AI Recommendations vs. Mentions vs. Citations: Which AI Visibility Metrics Could Matter Most to Investors?

Research status: Exploratory longitudinal research. AI recommendation momentum has not been validated as a predictor of revenue, earnings, analyst revisions, valuation, or stock returns.

Initial observation window: July through September 2026

Initial public-company panel: 25 mapped public parents

Broader company/entity momentum table: 406 companies and entities

AI platform families: ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, and Perplexity

Answer Capsule

AI mentions, citations, recommendations, rank, and sentiment are not interchangeable metrics. They describe different parts of an AI answer and can move in different directions for the same company.

For investors, the distinction matters because a company can be visible without being selected. An AI system may mention a company in an answer, cite its website as evidence, rank it highly when it appears, or recommend it as a preferred option. Each event has a different possible relationship to commercial consideration.

LLM Authority Index currently treats recommendation coverage as the primary V0 AI Commercial Momentum signal because it most directly measures whether a company enters the AI-generated consideration set across matched, commercially relevant prompts. Mentions, citations, rank, sentiment, presence, platform breadth, and persistence remain separate measurements that may improve interpretation or later prove predictive in their own right.

The initial public-company panel already shows why this separation is necessary. UWM Holdings had presence coverage increase by 2.3 percentage points while recommendation coverage fell by 11.5 points. American Express had recommendation coverage fall by 8.4 points while its average rank among observed recommendations improved from 3.88 to 2.59. Webull's recommendation coverage increased by 4.4 points while overall presence increased by only 0.5 points.

A dashboard that compressed those measurements into one visibility score could hide economically relevant differences.

This article explains what each metric measures, which questions it can answer, and why the investor research program described in the AI Commercial Momentum Hypothesis keeps them separate. The exact V0 measurement rules are documented in How We Measure AI Commercial Momentum, and the current public-company observations are maintained in the AI Investor Signal Tracker.

Key Findings

• A mention is not a recommendation. A company can be named in an answer without being selected as a preferred option.
• A citation is not a recommendation. A source can support an AI answer even when the source's company is not recommended, or even mentioned, in the answer text.
• Presence and recommendation can diverge. UWM Holdings gained 2.3 percentage points of presence coverage while losing 11.5 points of recommendation coverage in the initial matched panel.
• Rank can improve while recommendation frequency declines. American Express moved from an average observed recommendation rank of 3.88 to 2.59 while recommendation coverage fell 8.4 percentage points.
• Recommendation frequency can change more than mention-level visibility. Webull's recommendation coverage rose 4.4 points while presence rose only 0.5 points.
• Sentiment is a separate layer. Favorable framing can matter, but V0 does not use sentiment to assign investor-signal classifications because sentiment calibration has not yet been validated for investment use.
• Cross-platform breadth and persistence are not visibility metrics by themselves. They describe whether a measured signal is broad and durable across systems and time.
• No current AI visibility metric has been validated by LLM Authority Index as a standalone predictor of revenue or stock returns. The forward validation program will test which metrics, if any, add useful information.

AI Visibility Metrics at a Glance

Questions This Section Answers

• What is the difference between an AI mention, citation, and recommendation?
• Which AI visibility metric is closest to commercial consideration?
• Why should investors avoid using one blended AI visibility score?

Mentions, Citations, and Recommendations Measure Different Events

The generative-search industry often uses words such as visibility, share of voice, citation, inclusion, recommendation, and presence in overlapping ways.

That creates a measurement problem.

If two vendors both report an "AI visibility score" but one primarily counts mentions while the other counts citations or weights recommendation position, the resulting numbers may describe different behaviors even when the labels look similar.

Current industry publications increasingly acknowledge this distinction. CitedIntel separates mentions, citations, and recommendations as different AI-search events. DerivateX distinguishes citation share from recommendation share and argues that recommendation share is more directly related to being placed on a buyer shortlist. LLM Pulse notes that vendors calculate AI share of voice differently, with some counting brand mentions, others counting cited domains, and others weighting position. Presenc AI similarly treats mention rate, citation rate, framing, sentiment, and related metrics as distinct measurements rather than one underlying fact.

Those frameworks are useful examples of how the vocabulary is developing, but none of them proves that one specific metric predicts public-company revenue or investment returns.

For investor research, the safer starting point is to define each event separately and test it separately.

What is an AI mention?

An AI mention occurs when a company or brand is named in the generated answer.

Mention coverage can answer questions such as:

• How often is the company entering the answer at all?
• Is AI recognition of the company broadening across commercial prompts?
• Is the brand becoming more or less associated with a particular use case, category, price point, or audience?

Mention coverage is useful because a company generally cannot be considered if it is completely absent from the answer.

But mention is a weak semantic threshold.

A company might be mentioned as:

• a recommended option;
• a competitor;
• an example of what not to choose;
• a company with a limitation;
• a legacy provider;
• a source of background information;
• or a brand that is relevant to the category but not selected.

A mention therefore answers "Was the company in the answer?"

It does not necessarily answer "Was the company chosen?"

What is an AI citation?

An AI citation is a source attribution, supporting link, or cited page used to support an AI-generated response.

Citation data is especially valuable for understanding the information environment behind AI answers.

It can help answer:

• Which publishers, domains, and pages are repeatedly used as evidence?
• Is a company's own website being used as a source?
• Are third-party publishers shaping the model's description of the company?
• Which information sources appear portable across multiple AI systems?

This is the purpose of the broader LLM Authority Index citation research and Citation Index.

But a citation does not automatically create a commercial recommendation.

An AI answer can cite a company's documentation while recommending a competitor. It can cite a third-party review while mentioning several companies. It can cite a source that contains evidence relevant to the answer without naming the source's corporate owner in the generated response.

Citation frequency therefore answers "What evidence is the AI system using?"

It does not necessarily answer "Which company is the AI system selecting?"

What is an AI recommendation?

A recommendation occurs when the AI response affirmatively identifies a company, product, or brand as a valid option for the user's commercial question under the project's parsing rules.

Recommendation coverage is therefore closer to a consideration-set measure than a simple visibility measure.

For the V0 investor methodology:

Recommendation coverage = valid recommendation cells / eligible matched prompt-platform cells

The key word is eligible.

The denominator includes explicit not-mentioned company records rather than considering only responses where the company already appeared. This allows the metric to measure how often the company enters the recommendation set across the matched prompt universe.

The current research focuses on unbranded, commercially relevant prompts because those questions are closer to competitive discovery than branded queries such as "What does Bank of America offer?"

A recommendation therefore answers a narrower question:

When users ask comparable commercial questions without naming the company in advance, how often does the AI system select this company as an option?

That is why recommendation coverage is currently the primary V0 AI Commercial Momentum metric.

It still does not prove revenue impact.

The prospective revenue-growth validation study will test whether changes in recommendation coverage actually precede downstream commercial outcomes.

The V0 Data Already Shows Why Metric Separation Matters

Questions This Section Answers

• Can AI presence rise while recommendation visibility falls?
• Can rank improve while recommendation coverage declines?
• Can recommendation momentum move without much change in presence?

Yes to all three.

The initial 25-public-parent panel contains several examples where different AI metrics tell materially different stories.

Example 1: UWM Holdings, presence up but recommendations down

UWM Holdings, mapped through United Wholesale Mortgage, had:

• Base recommendation coverage: 21.3%
• September recommendation coverage: 9.8%
• Recommendation change: -11.5 percentage points
• Base presence coverage: 42.0%
• September presence coverage: 44.3%
• Presence change: +2.3 percentage points

This is a direct example of why "AI visibility" can be ambiguous.

If an analyst looked only at presence, the company became slightly more visible. If the analyst looked at recommendation coverage, the company became materially less likely to be selected as a recommended option in the matched panel.

Those are not contradictory measurements. They measure different things.

The company-specific result will be documented in United Wholesale Mortgage AI Search Visibility.

Example 2: American Express, rank improved while coverage declined

American Express had:

• Recommendation coverage change: -8.4 percentage points
• Base average observed recommendation rank: 3.88
• September average observed recommendation rank: 2.59

Because lower rank values are better, the observed rank improved.

At first glance, that might sound positive.

But rank is conditional on the responses where a ranked recommendation exists. Recommendation coverage asks how often the company is recommended at all.

American Express therefore became less frequently recommended in the matched panel while appearing at a better average position within the subset of recommendations where rank was observed.

A single rank metric could miss the decline in recommendation frequency.

The company-specific analysis will be published in American Express AI Search Visibility.

Example 3: Webull, recommendation coverage moved more than presence

Webull had:

• Recommendation coverage change: +4.4 percentage points
• Presence change: +0.5 percentage points

This suggests that the company was not simply appearing much more often. Within the matched observation set, it became more likely to receive a recommendation relative to its broader presence movement.

Webull does not qualify as a V0 positive AI divergence candidate because its current interval crosses zero and the full candidate rule requires more than magnitude alone. But the example demonstrates why recommendation rate and presence rate should be analyzed separately.

The future company page is Webull AI Search Visibility.

Other divergence examples

The initial panel contains several additional cases:

These examples are one reason the public AI Investor Signal Tracker publishes recommendation movement separately from related fields.

Why Rank Is Useful but Dangerous in Isolation

Questions This Section Answers

• Is being ranked first by AI more important than simply being recommended?
• Why can average recommendation rank be misleading?
• How should investors interpret rank together with recommendation coverage?

Rank can contain useful information, but it is usually conditional on being included.

Suppose Company A is recommended in 80% of eligible prompts with an average rank of 3.0, while Company B is recommended in only 10% of eligible prompts but has an average rank of 1.5 when it appears.

Company B has the better conditional rank but dramatically lower recommendation coverage.

Without both metrics, an analyst might incorrectly conclude that Company B has stronger AI commercial positioning.

This creates a form of selection bias:

• recommendation coverage measures how often the company enters the selected set;
• rank measures where the company sits after it enters the selected set.

Both can matter.

But they answer different questions.

A stronger future model may test interactions such as:

recommendation coverage x top-three recommendation rate x platform breadth x persistence

That would be more informative than relying on average rank alone.

For now, V0 retains average rank as a diagnostic field but does not use it in the watch-category rules.

Why Sentiment and Framing Should Remain Separate

Sentiment describes how the AI system talks about a company when it appears.

That may matter commercially.

A company could be recommended frequently but with recurring qualifications such as:

• best for advanced users but difficult for beginners;
• inexpensive but limited in features;
• strong national provider but weak local service;
• highly rated but expensive;
• convenient but with customer-service complaints.

Those distinctions could influence user behavior even if recommendation frequency remains unchanged.

But sentiment measurement has its own methodological problems:

• tone can depend heavily on prompt wording;
• mixed positive and negative framing can be difficult to compress into one score;
• recommendation language can be positive even when the surrounding explanation contains serious caveats;
• different sectors require different interpretations of risk language;
• automated sentiment labels may not map cleanly to purchase intent.

The source company table retains sentiment and framing fields, but the V0 investor watch-category rule does not use sentiment because sentiment calibration has not yet been validated for investment use.

Future research can test whether sentiment adds incremental information after controlling for recommendation coverage and presence.

Citation Metrics May Be More Useful for Explaining the Signal Than Replacing It

Citation data is central to understanding why an AI system might be recommending or describing a company in a particular way.

For example, a recommendation decline may coincide with:

• loss of citations from authoritative third-party review pages;
• new competitor pages becoming frequently cited;
• inconsistent information on company-owned pages;
• stale third-party descriptions;
• platform-specific source preferences;
• changes in which publishers dominate the answer ecosystem.

In that sense, citations can function as an explanatory layer beneath recommendation outcomes.

A useful research stack may eventually look like:

Citation environment -> company presence and framing -> recommendation probability and rank -> consumer response -> commercial outcome

But the arrows in that chain must be tested.

The existing LLM Authority Index citation research is designed to measure source authority and citation behavior. The investor series asks a different question: whether changes in company recommendation behavior contain information about future commercial or financial outcomes.

Those projects complement each other precisely because they do not collapse their metrics.

Questions This Section Answers

• Which AI visibility metric should investors prioritize today?
• Should citations or mentions be ignored?
• Could the best investor signal eventually combine several AI metrics?

Which Metric Should Investors Prioritize Today?

For the current V0 research program, recommendation coverage is the primary investor-facing AI metric.

That choice does not mean recommendation coverage has been proven to predict revenue.

It means recommendation coverage is the cleanest current measure of the behavior the hypothesis is trying to test: whether a company is becoming more or less likely to be selected when AI systems answer unbranded commercial questions.

The priority order for current research is roughly:

1. Recommendation coverage change for the primary AI Commercial Momentum signal.
2. Platform breadth to distinguish broad movement from one-platform noise.
3. Persistence as additional monthly observations accumulate.
4. Presence / mention coverage to identify whether selection changes reflect broader inclusion or a change in recommendation conditional on presence.
5. Rank and top-three rate to measure placement among selected companies.
6. Sentiment and framing to understand qualitative treatment.
7. Citation architecture to explain which sources may be shaping the answers.

This is a research priority order, not a proven ranking of financial predictive power.

The eventual empirical result may be different.

It is entirely possible that:

• presence predicts branded search better than recommendation coverage;
• recommendation coverage predicts customer acquisition better than mentions;
• sentiment matters only in high-trust sectors;
• citation changes precede recommendation changes;
• cross-platform breadth is more informative than the size of a one-platform move;
• or no AI metric adds meaningful predictive value after traditional variables are included.

The purpose of the longitudinal program is to find out.

Mentions and citations should not be ignored

Recommendation coverage is the current primary signal, but mentions and citations remain valuable.

Presence can reveal category association and broad awareness.

Citations can reveal information-source dependence and authority architecture.

Rank can identify competitive placement.

Sentiment can identify qualitative strengths and weaknesses.

Platform breadth can measure portability.

Persistence can measure durability.

The mistake is not tracking several metrics.

The mistake is pretending they are the same metric.

The Problem With Composite AI Visibility Scores

Composite scores are convenient.

They can summarize a dashboard into one number.

But they create three major problems for investor research.

1. Weighting assumptions can hide the economics

Suppose a score weights:

• mentions at 30%;
• citations at 25%;
• recommendations at 25%;
• rank at 10%;
• sentiment at 10%.

Those weights imply an economic theory.

They assert, implicitly, that a citation contributes a certain fraction as much information as a recommendation and that sentiment contributes some fixed amount to the total.

Unless those weights have been validated against downstream outcomes, the composite score can create false precision.

2. Opposing movements can cancel out

UWM Holdings demonstrates the problem.

Presence increased while recommendation coverage declined sharply.

A composite score might partially cancel those movements and report only a modest change, concealing the fact that the company was becoming more visible but less frequently selected.

That divergence may be more interesting than the blended score.

3. A composite can become impossible to falsify

If the score is built from many ingredients and the weights can be adjusted after seeing financial outcomes, the researcher can unintentionally overfit the metric to history.

The falsification framework is designed to avoid that problem.

For the initial research, LLM Authority Index therefore publishes underlying measurements rather than asking investors to accept a single opaque AI score.

A future composite metric would require empirical justification, transparent weights, and out-of-sample testing.

Cross-Platform Agreement Is a Separate Dimension

A metric can be strong on one AI platform and weak on another.

That does not necessarily mean one measurement is wrong.

Different systems can:

• retrieve different source sets;
• use different ranking logic;
• apply different model instructions;
• update at different speeds;
• answer with different levels of list compression;
• and interpret commercial prompts differently.

The investor question is therefore not only what metric moved, but also where it moved.

A recommendation gain across five of six platform families is different from the same aggregate change generated by one platform while the other five remain flat.

This is why platform breadth is part of the V0 candidate rule and why the dedicated cross-platform AI visibility study will examine recommendation portability and platform concentration risk directly.

The concept also connects to the broader LLM Authority Index Persistence-Portability Gap: a company or source can persist over time without appearing consistently across platforms, or appear broadly across platforms without remaining stable through time.

Persistence Is Different From Magnitude

A large one-month move may be less informative than a smaller move that continues for several months.

The current V0 observation window is only about three months, so persistence remains underdeveloped as a signal dimension.

Over time, the research will be able to distinguish:

• one-month spikes;
• two-month continuation;
• three-month persistence;
• reversals;
• plateauing;
• gradual accumulation;
• and structural regime shifts.

The AI Investor Signal Tracker will preserve those histories instead of overwriting prior values.

That history may eventually show that persistence matters more than raw magnitude, or it may show that rapid changes are useful earlier indicators. The current data is insufficient to decide.

How We Will Test Which Metric Actually Matters

The central validation question is not which metric sounds most commercially intuitive.

It is which metric improves prediction or explanation of later outcomes.

The planned testing sequence includes:

Stage 1: AI metric to consumer and digital behavior

Compare prior AI metrics with later changes in:

• branded search demand;
• direct web traffic;
• app traffic or engagement where available;
• retailer or marketplace navigation;
• customer acquisition indicators where observable.

Stage 2: AI metric to company financial outcomes

Compare prior AI metrics with later:

• revenue growth;
• segment growth;
• revenue surprise;
• EPS surprise;
• company guidance changes.

Stage 3: AI metric to analyst expectations

Test whether the signal precedes:

• consensus revenue revisions;
• consensus EPS revisions;
• target or expectation changes where available.

Stage 4: AI metric to market outcomes

Only after the earlier stages show useful information should the research test:

• sector-relative returns;
• event-window reactions;
• information coefficients;
• top-versus-bottom signal spreads;
• or market-expectation divergence.

The full validation protocol will be developed in Can AI Search Visibility Predict Revenue Growth?.

The key comparison should not be only:

Does recommendation coverage correlate with revenue?

It should also be:

Does recommendation coverage add more forward information than presence, citations, rank, sentiment, or conventional digital variables?

That is how the research can determine whether the current choice of recommendation coverage as the primary metric is justified by evidence rather than intuition.

What This Does Not Mean

This article does not establish that recommendation coverage is financially predictive.

It does not establish that mentions are unimportant.

It does not establish that citations are weak.

It does not establish that rank has no value.

It does not establish that sentiment is irrelevant.

It does not establish that a company with increasing recommendation coverage will grow revenue, beat earnings, or produce positive stock returns.

It does not establish that a company with declining recommendation coverage will lose market share or underperform.

The article establishes a measurement principle:

AI visibility is multidimensional, and each dimension should be measured and validated separately before being used in an investment model.

That principle is foundational to What Is an AI Investor Signal? and to the broader AI Search as Alternative Data research program.

Methodology Notes for This Article

The empirical examples in this article use the current V0 public-company matched panel.

The relevant methodology rules are:

• the same normalized prompt and platform family are compared across periods;
• July is the preferred base month when sufficient matched observations exist;
• Travelers uses August because July is unavailable in its matched panel;
• explicit extraction failures are excluded rather than treated as zero visibility;
• response-identical cross-vertical duplicate exports are collapsed in the primary panel;
• relevant brand and entity variants are mapped to public parents;
• recommendation coverage and presence coverage use eligible matched prompt-platform cells as denominators;
• average rank is a conditional diagnostic based on ranked recommendation observations;
• sentiment and framing fields are retained but not used in V0 candidate classification;
• platform breadth is evaluated separately;
• cleaning sensitivity is reported separately.

The complete technical specification is published in How We Measure AI Commercial Momentum.

Limitations

Several limitations matter when comparing AI visibility metrics.

Short time series

The core longitudinal archive currently covers only about three months. That is not enough to establish which metric is most predictive over economic or market cycles.

Metric definitions are still evolving across the industry

There is no universal market standard for terms such as visibility, share of voice, recommendation share, citation share, or rank. Comparisons across vendors require reviewing the underlying formula rather than relying on the label.

Recommendation extraction is interpretive

Determining whether text constitutes a valid recommendation requires parsing rules. Those rules should remain stable and auditable.

Rank is conditional

Average rank can improve even as recommendation frequency falls because it is calculated on a selected subset of observations.

Sentiment calibration remains immature

Generic positive, neutral, and negative labels may not capture commercially important nuance.

Citation behavior differs across platforms

Some AI systems expose explicit source links more consistently than others, complicating cross-platform citation comparisons.

Parent-company exposure varies

A signal for Coinbase Wallet, CVS Pharmacy, Marcus by Goldman Sachs, Labcorp OnDemand, UFB Direct, or another product or division may describe only part of the listed parent's economics.

None of the metrics has yet cleared financial validation

The project is measuring the AI side first. Financial relevance remains the outcome to be tested.

External Measurement Context

Several current industry sources illustrate why metric definitions need to be explicit:

• CitedIntel: Mentions vs Citations vs Recommendations separates the three event types and argues against treating them as interchangeable.
• DerivateX: AI Recommendation Share vs Citation Share distinguishes source citation from vendor recommendation and frames recommendation share as closer to the buyer shortlist.
• LLM Pulse: How to Measure AI Share of Voice documents that AI share-of-voice formulas differ materially across tools.
• Presenc AI: AI Visibility Metrics Explained treats mention rate, citation rate, framing, sentiment, and related measurements as separate fields.
• M+C Saatchi Performance: Measuring LLM Visibility recommends separately measuring inclusion, citations, recommendations, and competitors across a fixed prompt set.

These are examples of industry measurement practice, not evidence that any one metric predicts investment performance.

What We Will Test Next

The next research stage is to test whether the metrics separate not only conceptually but economically.

Questions include:

1. Does recommendation coverage predict branded search better than presence coverage?
2. Does presence coverage explain awareness while recommendation coverage explains consideration?
3. Do citation changes precede recommendation changes?
4. Does top-three recommendation rate add information beyond recommendation coverage?
5. Does sentiment improve prediction after recommendation frequency is controlled for?
6. Does cross-platform breadth improve robustness?
7. Does multi-month persistence matter more than one-month magnitude?
8. Are the relationships different across banking, insurance, fintech, mortgage, healthcare, and other sectors?
9. Which AI metrics add information beyond traditional financial and digital baselines?
10. Do any of these metrics survive genuine out-of-sample testing?

The companion cross-platform study will focus on portability, while the revenue validation study will focus on downstream outcomes.

Related LLM Authority Index Research

• Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis
• AI Search as Alternative Data: Could AI Recommendations Become a Leading Indicator for Investors?
• How We Measure AI Commercial Momentum: Methodology for AI Investor Signals
• [AI Investor Signal Tracker: Public Company AI Recommendation Momentum](https://llmauthorityindex.com/resources/ai-investor-signals/ai-investor-signal-tracker-public-companies-ai-search-recommendation-momentum)
• What Is an AI Investor Signal? How AI Search Visibility Could Become Financial Alternative Data
• Can AI Search Visibility Predict Revenue Growth? How We Plan to Test the Relationship
• Does Cross-Platform AI Visibility Matter? Measuring Recommendation Portability and Platform Concentration Risk
• Persistence-Portability Gap: Why AI Citation Authority Persists Over Time but Fragments Across Platforms
• LLM Authority Index Citation Index

Research Disclosure

This article reports exploratory AI-search measurement research. It is not investment advice and does not recommend buying, selling, or holding any security. Company-level AI visibility and recommendation measurements may describe only part of a public company's business and have not been validated as predictors of revenue, earnings, valuation, or stock returns.