own variant, and the variants differ in engines, questions, parser and sample size, so a number means something only with its label.
AI Recommendation Index
Public and independent: the same unbranded buyer questions per category, asked of several AI engines, with 95% intervals, quality gates and open data. The only source that ranks brands.
Index protocolPaid audit
One brand's own commercial question set across three providers, human-reviewed. Its numbers describe that brand's questions; they never rank brands, because every audit asks different questions.
Audit scopeFree check
The free snapshot RecoProof prepares for a prospect: five approved unbranded questions, one run each on one engine. Results are counts out of five, never percentages: a directional first look, not a rate. The public self-check reports the same way, but you read and tally the answers yourself.
Run the self-checkLabelling rule. Every public number carries the metric name, its source, the number of engines, n and the measurement date, for example Recommendation Gap (audit, 3 engines, n=74, measured 2026-09-10). A number without that label is not a RecoProof metric.
Sources are not interchangeable. Audit numbers never rank brands, because every audit asks its own questions. The index parser and the audit extractor are separate, so an index number and an audit number for the same brand are not directly comparable. Free-check results are shown as counts out of five.
Named metric
Shortlist Share
How often AI puts a brand on the buyer's shortlist.
The share of eligible answers to unbranded buyer questions that recommend the brand, meaning the answer presents it as a plausible choice for the buyer's stated need. Naming the brand is not enough; it has to be recommended.
Shortlist Share = eligible answers that recommend the brand ÷ eligible answers × 100
AI Recommendation Index
- Formula
- Per engine: answers recommending the brand ÷ the brand's eligible answers on that engine × 100. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers.
- Denominator
- The brand's eligible answers on one engine: completed, not empty, finish status STOP, parser confidence of at least 0.68, and offering options in the category. Labels still unresolved after human review leave the denominator; an ineligible answer never counts as "not recommended".
- Minimum sample
- At least 36 eligible answers on an engine for that engine's score, and at least 2 such engines for a combined score. Below that the brand is shown as "insufficient data" and is not ranked.
- How it is shown
- Whole-number percentage with its 95% question-bootstrap interval, n and engines. The headline metric: ranks and badges use it, badges on the interval's lower bound.
- Caveats
- Answers come from API requests, not consumer chat apps, so personalization and account history are not reflected.
- A fixed question set per category cannot represent every buyer question, which is why intervals are wide.
Paid audit
- Formula
- Per provider: unbranded evaluable runs recommending the brand ÷ unbranded evaluable runs × 100. Audit value: the equal-weight mean of the providers with at least 10 unbranded evaluable runs.
- Denominator
- Evaluable runs on one provider of the audit's approved questions that do not name the brand: evaluation status eligible (older runs without a status still qualify), with extraction confidence and brand-entity confidence each at least 0.68. Adaptive repeats count as runs.
- Minimum sample
- At least 10 unbranded evaluable runs on a provider for that provider's rate. A provider below that is shown but left out of the mean; with no qualifying provider the metric is not measurable and is not shown.
- How it is shown
- Whole-number percentage labelled with source, engines, n and measurement date. Never used to rank brands: every audit has its own question set.
- Caveats
- The paid report's Unbranded Recommendation Rate pools runs across providers; this public variant weights providers equally, so the two can differ.
- Adaptive repeats add runs for the highest-value and most disputed questions, so those questions weigh more than they do in the index.
- The audit extractor is not the index parser, so the same answer can be labelled differently by the two.
Free check
- Formula
- Answers that recommend the brand, out of 5.
- Denominator
- The 5 completed answers one engine (OpenAI) returned to 5 approved unbranded questions, one run each.
- Minimum sample
- Always exactly 5 answers; a check without 5 completed answers produces no result.
- How it is shown
- Shown as a count, "x of 5". Never shown as a percentage, compared with another brand, or ranked.
- Caveats
- 5 answers from one engine are a directional sample, not a rate: one answer moves the result by a fifth.
- The free check reports this count under the finding Recommendation coverage.
What it does not mean
- That the product is better, or that it wins sales.
- Market share or buyer demand.
- A prediction or guarantee of future answers.
Named metric
Recommendation Gap
How often AI mentions a brand without recommending it.
The share of eligible answers that name the brand but do not recommend it: the engine knows the brand but does not choose it. Every mention is either a recommendation or a gap, so Recommendation Gap = Mention Rate − Shortlist Share. It does not require another brand to be recommended instead; Displacement Share measures that.
Recommendation Gap = eligible answers that mention the brand without recommending it ÷ eligible answers × 100 = Mention Rate − Shortlist Share
AI Recommendation Index
- Formula
- Per engine: answers that mention the brand with recommendation NO ÷ eligible answers × 100. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers. The published value is the published Mention Rate minus the published Shortlist Share, so the identity holds on rounded numbers.
- Denominator
- The brand's eligible answers on one engine: completed, not empty, finish status STOP, parser confidence of at least 0.68, and offering options in the category. Labels still unresolved after human review leave the denominator; an ineligible answer never counts as "not recommended".
- Minimum sample
- At least 36 eligible answers on an engine for that engine's score, and at least 2 such engines for a combined score. Below that the brand is shown as "insufficient data" and is not ranked.
- How it is shown
- Percentage points with a 95% interval, published beside Displacement Share.
- Caveats
- Answers come from API requests, not consumer chat apps, so personalization and account history are not reflected.
Paid audit
- Formula
- Per provider: unbranded evaluable runs that mention the brand without recommending it ÷ unbranded evaluable runs × 100. Audit value: the equal-weight mean of the providers with at least 10 unbranded evaluable runs.
- Denominator
- Evaluable runs on one provider of the audit's approved questions that do not name the brand: evaluation status eligible (older runs without a status still qualify), with extraction confidence and brand-entity confidence each at least 0.68. Adaptive repeats count as runs.
- Minimum sample
- At least 10 unbranded evaluable runs on a provider for that provider's rate. A provider below that is shown but left out of the mean; with no qualifying provider the metric is not measurable and is not shown.
- How it is shown
- Whole-number percentage labelled with source, engines, n and measurement date. Never used to rank brands: every audit has its own question set.
- Caveats
- On pooled runs it equals the paid report's Unbranded Discovery Visibility minus its Unbranded Recommendation Rate; this public variant weights providers equally.
- Adaptive repeats add runs for the highest-value and most disputed questions, so those questions weigh more than they do in the index.
- The audit extractor is not the index parser, so the same answer can be labelled differently by the two.
Free check
- Formula
- Answers that mention the brand without recommending it, out of 5.
- Denominator
- The 5 completed answers one engine (OpenAI) returned to 5 approved unbranded questions, one run each.
- Minimum sample
- Always exactly 5 answers; a check without 5 completed answers produces no result.
- How it is shown
- Shown as a count, "x of 5". Never shown as a percentage, compared with another brand, or ranked.
- Caveats
- 5 answers from one engine are a directional sample, not a rate: one answer moves the result by a fifth.
What it does not mean
- That a competitor was recommended instead; Displacement Share measures that.
- That the answer was negative about the brand.
- Why the brand was not recommended.
Named metric
Citation Share
How often a brand's own website is part of the evidence AI shows.
The share of eligible answers that cite at least one page on the brand's own website. It is counted per answer, not per URL, so an engine that lists many sources cannot dominate.
Citation Share = eligible answers citing at least one page on the brand's own site ÷ eligible answers × 100
AI Recommendation Index
- Formula
- Per engine: answers citing at least one page on the brand's own site ÷ eligible answers × 100. Where one domain hosts several products, a citation counts only when its URL path starts with that product's prefix. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers.
- Denominator
- The brand's eligible answers on one engine: completed, not empty, finish status STOP, parser confidence of at least 0.68, and offering options in the category. Labels still unresolved after human review leave the denominator; an ineligible answer never counts as "not recommended".
- Minimum sample
- At least 36 eligible answers on an engine for that engine's score, and at least 2 such engines for a combined score. Below that the brand is shown as "insufficient data" and is not ranked. The combined value also needs 2 engines whose citation metadata passed quality gate 5.
- How it is shown
- Whole-number percentage with a 95% interval. Withheld for an engine whose citation metadata fails quality gate 5.
- Caveats
- Answers come from API requests, not consumer chat apps, so personalization and account history are not reflected.
Paid audit
- Formula
- Per provider: unbranded evaluable runs whose displayed answer citations include the customer-owned domain ÷ unbranded evaluable runs × 100. Audit value: the equal-weight mean of the providers with at least 10 unbranded evaluable runs.
- Denominator
- Evaluable runs on one provider of the audit's approved questions that do not name the brand: evaluation status eligible (older runs without a status still qualify), with extraction confidence and brand-entity confidence each at least 0.68. Adaptive repeats count as runs.
- Minimum sample
- At least 10 unbranded evaluable runs on a provider for that provider's rate. A provider below that is shown but left out of the mean; with no qualifying provider the metric is not measurable and is not shown.
- How it is shown
- Whole-number percentage labelled with source, engines, n and measurement date. Never used to rank brands: every audit has its own question set.
- Caveats
- The paid report's Owned Citation Rate counts branded and unbranded runs together; this public variant counts unbranded runs only, so the two differ.
- Sources the engine retrieved but did not show in the answer are not citations.
- The audit extractor is not the index parser, so the same answer can be labelled differently by the two.
Free check
- Formula
- Answers that visibly cite the brand's own domain, out of 5.
- Denominator
- The 5 completed answers one engine (OpenAI) returned to 5 approved unbranded questions, one run each.
- Minimum sample
- Always exactly 5 answers; a check without 5 completed answers produces no result.
- How it is shown
- Shown as a count, "x of 5". Never shown as a percentage, compared with another brand, or ranked.
- Caveats
- 5 answers from one engine are a directional sample, not a rate: one answer moves the result by a fifth.
- Only citations displayed in the retained answer, with a validated URL on the brand's domain, count.
What it does not mean
- That the brand was recommended: a citation is not a recommendation.
- That the cited page is accurate or persuasive.
- Traffic or referrals from AI answers.
Mention Rate
How often AI names a brand at all.
The share of eligible answers that name the brand, whether or not they recommend it. A recommendation always counts as a mention, so Mention Rate = Shortlist Share + Recommendation Gap.
Mention Rate = eligible answers that mention the brand ÷ eligible answers × 100
AI Recommendation Index
- Formula
- Per engine: answers mentioning the brand ÷ eligible answers × 100. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers.
- Denominator
- The brand's eligible answers on one engine: completed, not empty, finish status STOP, parser confidence of at least 0.68, and offering options in the category. Labels still unresolved after human review leave the denominator; an ineligible answer never counts as "not recommended".
- Minimum sample
- At least 36 eligible answers on an engine for that engine's score, and at least 2 such engines for a combined score. Below that the brand is shown as "insufficient data" and is not ranked.
- How it is shown
- Whole-number percentage with a 95% interval, published so that Mention Rate = Shortlist Share + Recommendation Gap holds on the published numbers.
- Caveats
- Answers come from API requests, not consumer chat apps, so personalization and account history are not reflected.
Paid audit
- Formula
- Per provider: unbranded evaluable runs mentioning the brand ÷ unbranded evaluable runs × 100. Audit value: the equal-weight mean of the providers with at least 10 unbranded evaluable runs.
- Denominator
- Evaluable runs on one provider of the audit's approved questions that do not name the brand: evaluation status eligible (older runs without a status still qualify), with extraction confidence and brand-entity confidence each at least 0.68. Adaptive repeats count as runs.
- Minimum sample
- At least 10 unbranded evaluable runs on a provider for that provider's rate. A provider below that is shown but left out of the mean; with no qualifying provider the metric is not measurable and is not shown.
- How it is shown
- Whole-number percentage labelled with source, engines, n and measurement date. Never used to rank brands: every audit has its own question set.
- Caveats
- On pooled runs this is the paid report's Unbranded Discovery Visibility; this public variant weights providers equally.
- The audit extractor is not the index parser, so the same answer can be labelled differently by the two.
Free check
- Formula
- Answers that mention the brand, out of 5.
- Denominator
- The 5 completed answers one engine (OpenAI) returned to 5 approved unbranded questions, one run each.
- Minimum sample
- Always exactly 5 answers; a check without 5 completed answers produces no result.
- How it is shown
- Shown as a count, "x of 5". Never shown as a percentage, compared with another brand, or ranked.
- Caveats
- 5 answers from one engine are a directional sample, not a rate: one answer moves the result by a fifth.
What it does not mean
- Being on the shortlist.
- Positive sentiment or an accurate description.
Displacement Rate
How often a tracked competitor is recommended while the brand is not.
The share of eligible answers in which the brand is not recommended and at least one predefined relevant competitor is. It is target-relative: the brand's absence alone is not displacement, and neither is a competitor mention or citation.
Displacement Rate = eligible answers where the brand is not recommended and at least one tracked competitor is recommended ÷ eligible answers × 100
Paid audit
- Formula
- Evaluable runs where the brand is not recommended and at least one tracked competitor is recommended ÷ evaluable runs × 100.
- Denominator
- Every evaluable run of the audit's approved questions, branded and unbranded, whether or not the brand is mentioned.
- Minimum sample
- No public minimum: reports show the count of displacement cases beside the evaluable runs it was counted from.
- How it is shown
- A count of displacement cases with its denominator, inside audit reports. Not a public headline number.
- Caveats
- Which competitors count depends on the audit's predefined competitor set.
- The audit extractor is not the index parser, so the same answer can be labelled differently by the two.
What it does not mean
- That the competitor caused the brand's absence.
- A share of gap answers only; Displacement Share measures that.
- Market share lost to a competitor.
Displacement Share
Of the answers that mention a brand without recommending it, how often another tracked brand is recommended.
Among Recommendation Gap answers only, the share in which at least one other tracked brand is recommended. It explains the gap; it is not a share of all answers.
Displacement Share = gap answers in which at least one other tracked brand is recommended ÷ gap answers × 100
AI Recommendation Index
- Formula
- Per engine: gap answers in which another tracked brand is recommended ÷ gap answers × 100. Combined: the equal-weight mean of the sufficient engines that had at least one gap answer.
- Denominator
- The brand's gap answers on one engine: eligible answers that mention it with recommendation NO.
- Minimum sample
- Published only for brands that meet the Shortlist Share minimum-data rule, and empty when there are no gap answers.
- How it is shown
- Whole-number percentage printed beside Recommendation Gap, without its own interval.
- Caveats
- Only brands in the category's tracked universe count as competitors.
What it does not mean
- That the competitor caused the omission.
- A share of all answers; Displacement Rate measures that.
First-Choice Rate
How often a brand is the single explicit top pick.
The share of eligible answers that name the brand as the single explicit first choice: exactly one recommended brand receives explicit preference or ranking language. Appearing first in a list is not enough.
First-Choice Rate = eligible answers naming the brand as the single explicit first choice ÷ eligible answers × 100
AI Recommendation Index
- Formula
- Per engine: answers naming the brand as the single explicit first choice ÷ eligible answers × 100. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers.
- Denominator
- The brand's eligible answers on one engine: completed, not empty, finish status STOP, parser confidence of at least 0.68, and offering options in the category. Labels still unresolved after human review leave the denominator; an ineligible answer never counts as "not recommended".
- Minimum sample
- At least 36 eligible answers on an engine for that engine's score, and at least 2 such engines for a combined score. Below that the brand is shown as "insufficient data" and is not ranked.
- How it is shown
- Whole-number percentage. Secondary: it never leads a headline.
- Caveats
- Checked each wave by blind human validation, which needs first-choice agreement of at least 90%.
What it does not mean
- Being mentioned or listed first.
- A ranking of products.
Shortlist Stability
How often the verdict stays the same when the question is asked again.
The share of question and engine pairs whose recommend or do-not-recommend verdict for the brand is identical in every repetition. It is a quality signal shown next to scores, not a performance score.
Shortlist Stability = question × engine pairs with the same YES/NO verdict in every repetition ÷ pairs where every repetition is eligible and resolved × 100
AI Recommendation Index
- Formula
- Per engine over one wave, where every question runs 2 times on every engine. Combined: the equal-weight mean of the engines that meet the minimum-data rule (3 engines are measured), never one pool of all answers.
- Denominator
- Question × engine pairs in which every repetition is eligible and its label resolved.
- Minimum sample
- Computed for brands that meet the Shortlist Share minimum-data rule.
- How it is shown
- Whole-number percentage shown next to every score. Below 70% the score is marked "variable" and no badge is awarded.
- Caveats
- Repetitions are collected on different days, so stability mixes sampling noise with day-to-day change.
What it does not mean
- That the result will persist next month.
- The paid report's Within-model Stability, which also compares mention and position and covers only adaptively repeated questions.
AI Share of Voice
A brand's share of all mentions of it and its tracked competitors.
The brand's mentions as a share of its own mentions plus the mentions of its tracked competitors in the measured sample. It counts mentions, not answers, and it depends on the competitor set.
AI Share of Voice = brand mentions ÷ (brand mentions + tracked-competitor mentions) × 100
Paid audit
- Formula
- Brand mentions ÷ (brand mentions + tracked-competitor mentions) across every evaluable run × 100.
- Denominator
- Mentions, not answers: every evaluable run of the audit's approved questions, branded and unbranded, adds one for the brand if it is named and one for each tracked competitor named.
- Minimum sample
- Shown when at least one mention exists; the paid report omits it otherwise.
- How it is shown
- Percentage inside paid reports only, with the competitor set named. Not a public headline number.
- Caveats
- Branded questions name the brand, which raises its share of voice.
- Changing the competitor set changes the number.
What it does not mean
- Market share.
- A recommendation rate.
- Comparable across audits with different competitor sets.
Measure who enters the recommendation set when the target brand does not.
For one defined commercial buyer question, RecoProof classifies an observed competitor-displacement case only when the answer is eligible, the target brand is deterministically not recommended, and at least one predefined relevant competitor is recommended. Target absence by itself is not displacement. A competitor mention or citation by itself is not displacement either.
As a share of all eligible answers this is Displacement Rate. The index publishes Displacement Share instead: the same kind of case, counted only among answers that mention the brand without recommending it.
Mention
The brand name appears in the retained answer.
Recommendation
The answer presents the brand as a plausible choice for the buyer's stated need.
First choice
Exactly one recommended brand receives explicit preference or ranking language. Mention order alone does not qualify.
Citation
The answer displays a source associated with a claim. A citation is not automatically a recommendation.
Competitor displacement
The target brand is deterministically not recommended while at least one predefined relevant competitor is recommended in the same eligible answer.
The denominator is eligible, deterministically classified observations.
It is not every prompt ever submitted. A completed answer must be eligible for classification and the target recommendation state must resolve to yes or no. Failed, excluded, invalid, or review-required observations do not silently become wins or losses. Review-required evidence stays unclassified until review resolves the ambiguity.
The result is target-relative. If Vendor A is the focal brand and Vendors B or C are recommended while A is not, that is an observed displacement case for A. It does not mean B or C caused A to disappear. The metric records the returned recommendation set; it does not assign causal influence or revenue attribution.
Mention rate asks, “Are we present?” Displacement asks, “When we are absent from a commercially important recommendation, who enters the considered set instead?” See the broader distinction between AI visibility metrics and denominators, then apply the framework with the practical guide to running an AI visibility audit for your brand.
One screenshot shows what happened in one run—not a stable market position.
AI visibility observations can disagree because the measured conditions or the tool’s classification design differ. Repetition improves the basis for judging stability, but it does not remove variability or reveal a single “real ranking.” Compare the measurement design before deciding that one dashboard is correct and another is fraudulent.
Prompt design
Wording, constraints, buyer context, and requested output can change the recommendation set even when two prompts represent the same underlying intent.
Run-to-run sampling
The same recorded request can produce different wording or selections. Repetition can reveal disagreement; it does not eliminate it.
Provider and model
Different providers, models, model versions, and interfaces are separate measurement conditions rather than interchangeable replicas.
Output and reasoning settings
Every run records its output token cap and reasoning setting. Paid audits give OpenAI models low reasoning effort with a 2,000-token cap and other engines a 1,000-token cap; Gemini thinking is set to minimal; five-question snapshots use an 800-token cap. An answer that reaches its cap is excluded as incomplete, never scored, and a change to these settings makes two measurements non-comparable.
Retrieval and source context
Where search or retrieval is available, source availability and retrieval behavior can change the evidence surfaced with an answer.
Execution context
Time, language, country, location, session context, and configuration matter only where the measured surface exposes or uses them, so they must be recorded rather than assumed.
Tool methodology
Prompt sets, repetitions, entity normalization, recommendation rules, ambiguous-answer handling, and aggregation logic can make two dashboards disagree.
Paraphrases are diagnostic comparisons, not independent demand observations.
Two naturally phrased questions can represent the same buyer intent and still return different recommendation sets. That difference is useful evidence about prompt sensitivity. It does not prove which wording is the “real ranking,” and paraphrases should not be counted as separate search-demand events.
Repetition is targeted where disagreement or decision value is highest.
The paid audit gives every approved question a first pass across three providers. It then repeats the five highest-value commercial questions plus questions with provider disagreement, entity or extraction confidence risk, accuracy risk, or owned-citation conflict, capped at ten repeated questions. Retained answers support separate within-model, cross-model, recommendation-agreement, and citation-stability views.
RecoProof’s confidence label is a decision rubric based on sample size, entity matching, extraction quality, and provider agreement. It is not a validated statistical confidence interval. Low-confidence evidence remains low priority or review-required rather than being upgraded by presentation alone.
Before comparing tools, compare their measurement design.
- What exact commercial questions were tested?
- Which providers, models, interfaces, dates, countries, and languages were used?
- How many first-pass and repeated observations were collected?
- What exact evidence qualifies a brand as recommended or first choice?
- How are predefined competitors and entity aliases normalized?
- Are mentions, recommendations, first choice, and citations reported separately?
- How are failed, ambiguous, and review-required answers handled in denominators?
- Can I inspect the retained answer and displayed source evidence behind a finding?
- How are results aggregated, and what does the confidence label actually mean?
Once a reviewable baseline and response process exist, the same design questions determine whether recurring monitoring is worth paying for. If branded and unbranded prompts diverge, use the focused diagnosis for search-to-recommendation variability.
These model-level measurements need a separate protocol for Google Search. Learn how to measure Google AI search visibility, then compare Google AI Overview tracking tools by retained answers, citation evidence, and collection conditions. The broader AI visibility tools comparison explains which measurement model each purchase supports.
RecoProof publishes no composite visibility score.
A number that blends recommendations, citations and stability with chosen weights is only as meaningful as its weights, so every public RecoProof number is one of the metrics above, with its label. Paid reports include one clearly labelled convenience index so a customer can track its own audits over time. It is for internal tracking only: it is not a public metric, it is never used to compare or rank brands, and its weights are disclosed in the report itself.
See how the evidence becomes an AI visibility audit.
Review the commercial audit scope, compare the fictional sample output, or start with a bounded five-question check.