AI Recommendation Rate: Formula and Measurement Guide

By maxaeo

AI recommendation rate measures how often an AI assistant actively recommends your brand—not merely whether it mentions you. It separates meaningful endorsements from neutral descriptions, citations, conditional suggestions, exclusions, and complete absence.

That distinction matters because all of these responses contain a brand signal, but only one puts the company into the buyer’s choice set:

• “I recommend maxaeo for this use case.”
• “maxaeo is an AI visibility platform.”
• “Consider maxaeo if it supports your required markets.”
• “maxaeo would not meet this requirement.”
• A maxaeo article appears as a source, but the answer recommends another provider.

AI recommendation rate is not yet governed by an industry or ISO standard. This guide therefore defines a transparent measurement method—the maxaeo EQNXA framework—that another analyst can reproduce from the same prompts and responses.

What is AI recommendation rate?

AI recommendation rate is the percentage of eligible AI responses that explicitly endorse a brand for the user’s stated need. The numerator includes direct recommendations, requested-shortlist inclusion, and unambiguous suitability statements. The denominator includes every usable response that had a fair opportunity to recommend the brand, including responses where it was absent.

The strict formula is:

AI recommendation rate = Explicit recommendations ÷ Eligible responses × 100

If a company receives 18 explicit recommendations across 100 eligible responses, its AI recommendation rate is 18%.

The observation unit should be:

Prompt × model or AI surface × persona/market × run × timestamp

That structure prevents results from different models, markets, or collection periods from being treated as interchangeable.

What does “eligible” mean?

An eligible response comes from a prompt where recommending a provider was a reasonable outcome. Examples include:

• “What are the best AI visibility platforms for a global SaaS company?”
• “Which tools can monitor brand recommendations across ChatGPT and Gemini?”
• “Recommend three platforms for measuring visibility in AI-generated answers.”
• “What should an enterprise team use to track AI search visibility?”

A general definition prompt such as “What is generative engine optimization?” is not eligible for the primary rate because it does not ask the model to choose or evaluate providers.

Why the denominator matters

The denominator must include eligible responses in which the brand was not mentioned. Otherwise, the calculation becomes:

Recommendations among brand mentions

That companion metric can be useful, but it overstates market visibility because it ignores every occasion when the model selected competitors or named no vendors at all.

How is recommendation rate different from mention rate?

Mention rate measures presence. Recommendation rate measures endorsement. A brand can be widely recognized yet rarely presented as an appropriate choice.

Model language	Mention	Strict recommendation	Interpretation
“My top choices are A, B, and maxaeo.”	Yes	Yes	Explicit shortlist inclusion
“maxaeo is a strong option for multi-model monitoring.”	Yes	Yes	Clear suitability statement
“Consider maxaeo if it supports your required markets.”	Yes	No	Qualified consideration
“maxaeo monitors how AI systems describe brands.”	Yes	No	Neutral description
“maxaeo is not suitable for this requirement.”	Yes	No	Explicit exclusion
A maxaeo URL appears only in the citations	No, under the strict policy	No	Source visibility without brand endorsement
The response recommends competitors and omits maxaeo	No	No	Eligible absence

Citation rate answers a different question: whether the AI answer uses or displays your content as evidence. A model can cite a company’s educational article while recommending a competing product.

Prominence is also separate. A brand recommended in first position creates a different impression from one listed eighth, but both qualify as endorsements. Record position with a separate depth-of-mention framework.

A useful measurement funnel is:

1. Presence: Was the brand named?
2. Consideration: Was it presented as potentially suitable?
3. Recommendation: Was it affirmatively endorsed?
4. Prominence: How prominently did it appear?
5. Evidence: Was the brand or its content cited?
6. Accuracy: Were the claims about the brand correct?

No single “AI visibility score” can answer all six questions without obscuring important differences.

What counts as an AI recommendation?

A response is an explicit recommendation when it affirmatively places the brand in the user’s choice set. Direct advice, inclusion in a requested “best” list, a “best for” designation, or a clear statement of suitability qualifies. Positive adjectives alone do not.

The maxaeo EQNXA framework assigns one of six states to each brand-level observation:

Code	Classification	Operational rule	Example
E	Explicit recommendation	The answer recommends the brand, includes it in a requested shortlist, or clearly says it fits the need	“For cross-model tracking, maxaeo is a strong option.”
Q	Qualified consideration	The brand may fit, but a meaningful condition, uncertainty, or verification step prevents a clear endorsement	“Consider maxaeo if its regional coverage matches your requirements.”
N	Neutral reference	The answer names or describes the brand without advising the user to consider it	“maxaeo is an AI search visibility platform.”
X	Exclusion	The answer advises against the brand or states that it does not satisfy the requirement	“maxaeo is not a fit for this use case.”
A	Absent	The response is eligible and usable, but the brand does not appear	A competitor shortlist that omits maxaeo
U	Unscorable	No usable answer exists because of a refusal, truncation, error, wrong locale, or collection failure	Empty response or system error

Rules for difficult cases

Use these rules consistently before reviewing performance totals:

• Judge the action, not the adjective. “Popular,” “well-known,” and “established” are descriptions. “I recommend” and “a good fit for your requirement” are endorsements.
• Interpret lists from the prompt’s context. Inclusion in a requested “best tools” shortlist is E. Appearance in a historical or comprehensive vendor directory is N.
• Treat meaningful conditions as Q. “Choose maxaeo if you need daily tracking” can be E when it identifies a best-fit use case. “Consider maxaeo if it supports your market” is Q because an unresolved condition remains.
• Use the final conclusion when language conflicts. If the response praises the product but ultimately advises against it, label X.
• Do not infer endorsement from a citation. A source-domain appearance belongs in the citation field unless the user-visible answer also recommends the brand.
• Resolve the entity before assigning sentiment. Product names, former names, abbreviations, and similarly named companies require an alias dictionary. If the model refers to the wrong company, flag an entity error rather than crediting a recommendation.
• Code each brand independently. A single response may explicitly recommend several companies. Each can receive E.

Keep rank, sentiment, citation, factual accuracy, and evidence strength in separate columns. The EQNXA label should answer only one question: Did the model endorse this brand for the stated need?

Which responses belong in the denominator?

Include every successfully collected response to an eligible decision prompt, whether the target brand is recommended, mentioned neutrally, excluded, or absent. Exclude only responses that cannot be scored or prompts that did not create a genuine recommendation opportunity.

For one target brand:

Eligible responses (n) = E + Q + N + X + A

Unscorable responses are excluded from (n), but their count and failure rate must be disclosed.

Observation	Primary denominator	Reason
Requested category shortlist	Include	A provider recommendation is expected
“Best tool for…” prompt	Include	The model can select suitable brands
Non-branded problem prompt that asks for solutions	Include	A vendor recommendation is possible
Competitor shortlist that omits the brand	Include as A	Absence is part of market visibility
Neutral brand reference in an eligible answer	Include as N	The response was usable but did not endorse
General educational definition	Exclude	No provider decision was requested
Prompt that names the target brand	Separate cohort	Measures branded reputation, not unprompted discovery
Empty, refused, corrupted, or truncated response	Exclude as U	No reliable classification is possible
Wrong language, market, or persona response	Exclude as U and rerun	Collection conditions were not met

Use non-branded prompts for the headline discovery metric. “Should I use maxaeo?” tests how the model evaluates an already-known company; it does not show whether the brand enters an unprompted choice set. Keep branded and non-branded results separate using a branded-versus-non-branded prompt design.

How do you calculate AI recommendation rate?

Use the strict binary rate as the headline KPI and report companion metrics beside it.

For (n) eligible observations:

• Strict recommendation rate: (E / n)
• Consideration rate: ((E + Q) / n)
• Mention rate: ((E + Q + N + X) / n)
• Qualified consideration rate: (Q / n)
• Exclusion rate: (X / n)
• Absence rate: (A / n)
• Recommendation conversion: (E / (E + Q + N + X))
• Collection failure rate: (U / (n + U))

Recommendation conversion means “endorsements among mentions.” It is diagnostically useful, but it should not replace the headline recommendation rate.

The two-lever decomposition

A particularly useful diagnostic identity is:

Recommendation rate = Mention rate × Recommendation conversion

This separates two different problems:

1. Candidate-set visibility: Does the brand appear at all?
2. Endorsement conversion: When it appears, does the model recommend it?

For example:

• Mention rate: 51.8%
• Recommendation conversion: 36.2%
• Recommendation rate: (51.8% × 36.2% ≈ 18.75%)

A low mention rate calls for stronger entity recognition, category association, and third-party corroboration. A low recommendation conversion means the model recognizes the company but lacks sufficient reasons to endorse it.

Recommendation rate versus competitive recommendation share

AI recommendation rate is not market share. Multiple brands can receive E in the same answer, so competitors’ rates do not have to sum to 100%.

If you need a normalized competitive metric, calculate:

Competitive recommendation share = Target brand’s E labels ÷ E labels across all tracked brands

Use the same prompt set, brand dictionary, and annotation policy for every competitor. Report recommendation rate beside the share because a brand can gain share while the entire category receives fewer recommendations.

Weighted recommendation rate

If some prompts are more commercially important, assign weights before collection:

Weighted recommendation rate = Σ((w_i × E_i)) ÷ Σ(w_i)

A procurement prompt may deserve more weight than an early educational prompt. The weight should reflect audience value or buying intent—not which prompts produced favorable answers.

Always publish both the weighted and unweighted results. Post-hoc weighting makes the metric easy to manipulate.

Worked example: 120 AI responses

In this synthetic example, 120 collected responses produce 112 eligible observations after eight failures. Twenty-one explicit endorsements result in an AI recommendation rate of 18.75%.

The dataset demonstrates the calculation; it is not a maxaeo customer result or an industry benchmark.

Model	Eligible	E	Q	N	X	A	Strict rate
ChatGPT	28	7	5	4	2	10	25.0%
Gemini	28	4	6	3	1	14	14.3%
Perplexity	28	6	3	5	2	12	21.4%
Claude	28	4	3	2	1	18	14.3%
Total	112	21	17	14	6	54	18.75%

The companion calculations are:

• Consideration rate: ((21 + 17) / 112 = 33.9%)
• Mention rate: ((21 + 17 + 14 + 6) / 112 = 51.8%)
• Exclusion rate: (6 / 112 = 5.4%)
• Absence rate: (54 / 112 = 48.2%)
• Recommendation conversion: (21 / 58 = 36.2%)
• Collection failure rate: (8 / 120 = 6.7%)

A 95% Wilson interval for 21 successes in 112 observations is approximately 12.6% to 27.0%. The interval makes clear that 18.75% is an estimate, not a permanent property of the brand.

If repeated runs share the same underlying prompts, the observations are correlated. For consequential decisions, calculate uncertainty by bootstrapping complete prompt clusters rather than treating every run as independent.

How weighting changes the answer

Suppose 16 of 50 high-intent observations receive E, compared with five of 62 problem-awareness observations. The unweighted result remains 18.75%.

If high-intent prompts were assigned three times the weight before collection:

((16 × 3 + 5) ÷ (50 × 3 + 62) = 25.0%)

Both results are mathematically valid:

• 18.75% describes the collected response set.
• 25.0% describes the predefined intent-weighted portfolio.

The report must state which question each number answers.

How should the prompt sample be designed?

A valid prompt portfolio represents real buyer decisions, not a collection of phrasings selected because they mention the brand. Define the audience, market, category, constraints, and decision stage before generating prompts.

A practical starter portfolio

The following 50-prompt structure is a defensible starting point, not a universal standard:

Prompt group	Unique prompts	Purpose
Category shortlists	15	Test whether the brand enters broad candidate sets
Job-to-be-done prompts	15	Measure recommendation for specific outcomes
Constraint prompts	10	Test requirements such as market, integration, scale, or budget
Alternative and comparison prompts	5	Measure competitive substitution
Objection and risk prompts	5	Surface reasons for qualified or exclusionary answers
Total	50

Running 50 prompts across four models twice creates 400 observations. Those observations are not fully independent because each prompt is repeated, so reports should show both total responses and unique prompt count.

Useful non-branded prompt templates include:

• “Recommend three tools for [job] used by [persona].”
• “What are the best [category] platforms for a company operating in [market]?”
• “Which providers meet [constraint 1] and [constraint 2]?”
• “What alternatives should a buyer evaluate if [current approach] is insufficient?”
• “Which solution would you choose for [use case], and why?”

Do not merely paraphrase one favorable prompt 50 times. Prompt diversity should change the underlying decision, not just the wording.

Separate prompt cohorts

Maintain at least three cohorts:

1. Non-branded discovery prompts: Can the brand enter an unprompted shortlist?
2. Branded reputation prompts: How does the model evaluate the company when named?
3. Stable control prompts: Do model-wide changes affect unrelated categories or requirements?

Keep a stable core cohort for trend reporting and a smaller exploratory cohort for new buyer language. Do not merge exploratory prompts into the historical baseline without versioning the change.

How do you collect reproducible AI responses?

Reproducible tracking requires frozen prompts, documented collection conditions, complete raw responses, and stable scoring rules. A percentage without response-level evidence cannot be audited.

Record these fields for every observation:

Field	Why it is required
Prompt ID and version	Detects changes to the test
Full prompt	Preserves the exact decision context
Model and visible version	Separates different AI systems and updates
AI surface	Distinguishes chat, search, overview, or API experiences
Market, locale, and language	Controls geographic and linguistic variation
Browsing or retrieval state	Retrieval can change sources and recommendations
Account and personalization state	Logged-in history may affect output
Timestamp and run number	Supports trend and volatility analysis
Full response	Provides auditable evidence
Citations and destination URLs	Separates evidence visibility from endorsement
EQNXA label	Stores the recommendation classification
Evidence excerpt	Shows the language supporting the label
Coder and codebook version	Makes annotation traceable
Entity-match confidence	Flags potential brand-name collisions

For manual tracking, these fields fit in a spreadsheet. An AI visibility tool can automate collection and pre-classification, but borderline E-versus-Q and N-versus-X cases still need documented quality control.

Model recommendations can move after a model or retrieval update even when the website has not changed. Preserve model-level baselines and use a model-update change log before attributing every movement to marketing activity.

How do you build a reliable annotation workflow?

A reliable annotation workflow uses a versioned codebook, hides aggregate performance during labeling, independently reviews a sample, and preserves every disagreement.

1. Freeze the codebook. Define E, Q, N, X, A, and U before reviewing the response set.

2. Create boundary examples. Include ranked lists, conditional language, conflicting conclusions, citation-only appearances, and entity collisions.

3. Annotate at brand level. One multi-brand response may create separate observations for several tracked companies.

4. Blind coders to totals. Reviewers should not see whether a borderline decision will improve the dashboard.

5. Double-code a quality-control sample. Begin with 20% when introducing a codebook or new market.

6. Adjudicate disagreements. Preserve both initial labels, the final label, the rule applied, and the reason for the decision.

7. Recalibrate recurring boundaries. If E-versus-Q disagreements recur, add examples and repeat the affected sample.

8. Compute from frozen data. Publish the codebook version, collection window, numerator, denominator, unique prompt count, and failures with the result.

How much reviewer agreement is enough?

Report both raw agreement and Cohen’s kappa. Raw agreement is intuitive, while kappa adjusts for chance agreement. Kappa can be sensitive to label prevalence, so neither measure should be interpreted alone; this peer-reviewed overview of inter-rater reliability explains the limitation.

A practical internal publishing gate is:

• At least 90% raw agreement
• Cohen’s κ of 0.80 or higher
• No unresolved systematic disagreement over E versus Q

These are operating thresholds, not universal statistical laws. A lower score should trigger calibration, not selective removal of inconvenient responses.

How many responses are needed?

Sample size depends on the precision required, the expected rate, and the amount of clustering. There is no single minimum that fits every prompt portfolio.

For a simple independent proportion, an approximate starting formula is:

(n = z^2 × p(1-p) / m^2)

Where:

• (z = 1.96) for an approximate 95% confidence level
• (p) is the expected recommendation rate
• (m) is the desired margin of error

At a rate near 20%:

Independent observations	Approximate 95% margin
100	±7.8 percentage points
250	±5.0 percentage points
400	±3.9 percentage points

These are approximations. Repeated runs of the same prompt provide information about model volatility, but they do not add as much independent evidence as new prompt clusters.

For practical reporting:

• Show the total number of responses.
• Show the number of unique prompts.
• Show the number of models, markets, and runs.
• Use Wilson intervals for a simple binary rate.
• Use prompt-cluster bootstrap intervals when prompts are repeated.
• Avoid ranking small segments on percentage alone.

Two recommendations from five responses should never be presented as stronger evidence than 40 from 200 without displaying the sample sizes and uncertainty.

What is a good AI recommendation rate?

There is no credible universal benchmark. A good AI recommendation rate is one that improves against a frozen, relevant baseline without increasing exclusions, measurement failures, or dependence on branded prompts.

The result depends on:

• Category maturity and competitive density
• Prompt intent and buyer constraints
• Brand awareness and product maturity
• Market and language
• Model and AI surface
• Whether the prompt names the brand
• Denominator and annotation rules
• Collection date and retrieval conditions

A 15% rate across non-branded enterprise shortlists may be more valuable than a 60% rate on prompts that already contain the company name.

Use three comparisons:

1. Historical: Did the rate improve on the same stable prompt cohort?
2. Segment: Where is the company strong or absent?
3. Competitive: How does its rate compare with named competitors under the same method?

Treat public benchmarks cautiously unless they disclose prompts, model versions, markets, run counts, denominator rules, and labeling criteria.

Which segments should appear in a report?

Always show the overall rate with its numerator, denominator, and uncertainty, then segment by factors capable of changing the recommendation.

Segment	Question answered
Model or AI surface	Which systems recommend the brand?
Prompt intent	Does the brand appear during learning, comparison, or selection?
Persona	Is the product recommended to the intended buyer?
Market and language	Does visibility change by locale or availability?
Use case	Which jobs does the brand credibly own?
Constraint	Which requirements create qualified or exclusionary answers?
Funnel stage	Does recognition survive into final shortlisting?
Time window	Is a change persistent or short-lived?

For each segment, show:

• E count
• Eligible (n)
• Recommendation rate
• Q, X, and A rates
• Confidence interval
• Change from the stable baseline

Do not average model-level percentages unless their sample sizes and intended weights are equal. Aggregate the underlying counts or apply declared weights.

How should changes over time be interpreted?

A change in recommendation rate can result from market evidence, model updates, prompt changes, collection conditions, annotation changes, or genuine optimization. Attribution requires controlling these alternatives.

Use these controls:

• Maintain a stable core prompt set.
• Version every prompt and codebook change.
• Log model, surface, retrieval state, locale, and collection time.
• Mark product launches, content releases, PR activity, and entity corrections.
• Compare rolling windows instead of isolated daily results.
• Inspect response evidence behind every material movement.
• Use control prompts when evaluating an intervention.

If a lift appears only in newly introduced prompts, treat it as a hypothesis. If it persists across a stable cohort, multiple runs, relevant models, and unchanged annotation rules, it becomes stronger evidence.

Daily AI search monitoring is useful for detecting breakpoints. Weekly or 28-day windows are usually more appropriate for strategic interpretation because individual model outputs can vary between runs.

How do you improve AI recommendation rate?

Use the label distribution to identify the limiting stage. Absence, neutral recognition, qualified consideration, and exclusion require different interventions.

Pattern	Diagnosis	First action
High A	The brand is not entering the candidate set	Strengthen entity clarity, category association, third-party corroboration, and answer-ready coverage
High N	The model recognizes the brand but lacks reasons to select it	Publish verifiable differentiators, best-fit use cases, limitations, and comparison evidence
High Q	An unresolved condition blocks endorsement	Extract repeated caveats and publish evidence that resolves them
High X	The model sees the brand as unsuitable	Determine whether the reason is accurate, outdated, or attached to the wrong entity
High E, weak prominence	The brand is recommended but rarely leads the list	Strengthen evidence for the specific criteria buyers prioritize
High E, weak business response	AI endorsement is not translating downstream	Audit landing-page continuity, audience fit, offer clarity, and attribution

If the brand is absent

Improve the evidence that connects the entity to its category and use cases:

• Use one consistent organization and product name.
• State clearly what the product does, who it serves, and where it is available.
• Align Organization, Product, Person, and Article entities where relevant.
• Earn independent descriptions from credible industry sources.
• Publish pages answering the exact problems and constraints in the prompt set.
• Correct company-name collisions and obsolete profiles.

If an AI system confuses the company with a similarly named entity, use an entity-disambiguation playbook before producing more generic content.

If mentions are neutral

Neutral recognition means the model knows the company exists but lacks evidence for preference. Add:

• Specific best-fit and non-fit statements
• Verifiable feature and coverage details
• Transparent pricing or evaluation criteria where appropriate
• Original comparisons based on buyer requirements
• Named use cases with constraints
• First-party research or documented methodology
• Independent proof that supports important claims

Google’s people-first content guidance recommends original information, substantial value, clear sourcing, and content designed for an existing audience. Those principles also make passages easier for AI systems to interpret and cite.

If qualified consideration is high

Collect the exact caveat attached to every Q label. Group repeated conditions such as:

• “If it supports your target models”
• “If enterprise security is available”
• “If the product operates in your market”
• “If pricing fits a small team”
• “If the integration supports your stack”

A repeated condition is a content and evidence requirement—not merely a sentiment problem. Resolve it with a clear product page, documentation, independent validation, or an honest statement that the requirement is not supported.

If exclusions are high

Separate four causes:

1. Correct non-fit: The product genuinely does not meet the requirement.
2. Outdated information: The model describes an old feature set, market, or price.
3. Unsupported inference: The answer makes a claim without reliable evidence.
4. Entity collision: The answer attributes another company’s characteristics to the brand.

Do not try to erase legitimate limitations. Accurate non-fit language improves trust and prevents unqualified demand. Correct false or obsolete claims with specific, corroborated evidence.

How should leaders report the metric?

Report AI recommendation rate as evidence of AI-mediated consideration, not as proof of revenue or causal impact.

A defensible executive statement is:

“Across 200 eligible, non-branded category responses, the brand received 48 explicit recommendations: a 24% recommendation rate. That is six percentage points above the frozen baseline. The increase appeared across three models and 31 additional responses; sample sizes, confidence intervals, failures, and source evidence are attached.”

This is a reporting template, not a maxaeo benchmark.

A complete report should disclose:

• Collection window
• Prompt cohort and version
• Models and AI surfaces
• Markets and personas
• Runs per prompt
• E numerator and eligible denominator
• Unique prompt count
• Q, N, X, A, and U counts
• Weighted and unweighted rates
• Confidence interval
• Codebook version and reviewer agreement
• Representative response evidence

Connect the metric to commercial outcomes through AI referral sessions, self-reported attribution, assisted conversions, branded search changes, demo-request language, and win-loss interviews. Recommendation rate alone does not show that an endorsement caused a purchase.

For a broader executive dashboard, place it beside the other AI search visibility metrics rather than combining all signals into one opaque score.

Common measurement mistakes

The most serious errors inflate the numerator, shrink the denominator, or change the test between periods.

Avoid these mistakes:

• Counting every mention as an endorsement. Neutral references and exclusions are not recommendations.
• Using mentions as the denominator. This removes every eligible response where the brand was absent.
• Treating citations as advocacy. An answer may cite your article while recommending a competitor.
• Mixing branded and non-branded prompts. They measure different stages of visibility.
• Treating all lists as endorsements. A requested “best tools” list differs from a factual vendor directory.
• Ignoring multi-brand answers. Several brands can receive E in one response.
• Changing prompts after seeing weak results. The new prompt set no longer supports a historical comparison.
• Pooling incompatible markets. A provider may be available in one country and unsuitable in another.
• Hiding failures. Changes in error or refusal rates can alter the denominator.
• Applying weights after collection. Post-hoc weights invite a preferred result.
• Reporting percentages without counts. Small samples create false precision.
• Averaging percentages with unequal sample sizes. Aggregate counts or use predefined weights.
• Treating recommendation as conversion. AI output is an exposure signal, not revenue attribution.
• Ignoring inaccurate exclusions. The cause may be outdated information or an entity collision.
• Automating labels without audits. Small classification biases compound across thousands of responses.

Frequently asked questions

Is AI recommendation rate the same as mention rate?

No. Mention rate counts eligible responses that name the brand, regardless of context. AI recommendation rate counts only responses that explicitly endorse the brand for the user’s stated need.

In the worked example, mention rate is 51.8% while recommendation rate is 18.75%. The gap consists of qualified, neutral, and exclusionary mentions.

Should qualified consideration count as a recommendation?

Not in the strict headline metric. Keep Q separate so “consider this if…” cannot be silently upgraded into an unconditional endorsement.

Report ((E + Q) / n) as consideration rate. Then analyze the conditions attached to Q responses to identify the evidence, feature, market constraint, or trust signal blocking a clear recommendation.

What is a good AI recommendation rate?

There is no reliable universal benchmark. Compare results with a frozen internal baseline and competitors measured with the same prompts, models, markets, and annotation policy.

A lower rate on relevant non-branded buying prompts can be more meaningful than a high rate on prompts that already name the company.

How many responses do I need?

Choose the sample based on the precision required. At an observed rate near 20%, 100 independent observations have an approximate 95% margin of error of eight percentage points; 400 reduce it to about four points.

Because repeated prompt runs are correlated, also report unique prompt count and use prompt-cluster intervals for consequential decisions.

How often should AI recommendation rate be measured?

Use frequent collection for detection and longer windows for decisions. Daily monitoring can reveal sudden model or retrieval changes, while weekly and 28-day rolling results reduce the risk of reacting to individual outputs.

Keep a frozen baseline cohort and record model, prompt, and codebook changes.

Can AI recommendation rate be negative?

No. The strict rate ranges from 0% to 100% because it is the share of eligible responses labeled E.

Do not subtract exclusions from recommendations. Report exclusion rate separately so a shift from neutral recognition to active rejection remains visible.

Do recommendation rates across competitors add up to 100%?

Usually not. One AI response can recommend several brands, so each company may receive E.

Use competitive recommendation share when you need a normalized distribution, but retain absolute recommendation rate because share alone can conceal a category-wide decline.

Can an AI visibility tool calculate the rate automatically?

It can automate prompt collection, entity matching, initial classification, segmentation, and trend reporting. Human-reviewed boundary rules remain important for conditional language, mixed conclusions, similarly named entities, and unusual list formats.

Any tool should preserve the raw prompt, complete response, evidence excerpt, model metadata, label, and codebook version.

What should teams do next?

Start with non-branded decision prompts, preserve every raw response, classify each target-brand observation with EQNXA, and publish explicit recommendations over all eligible responses.

Then diagnose the result with:

• Mention rate
• Recommendation conversion
• Qualified consideration rate
• Exclusion rate
• Absence rate
• Citation rate
• Recommendation prominence
• Competitive recommendation share

maxaeo monitors how major AI assistants mention, rank, cite, and recommend brands over time. The measurement method in this guide turns those responses into an auditable KPI that SEO, marketing, PR, product, and finance teams can examine—and improve—without confusing visibility with endorsement.