How We Measure AI Commercial Momentum: Methodology for AI Investor Signals

Research status: Exploratory methodology for longitudinal AI recommendation research. The method measures changes in AI recommendation visibility. It has not been validated as a predictor of revenue, earnings, analyst revisions, valuation, or stock returns.

Current methodology version: V0

Primary observation window for the initial panel: 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

LLM Authority Index measures AI Commercial Momentum as the change in a company’s recommendation coverage across matched commercial prompts and AI platforms over time. The V0 methodology compares the same normalized prompt on the same AI platform family between a base month and September 2026, excludes explicit extraction failures, collapses response-identical duplicate exports, separates recommendations from mentions and citations, and maps relevant brands or product entities to public-company parents.

The primary statistic is the change in recommendation coverage, expressed in percentage points. A company is classified as a positive AI divergence candidate only when its recommendation coverage increases by at least 5 percentage points, the exploratory 95% interval remains above zero, and at least four platform families improve. A negative AI divergence candidate requires the reverse pattern. Everything else is classified as mixed or neutral.

These labels describe unusual AI recommendation movement. They are not stock ratings, valuation conclusions, or forecasts of financial performance.

The purpose of the methodology is to freeze a transparent AI-side signal before downstream financial outcomes are known. That prospective design allows later work to test the AI Commercial Momentum Hypothesis without rewriting the original signal after observing revenue, analyst revisions, earnings, or stock returns.

For the first 25-company output produced by this methodology, see Initial Findings From 25 Public Companies. The ongoing panel will be maintained through the AI Investor Signal Tracker.

Methodology at a Glance

Questions This Section Answers

• What exactly is AI Commercial Momentum?
• What is recommendation coverage?
• Why is this methodology based on recommendations rather than citations alone?

What Is AI Commercial Momentum?

AI Commercial Momentum is the change over time in how frequently a company is recommended by AI systems for commercially relevant, unbranded questions.

The key idea is not simply whether a company appears in an AI response. It is whether the company enters the AI-generated consideration set as a valid recommendation.

A company can be:

• cited as a source without being recommended;
• mentioned without being recommended;
• recommended but ranked differently across platforms;
• present in a response with neutral or negative framing;
• absent entirely;
• or explicitly recorded as not mentioned in the underlying company array.

Those are different states and should not be collapsed into a single visibility number.

The V0 investor signal therefore uses company-level recommendation fields as its primary signal source. Citation data remains valuable for understanding source authority and retrieval behavior, but citation count is not the denominator for this methodology.

The broader conceptual distinction is developed in AI Recommendations vs. Mentions vs. Citations.

Recommendation coverage

For a company or mapped public parent:

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

If a company is eligible in 200 matched prompt-platform cells and receives a valid recommendation in 80 of those cells, its recommendation coverage is 40%.

If the same company was recommended in 60 of 200 eligible matched cells in the base month, its recommendation coverage change would be:

40% - 30% = +10 percentage points

The unit is percentage points, not percent growth.

Recommendation coverage is intentionally different from:

• citation frequency;
• mention share of voice;
• average rank;
• sentiment;
• market share;
• revenue share;
• conversion rate;
• stock return.

The methodology does not assume that movement in one of those variables implies movement in another.

Source Corpus and Provenance

The V0 investor prototype is derived from the preserved LLM Authority Index research corpus.

The source archive contains:

• 68,203 raw observations;
• 310,114 raw citation-array entries;
• 624,698 company-array entries, including explicit not-mentioned records;
• 1,278 observations explicitly flagged as extraction failures;
• 5,976 exact cross-vertical duplicate observations identified in QA;
• 31,120 citation events attached to those duplicate observations.

The investor signal does not use the 310,114 raw citation entries as its primary denominator. It uses the company arrays because the company arrays contain the states required to distinguish presence, recommendation, rank, sentiment, and explicit non-mention.

The current company-momentum summary contains 406 company/entity signals with enough overlap to calculate the V0 summary, while the public-company prototype maps 25 public parents.

These numbers describe the preserved research infrastructure and current V0 output. They should not be confused with a claim that all raw observations are equally eligible for every investor analysis.

Questions This Section Answers

• How are failed AI responses handled?
• How are duplicated prompt exports handled?
• Why do explicit not-mentioned rows matter?

Failure Handling: A Failed Extraction Is Not a Zero

The source QA archive identifies 1,278 explicit extraction failures.

Those observations are excluded from valid AI visibility denominators.

This is important because treating a failed extraction as a legitimate “not recommended” response would mechanically push recommendation coverage downward. A parser failure, malformed source response, or extraction failure does not tell us whether the underlying AI system recommended a company.

The methodological rule is therefore:

Unknown because extraction failed is not equivalent to observed and not recommended.

This distinction matters most in longitudinal work. If one month had a higher failure rate than another and failed records were treated as zeros, the apparent month-to-month change could reflect extraction quality rather than AI behavior.

Some original stage0 datasets had particularly severe August extraction failure rates. Explicit failures can be removed, but missing upstream requests and silent losses cannot always be reconstructed. That remains a limitation of the current archive.

Duplicate Handling: Preventing Overlapping Exports From Becoming Extra Votes

The preserved corpus contains prompt-bank overlap across some vertical datasets.

QA identified 5,976 exact cross-vertical duplicate observations. In the primary cleaned company panel, response-identical duplicates exported into multiple vertical datasets are collapsed so that the same underlying response state does not receive extra weight simply because it appeared in more than one source export.

Vertical memberships are retained as metadata.

This matters because a repeated response should not become more influential merely because the data pipeline stored it twice under different vertical contexts.

The workbook also retains a no-dedupe sensitivity result for each public parent. That allows us to ask a simple robustness question:

If we do not perform this deduplication step, does the company’s measured direction materially change?

In the initial 25-public-parent panel, the median difference between the primary clean signal and the no-dedupe sensitivity calculation was 0.0 percentage points. The largest observed public-parent difference was about 3.5 percentage points.

That does not prove the cleaning rule is perfect, but it shows that most public-parent directions were not created by the cross-vertical deduplication choice.

Why Explicit Not-Mentioned Company Rows Matter

The source company arrays include records where a tracked company was explicitly not mentioned.

Those rows are analytically important because they provide a denominator.

If we analyzed only responses in which a company appeared, we could describe how the company was framed when present, but we could not estimate how frequently it entered the recommendation set across the whole matched prompt universe.

Recommendation coverage requires both states:

• recommended;
• not recommended.

Presence coverage similarly requires both:

• present;
• not present.

This is one reason the investor prototype is based on the company arrays rather than citation events.

Matched-Panel Design

Questions This Section Answers

• How do we reduce prompt-mix bias between months?
• Why do we compare the same platform family over time?
• What happens when July data is not available?

The primary longitudinal comparison uses a matched prompt-platform panel.

For each company or mapped public parent, the V0 methodology compares the same normalized prompt on the same platform family in the base month and September 2026.

July 2026 is the preferred base month when available. In the current public-company panel, Travelers uses August because July is unavailable in its matched panel.

Why matching matters

Suppose a bank appears in 20% of responses in July and 40% in September.

That raw change would be difficult to interpret if July contained mostly generic banking prompts while September contained a large new block of prompts specifically favorable to that bank’s niche.

The matched-panel design asks a narrower question:

For prompts and platform families observed in both periods, did the company become more or less likely to receive a valid recommendation?

That does not eliminate every form of sampling bias, but it reduces a major source of false movement.

Prompt normalization

Prompt text is normalized so that comparable prompt identities can be matched across periods.

The methodology then compares matched cells rather than simply comparing all July rows with all September rows.

The exact prompt universes are not identical across all verticals and months, which is why the matched design is required in the first place.

Public-Parent Rollups and Entity Resolution

AI systems frequently refer to brands, subsidiaries, products, operating businesses, or entity variants rather than the exact legal name of a public parent.

The V0 public-company layer therefore uses a manually curated parent map.

Examples include:

• Axos Financial: Axos Bank and UFB Direct;
• Cigna Group: Cigna and Express Scripts;
• UnitedHealth Group: UnitedHealthcare and related UnitedHealthcare consumer entities;
• Goldman Sachs Group: Goldman Sachs and Marcus by Goldman Sachs;
• Citigroup: Citi and Citigroup entity variants;
• American Express: American Express and American Express Co. entity variants.

For a prompt-platform cell, the public parent is treated as recommended when any tracked mapped entity for that parent receives a valid recommendation.

This prevents multiple labels for the same listed parent from being counted as separate stocks.

The important limitation

Parent rollups do not make a product-level signal equivalent to whole-company demand.

A movement in:

• Marcus by Goldman Sachs;
• Coinbase Wallet;
• CVS Pharmacy;
• Labcorp OnDemand;
• UFB Direct;

may be economically important, but its relevance depends on how much that business contributes to the parent company’s revenue, growth, margin, strategic positioning, and customer acquisition.

The V0 methodology therefore keeps an Exposure type field and requires later financial validation to account for business-segment importance.

The Happen/LendingClub row requires an additional transition caveat. The tracked entity is the legacy LendingClub brand during a 2026 rebrand period, so declining legacy-brand recommendation visibility may partly measure brand migration or AI staleness rather than weakening commercial demand.

A governed entity master is a planned methodological improvement.

Primary Point Estimate

For each public parent, the main reported movement is:

Recommendation change = September recommendation coverage - base-month recommendation coverage

The result is reported in percentage points.

The point estimate is cell-weighted across the eligible matched prompt-platform cells.

This means the primary reported change reflects the matched observation structure rather than giving every prompt family an identical weight regardless of how many matched platform cells it contributes.

The V0 method separately reports:

• base recommendation coverage;
• September recommendation coverage;
• recommendation change in percentage points;
• base presence coverage;
• September presence coverage;
• presence change;
• platform-specific recommendation changes;
• average recommendation rank when available;
• sentiment fields for descriptive analysis.

Only recommendation movement is used in the current watch-category rule.

Why Presence and Recommendation Are Separate

A company can be mentioned without being recommended.

A company can also remain present in roughly the same share of AI answers while becoming less likely to be selected as a preferred option.

For that reason, V0 reports presence coverage separately from recommendation coverage.

Presence is useful diagnostic context. Recommendation is the primary investor-signal variable.

The methodology does not combine presence, recommendation, citations, sentiment, and rank into one opaque score.

That choice is deliberate. If future validation shows that a composite model predicts downstream outcomes better, the weights should be learned from out-of-sample evidence rather than assigned arbitrarily in advance.

Platform Breadth

The current panel evaluates six AI platform families:

• ChatGPT;
• Gemini;
• Google AI Mode;
• Google AI Overviews;
• Microsoft Copilot;
• Perplexity.

For each platform family, the company’s matched recommendation change is classified by sign:

• improving;
• worsening;
• stable.

A broad movement across several systems is treated as more noteworthy than a change isolated to one platform.

This is not because all platforms are assumed to have equal economic impact. V0 does not yet have a validated market-impact weight for each platform.

Instead, platform breadth is used as a robustness condition.

If a company gains sharply on one system but declines on four others, the methodology resists labeling that move as a broad positive candidate.

This connects the investor work to the broader LLM Authority Index Persistence-Portability Gap. A signal can persist through time but fail to travel across platforms, or it can appear across platforms without persisting through time.

The dedicated investor-focused treatment is Does Cross-Platform AI Visibility Matter?.

Exploratory 95% Interval

The V0 public-company table includes a 95% interval around the recommendation-change estimate.

The interval is intentionally more conservative than treating every matched platform cell as fully independent.

The process is:

1. calculate matched-cell recommendation changes;
2. average those changes within each normalized prompt;
3. compute the standard deviation of the prompt-level means;
4. divide by the square root of the number of prompt clusters to estimate the standard error;
5. apply a normal 1.96 multiplier around the primary point estimate.

In shorthand:

SE = SD(prompt-level mean changes) / sqrt(number of prompt clusters)

95% interval = primary point estimate ± 1.96 × SE

Important interpretation caveat

The primary point estimate is cell-weighted, while the uncertainty estimate is based on prompt-clustered means.

This is an exploratory measurement interval, not a causal confidence interval and not a fully specified cluster-robust regression result.

The purpose is to prevent a company from being labeled directional when the observed movement is too noisy relative to the number and variation of matched prompts.

Future research should compare this V0 interval with more formal clustered regression and bootstrap approaches.

Statistical Direction

The current public watchlist uses the exploratory interval to define simple statistical direction:

• Positive if the 95% interval lower bound is greater than zero;
• Negative if the 95% interval upper bound is less than zero;
• Inconclusive otherwise.

This directional flag is only one input to the watch-category rule.

A company can have a positive point estimate and still be inconclusive if the interval crosses zero.

Likewise, a company can have a large negative raw change and remain mixed or neutral if the directional interval or platform-breadth conditions are not met.

Watch-Category Rule

The V0 watch category is rule-based and predeclared.

Positive AI divergence candidate

A company must meet all three conditions:

1. recommendation change of at least +5 percentage points;
2. 95% interval lower bound greater than 0;
3. at least 4 platform families improving.

Negative AI divergence candidate

A company must meet all three conditions:

1. recommendation change of at least -5 percentage points;
2. 95% interval upper bound less than 0;
3. at least 4 platform families worsening.

Mixed / neutral

Everything else is classified as mixed or neutral.

This three-part rule is designed to require:

• meaningful magnitude;
• directional consistency relative to prompt-level variation;
• cross-platform breadth.

The rule is intentionally simple enough to audit.

It is not optimized against stock returns, revenue growth, or analyst revisions. That is important. If the category thresholds had been tuned to maximize later market performance, the resulting backtest would be much harder to interpret honestly.

The current labels should therefore be understood as AI-side candidate classifications frozen before downstream validation.

The next conceptual step is described in What Would Prove the AI Commercial Momentum Hypothesis Wrong?.

Confidence Labels

The methodology separates signal direction from confidence in the AI measurement.

A company can have a positive or negative candidate classification while still carrying only Medium confidence.

High confidence

V0 requires:

• at least 200 matched cells;
• all 6 platform families represented;
• a directional 95% interval;
• absolute no-dedupe sensitivity no greater than 2 percentage points;
• absolute capture-average sensitivity no greater than 2 percentage points.

Medium confidence

V0 requires:

• at least 100 matched cells;
• at least 5 platform families.

Exploratory

Rows not meeting the Medium threshold are labeled Exploratory.

In the initial 25-public-parent panel:

• 7 rows are High confidence;
• 16 are Medium;
• 2 are Exploratory.

These labels describe confidence in the measured AI movement under the V0 methodology. They do not describe confidence that the movement predicts company fundamentals or stock returns.

Sensitivity Analysis

Questions This Section Answers

• Could the results be caused by duplicate-cleaning choices?
• Could repeated AI captures change the measured direction?
• Why do we report sensitivity instead of hiding alternative calculations?

Two sensitivity calculations are retained for every public parent.

1. No-dedupe sensitivity

The primary panel collapses response-identical duplicate exports across overlapping vertical datasets.

The no-dedupe sensitivity calculation removes explicit failures but otherwise retains the raw source observations without that response-level deduplication.

The difference between the sensitivity result and the clean primary result is reported in percentage points.

In the initial public-parent panel, the median difference was 0.0 percentage points. Most rows were therefore directionally stable to this cleaning choice.

2. Capture-average sensitivity

Some company/month/prompt/platform cells contain more than one distinct retained response signature.

The capture-average sensitivity model first averages recommendation state across those distinct retained response variants within the cell, then compares periods.

This asks whether the primary result depends on a single consolidated response state when multiple legitimate captures exist.

In the initial panel, the median absolute difference from the primary calculation was essentially zero, with the largest public-parent differences concentrated in a small number of rows.

Why publish alternative calculations?

Alternative reasonable cleaning choices are part of the uncertainty of this kind of research.

A signal that disappears under a modest cleaning change should be treated more cautiously than one that survives multiple specifications.

That is why the High confidence rule explicitly incorporates both sensitivity checks.

Why V0 Does Not Use Sentiment in the Watch Category

Sentiment fields are preserved in the company-level data but are not used in the current positive/negative candidate rule.

The reason is methodological, not philosophical.

Sentiment has not yet been calibrated for investor use in this dataset.

A model might describe a company positively while recommending a competitor. Another response might recommend a company while discussing risks or tradeoffs. Compressing those patterns into a single sentiment contribution before validating the labels could create false precision.

V0 therefore keeps sentiment available for descriptive analysis while excluding it from the directional watch-category rule.

The same logic applies to recommendation rank. Rank is recorded, but it is not currently combined with recommendation coverage through an arbitrary weighting formula.

Why V0 Does Not Build a Single “AI Score”

A single score is tempting because it is easy to display.

It can also hide methodological assumptions.

For example, an opaque composite might assign:

• 40% to recommendation coverage;
• 20% to citation share;
• 15% to average rank;
• 15% to sentiment;
• 10% to platform breadth.

Without validated weights, those percentages would reflect researcher preference rather than demonstrated predictive value.

The V0 design instead exposes the components separately.

If later longitudinal testing shows that recommendation coverage, platform breadth, rank, sentiment, or another variable adds incremental predictive value, a formal model can learn those relationships from historical data.

That future model should then be tested out of sample.

This distinction is central to the planned longitudinal revenue-growth validation study and investor backtesting framework.

What the Signal Does Not Mean

The V0 methodology does not establish:

• causality;
• revenue impact;
• earnings impact;
• analyst revision direction;
• market share;
• purchase share;
• intrinsic value;
• expected stock return;
• whether a stock is undervalued or overvalued;
• whether an investor should buy, sell, or hold a security.

“Positive AI divergence candidate” means only that the company’s AI recommendation momentum is directionally unusual enough under the V0 rule to merit financial follow-up.

“Negative AI divergence candidate” means the same in the opposite direction.

This is why AI Search as Alternative Data describes the signal as a candidate alternative-data layer rather than a validated investment factor.

Known Limitations

The V0 methodology has several important limitations.

1. Short longitudinal history

Only about three months of core longitudinal data are currently available.

That is not enough to validate a stock-selection model or a durable financial leading indicator.

2. Prompt universes vary

Prompt universes are not identical across all verticals and months.

The matched prompt-platform design reduces this problem but does not eliminate every sampling difference.

3. Upstream missingness cannot always be reconstructed

Explicit extraction failures are excluded, but missing upstream requests and silent extraction losses cannot always be recovered.

4. Some prompt banks overlap across verticals

The primary panel collapses response-identical overlaps and publishes sensitivity results, but corpus construction still matters.

5. Entity resolution is imperfect

The 25 public-parent rollups are manually curated.

A governed entity master is needed as the public-company universe expands.

6. Product-level exposure is not whole-company exposure

Some measured entities represent a segment, product, subsidiary, or legacy brand rather than the full listed company.

Financial validation must account for economic materiality.

7. Platforms are not economically weighted

V0 treats platform breadth as a robustness feature, not as a traffic- or revenue-weighted market-impact model.

8. Recommendation visibility is not financial performance

The causal bridge from AI recommendation behavior to consumer action, company revenue, earnings, estimates, and valuation remains unproven.

What We Will Test Next

The methodology is designed so that the AI-side signal can be frozen before later outcomes are observed.

The planned validation ladder is:

Stage 1: digital and consumer behavior

Test whether AI recommendation momentum precedes changes in:

• branded search demand;
• direct and organic website traffic;
• app traffic or engagement where observable;
• retailer or marketplace interest;
• AI-referred traffic where measurable.

Stage 2: analyst expectations

Test whether AI momentum precedes:

• revenue-estimate revisions;
• EPS-estimate revisions;
• changes in expected growth rates.

Stage 3: reported company outcomes

Test against:

• revenue growth;
• revenue surprise;
• EPS surprise;
• segment-specific outcomes where the measured brand is only part of the parent company.

Stage 4: market outcomes

Only after the earlier links are studied should the research test:

• sector-relative stock returns;
• factor-adjusted excess returns;
• whether AI momentum adds information after controlling for ordinary financial variables.

The core question is not whether a clever retrospective story can be constructed. It is whether a pre-specified AI signal at time T contains incremental information about outcomes observed later.

The formal design for that work is described in How Investors Could Backtest AI Search Signals Against Revenue, Analyst Estimates and Stock Performance.

How the Methodology Should Evolve

V0 is intentionally simple and auditable.

As the monthly history grows, likely improvements include:

1. Governed entity resolution. Replace manual mappings with a maintained parent, subsidiary, product, and brand hierarchy.
2. Fixed prompt-set identifiers. Version prompt universes so additions, retirements, and revisions are explicit.
3. Formal clustered inference. Compare the exploratory prompt-clustered interval with cluster-robust regression and bootstrap methods.
4. Sector-neutral analysis. Separate company-specific movement from category-wide changes in AI behavior.
5. Persistence metrics. Distinguish one-month spikes from sustained multi-month momentum.
6. Platform concentration metrics. Measure how dependent a signal is on one platform family.
7. Economic exposure weights. Where possible, relate product or segment signals to the share of parent-company revenue or customer acquisition they represent.
8. Out-of-sample model selection. Learn predictive weights only from historical training periods and evaluate them on later unseen periods.
9. Benchmark models. Compare AI variables with simple financial, search-demand, traffic, and market baselines.
10. Versioned methodology. Preserve historical results under the method that produced them rather than silently rewriting old signals when the model changes.

Methodology Versioning and Historical Integrity

The AI Investor Signals project is designed as a dated longitudinal record.

That means methodology changes should be versioned, not retroactively hidden.

If V1.1 changes the entity map, prompt matching, inference method, or candidate thresholds, future publications should state that explicitly.

Historical V0 articles should remain available with their original definitions.

Where practical, later research can show both:

• the original historical V0 signal;
• a restated value under the newer methodology.

The distinction matters because the credibility of a predictive research program depends on showing what was actually known and measured at the time.

The AI Investor Signal Tracker will serve as the permanent index for that history.

Reproducibility Checklist for a Public-Company Signal

A company-level AI momentum claim should not be considered reproducible unless the researcher can specify:

• the company and mapped entity set;
• the observation months;
• the normalized prompt identities;
• the platform families;
• the eligible denominator;
• extraction-failure exclusions;
• duplicate-handling rule;
• repeated-capture handling;
• base and later recommendation coverage;
• recommendation change in percentage points;
• platform-specific changes;
• prompt-cluster count;
• exploratory interval;
• sensitivity calculations;
• watch-category rule;
• confidence rule;
• methodology version.

That level of transparency is more important at this stage than producing a polished single-number score.

Related LLM Authority Index Research

This methodology is the technical foundation for the AI Investor Signals series.

• Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis defines the hypothesis being tested.
• Can AI Recommendation Momentum Identify Emerging Company Trends? Initial Findings From 25 Public Companies applies the V0 rules to the initial public-company panel.
• AI Search as Alternative Data: Could AI Recommendations Become a Leading Indicator for Investors? explains why the measurement could matter if later outcomes validate it.
• What Would Prove the AI Commercial Momentum Hypothesis Wrong? defines the falsification standard.
• AI Investor Signal Tracker will preserve the longitudinal record.
• AI Recommendations vs. Mentions vs. Citations explains why V0 keeps AI visibility metrics separate.
• Can AI Search Visibility Predict Revenue Growth? describes the planned outcome validation.
• Does Cross-Platform AI Visibility Matter? focuses on platform portability and concentration.
• How Investors Could Backtest AI Search Signals defines the future financial testing framework.

Research Disclosure

LLM Authority Index is publishing this methodology before the AI Commercial Momentum Hypothesis has been validated against a sufficiently long history of downstream financial outcomes.

The purpose is to create a timestamped, auditable record of the signal definition before later revenue, analyst, earnings, and market outcomes are known.

The methodology may change as evidence accumulates. Material changes should be versioned and disclosed rather than silently applied to historical publications.

Nothing in this methodology constitutes investment advice, a security rating, a valuation conclusion, or a recommendation to buy, sell, or hold any security.