Can AI Recommendation Momentum Identify Emerging Company Trends? Initial Findings From 25 Public Companies

Research status: Exploratory longitudinal research. These findings have not been validated as predictors of revenue, earnings, analyst revisions, valuation, or stock returns.

Primary observation window: July through September 2026

Public-company panel: 25 mapped public parents

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

Current V0 classifications: 2 positive AI divergence candidates, 11 negative AI divergence candidates, and 12 mixed or neutral observations

Answer Capsule

In the first LLM Authority Index public-company AI Investor Signals panel, AI recommendation momentum varied substantially across 25 public companies from the base month to September 2026. Under the initial V0 classification rules, Axos Financial and MetLife qualified as positive AI divergence candidates, while 11 companies qualified as negative AI divergence candidates and 12 remained mixed or neutral.

These are not stock ratings. They are frozen observations of how often a company or mapped public parent was recommended across comparable unbranded prompt and AI-platform cells. The central research question is whether sustained changes in this kind of recommendation coverage later correspond with changes in branded search, web traffic, revenue expectations, reported revenue, earnings, or market performance.

The result that matters most today is not that one company moved up and another moved down. It is that the signal can be measured prospectively, with fixed definitions, before later financial outcomes are known.

This article freezes that starting point.

For the theory behind the research, see Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis. For the full measurement rules, see How We Measure AI Commercial Momentum.

Key Findings

• 2 of 25 public parents qualified as positive AI divergence candidates under the V0 rules.
• 11 of 25 qualified as negative AI divergence candidates.
• 12 of 25 remained mixed or neutral.
• 7 companies received a High confidence classification, 16 Medium, and 2 Exploratory.
• Axos Financial recorded the largest positive recommendation-coverage change in the panel at +19.7 percentage points.
• MetLife increased +11.3 percentage points and improved on 5 of 6 platform families.
• United Wholesale Mortgage recorded the largest negative change at -11.5 percentage points.
• Goldman Sachs / Marcus declined -11.3 percentage points and worsened on all 6 platform families.
• Bank of America declined -8.6 percentage points and worsened on all 6 platform families.
• The panel's median recommendation-coverage change was approximately -6.4 percentage points, but this should not be interpreted as a market-wide decline because the 25-company panel is not a representative sample of all listed companies.
• The primary signal was generally stable under two sensitivity approaches, including a no-dedupe comparison and a repeated-capture averaging comparison.

Questions This Section Answers

• Which public companies gained the most AI recommendation visibility in the initial panel?
• Which public companies lost the most AI recommendation visibility?
• Does a positive or negative AI divergence classification mean a stock is undervalued or overvalued?

The strongest positive movements in this initial panel were Axos Financial and MetLife. The strongest negative movements included United Wholesale Mortgage, Goldman Sachs / Marcus, Labcorp OnDemand, Chime, CVS Pharmacy, Coinbase Wallet, Bank of America, and American Express, although not all of those rows qualified for the same confidence level or classification.

A positive or negative AI divergence classification does not mean that a stock is undervalued, overvalued, a buy, a sell, or likely to outperform. The classification describes AI recommendation movement only. The purpose of publishing it now is to create a timestamped record that can later be compared with real commercial and financial outcomes.

The Initial 25-Company Public Panel

The table below records the starting signal for all 25 public parents in the V0 panel. Publishing the full panel is intentional. Future validation should be able to compare later outcomes with the complete original set, not only with the companies that eventually appear to confirm the hypothesis.

Important: Travelers uses August as its base month because July was not available in the matched panel used for this analysis. All other rows shown here use July as the base month.

What the Two Positive AI Divergence Candidates Actually Show

Axos Financial: the largest positive movement in V0

Axos Financial increased from 15.3% to 35.0% recommendation coverage, a gain of 19.7 percentage points across 157 matched cells. Four of six platform families improved, one worsened, and one was stable.

The interval around the primary estimate was +9.8 to +29.7 percentage points, so the V0 signal met the positive directional requirement. Axos received a Medium confidence classification rather than High because one sensitivity treatment moved the estimate by more than the tighter High-confidence threshold.

This is exactly why the project separates signal strength from confidence. A large movement is not automatically a high-confidence movement.

The result also needs entity context. The public-parent rollup includes Axos Bank and UFB Direct. That makes the signal commercially relevant to the parent, but it does not mean every business line at Axos Financial experienced the same change.

MetLife: a smaller change with broader support

MetLife increased from 27.5% to 38.7% recommendation coverage, a gain of 11.3 percentage points across 222 matched cells.

Five of six AI platform families improved. Its interval was +4.4 to +18.1 percentage points, and both sensitivity checks were essentially unchanged from the primary estimate. Under the V0 framework, that combination qualified MetLife as a High-confidence positive AI divergence candidate.

Again, this is not evidence that MetLife's revenue or stock price will increase. It is evidence that, within this measured prompt population, MetLife became more likely to be recommended in September than in July.

The next question is whether that type of movement has economic information content.

The Negative AI Divergence Candidates

Eleven companies met the initial negative-candidate rule. The rule requires more than a negative point estimate. The recommendation-coverage change must be at least -5 percentage points, the upper bound of the exploratory 95% interval must remain below zero, and at least four platform families must be worsening.

That rule produced the following negative candidates:

• Ally Financial
• PennyMac Financial Services
• Upstart
• Happen / LendingClub
• American Express
• Bank of America
• Coinbase
• CVS Health
• Chime
• Goldman Sachs / Marcus
• United Wholesale Mortgage

Broad cross-platform declines are more interesting than isolated declines

Bank of America and Goldman Sachs / Marcus both worsened across all six platform families in the matched panel. Happen / LendingClub also worsened across all six.

That does not make those companies stronger investment signals automatically, but it makes their AI movement more portable across platforms than a decline driven primarily by one engine.

This distinction is central to the existing LLM Authority Index Persistence-Portability Gap. A pattern can persist over time without appearing consistently across platforms, and a broad cross-platform pattern can still fail to persist in later months.

The dedicated cross-platform AI visibility study will examine that issue directly.

Why 12 Companies Stayed Mixed or Neutral

A mixed or neutral label does not mean nothing happened.

Several companies moved materially but did not satisfy all V0 directional requirements.

For example:

• Principal Financial Group increased +5.4 percentage points, but its interval crossed zero and only three platform families improved.
• Webull increased +4.4 percentage points and improved on four platforms, but the point estimate did not reach the +5 point threshold and the interval crossed zero.
• Allstate declined -6.4 percentage points and worsened on five platform families, but its interval still crossed zero.
• Labcorp declined -10.0 percentage points, but the panel was smaller and the interval was wide enough to cross zero.
• Lincoln Financial had a negative overall point estimate while four of six platform families improved, illustrating why one aggregate number can hide important platform disagreement.

These mixed cases may eventually be more useful for learning than the cleanest candidates. They can show whether platform breadth, sample size, persistence, or magnitude is most important if any predictive relationship emerges.

Why This Research Could Matter to Investors

Questions This Section Answers

• Why might AI recommendation momentum contain economic information?
• What would make it useful as alternative data?
• Why is this different from simply measuring AI mentions or citations?

AI systems increasingly participate in product and service discovery. NielsenIQ reported on September 24, 2026 that 51% of U.S. consumers had used at least one AI-powered tool to support shopping in the prior month. NIQ has also argued that AI-mediated discovery is becoming part of the path to purchase, with recommendation visibility emerging as a new commercial measurement problem.

That does not establish a relationship between AI recommendation coverage and public-company revenue. It does establish a reason to test one.

A 2026 Marketing Science study of 973 e-commerce websites, representing more than $20 billion in revenue, analyzed more than 50,000 transactions attributed to ChatGPT referrals. The study found that organic LLM traffic remained a very small share of overall traffic, but that measurable commercial behavior already existed downstream of LLM referrals. It also found meaningful differences in conversion and revenue-per-session performance relative to traditional channels.

Our question sits one step earlier in the funnel.

We are not asking only whether a person who arrives from ChatGPT converts. We are asking whether the probability of being recommended before the click occurs contains information about future consumer consideration.

If recommendation momentum is eventually shown to precede branded search, traffic, revenue expectations, or revenue itself, the signal could become a form of alternative data. If it does not, the historical record should show that too.

For the broader framework, see AI Search as Alternative Data: Could AI Recommendations Become a Leading Indicator for Investors?.

Recommendation Coverage Is Not the Same as Mention Share, Citation Share, or Sentiment

One of the most important design rules in this series is metric separation.

A company can be:

• mentioned but not recommended,
• recommended without being ranked first,
• cited frequently while rarely being selected,
• described positively while losing recommendation coverage,
• visible on one AI platform and absent on another.

Those are different states.

The investor prototype therefore treats recommendation coverage as its primary commercial-consideration signal while keeping mentions, citations, rank, sentiment, and downstream financial outcomes separate.

The dedicated study AI Recommendations vs. Mentions vs. Citations: Which AI Visibility Metrics Could Matter Most to Investors? will test these distinctions more directly.

This separation also distinguishes the current work from some existing AI visibility research aimed at capital markets. For example, 5W's 2026 IPO AI Visibility Index evaluates recent and pending IPO candidates using a modeled AI visibility or readiness framework built around how companies appear in AI-generated answers. Q4 has separately launched AEO capabilities for investor-relations websites to improve how public companies' own materials surface in AI answers used by investors and analysts.

Those efforts demonstrate that AI visibility is already becoming relevant to investor communications. Our present study asks a different question: whether consumer-facing recommendation movement in unbranded commercial prompts can become a prospective commercial signal.

Why We Are Not Converting These Findings Into Valuation Yet

AIVO's September 2026 revision of its LLM Equity Valuation framework illustrates the risk of converting AI recommendation behavior into financial valuation before the assumptions have been validated.

AIVO withdrew earlier Grüns valuation figures after concluding that the original formula overstated AI-reachable revenue and compared incompatible financial quantities. Its revised framework now emphasizes unbranded recommendation behavior and explicitly calls for longitudinal testing of whether changes in AI recommendation share precede changes in AI-referred revenue.

That is an important methodological lesson.

Our V0 findings therefore stop at the signal layer.

We measure:

AI recommendation movement

We do not yet infer:

revenue impact

We do not yet infer:

earnings impact

We do not yet infer:

intrinsic value

We do not yet infer:

stock return

The later AI Visibility Market Divergence study, AI recommendation share versus market share study, and backtesting framework will only become meaningful after the time ordering is preserved and subsequent outcomes can be measured.

Sector Research Will Matter Because AI Exposure Is Not Economically Uniform

The initial public-company panel is concentrated in financial services, insurance, lending, healthcare, fintech, brokerage, and related consumer categories. It should not be treated as a random sample of the stock market.

That concentration is useful for one reason: the relationship between AI recommendations and business outcomes is unlikely to be identical across sectors.

A recommendation for a bank account, mortgage lender, pet insurer, brokerage platform, or consumer health service may sit relatively close to an eventual commercial action. A recommendation for a diversified public company may represent only one segment of the parent.

The initial sector studies will therefore examine the signal within more coherent competitive groups:

• Bank Stocks and AI Search: Which Banks Are Gaining or Losing AI Recommendation Visibility?
• Insurance Stocks and AI Search: Which Insurers Are Gaining or Losing AI Recommendation Visibility?
• Fintech, Brokerage and Crypto Stocks: What AI Recommendation Momentum Shows So Far
• Mortgage and Lending Stocks in AI Search: Initial Recommendation Momentum Signals
• Healthcare Stocks and AI Search: Which Companies Are Gaining or Losing AI Recommendation Visibility?

Sector analysis should also make it easier to compare companies against economically relevant peers rather than against the entire 25-company panel.

Platform Differences in the Initial Panel

Cross-platform breadth is one of the most important parts of the V0 design.

A large aggregate change concentrated on a single platform may represent platform-specific retrieval or ranking behavior. A similar change appearing across ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, and Perplexity is a different type of signal.

The strongest positive example of breadth in the initial panel is MetLife, which improved on five of six platform families.

Several negative candidates showed similarly broad declines:

• Bank of America: 0 improving, 6 worsening
• Goldman Sachs / Marcus: 0 improving, 6 worsening
• Happen / LendingClub: 0 improving, 6 worsening
• Ally Financial: 1 improving, 5 worsening
• American Express: 1 improving, 5 worsening
• Coinbase: 1 improving, 5 worsening
• Chime: 1 improving, 5 worsening

The platform-level pattern can also reveal disagreement. Lincoln Financial, for example, had a negative overall recommendation-coverage change while four platform families improved and two worsened. That kind of result cautions against interpreting the aggregate metric without inspecting the underlying platform distribution.

Persistence Is the Next Test

This article measures an initial change. It does not yet prove persistence.

A useful investor signal would likely need to survive more than one monthly comparison. If a company increases sharply in September and reverses completely in October, the original movement may have little commercial meaning.

The AI Investor Signal Tracker is intended to preserve that history month by month.

The central questions will include:

1. Does the direction persist?
2. Does the movement broaden or narrow across platforms?
3. Does it expand into additional prompt clusters or categories?
4. Does branded search or traffic move afterward?
5. Do analyst revenue estimates move afterward?
6. Does reported revenue eventually move in the same direction?
7. Does the relationship survive sector and market controls?

Only repeated measurement can answer those questions.

Brand-to-Parent Mapping Is a Major Limitation

Public-company mapping creates both usefulness and risk.

Investors think in terms of listed parent companies, but AI systems often recommend brands, products, divisions, or operating subsidiaries.

Several rows require special caution:

• Goldman Sachs includes Marcus by Goldman Sachs.
• Coinbase Global is represented by Coinbase Wallet in the measured consumer context.
• CVS Health is represented by CVS Pharmacy.
• Labcorp Holdings is represented by Labcorp OnDemand.
• UnitedHealth Group includes consumer-facing UnitedHealthcare entities.
• Axos Financial includes Axos Bank and UFB Direct.

A large change in one consumer-facing brand should not be interpreted as a proportionate change in the economics of the entire public parent.

Future versions should add segment weighting where reliable revenue or customer-acquisition exposure can be mapped to the measured entity.

The Happen / LendingClub row has an additional complication. The tracked entity is the legacy LendingClub brand during a 2026 rebrand. A decline in AI recommendation coverage may partly reflect brand migration or model staleness rather than weakening consumer demand. That row is intentionally preserved because rebrands themselves are an important test of how quickly AI systems update commercial entity knowledge.

Methodology Summary

The complete method is documented in How We Measure AI Commercial Momentum: Methodology for AI Investor Signals. The main V0 rules are summarized here.

Source corpus

The preserved LLM Authority Index research 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

Those are preservation counts, not the final screened public-company sample.

Primary outcome

Recommendation coverage is the proportion of eligible matched prompt/platform cells in which the company or mapped public parent received a valid recommendation.

Matched-panel design

For each company, the primary comparison uses the same normalized prompt and AI platform surface in the base month and September 2026. July is the preferred base month. Travelers uses August because July was unavailable in its matched panel.

This reduces prompt-mix bias between months.

Extraction failures

Explicit extraction failures are excluded rather than treated as true zero-visibility responses.

Duplicate handling

The primary panel collapses repeated response states inferred to represent duplicate exports. A no-dedupe sensitivity result is retained. A second sensitivity approach averages distinct retained response captures within comparable prompt/platform/month cells.

Public-parent rollup

Brand and entity variants are mapped to public parents. A parent counts as recommended in a matched cell when any included tracked entity is recommended.

V0 directional classification

Positive AI divergence candidate:

• recommendation change of at least +5 percentage points,
• lower bound of the exploratory 95% interval above zero,
• at least four platform families improving.

Negative AI divergence candidate:

• recommendation change of at least -5 percentage points,
• upper bound of the exploratory 95% interval below zero,
• at least four platform families worsening.

Everything else is classified as mixed or neutral.

Confidence classification

High confidence requires at least 200 matched cells, all six platform families, a directional interval, and limited movement under both sensitivity treatments.

Medium confidence requires at least 100 matched cells and at least five platforms.

Other rows are classified as Exploratory.

These are internal research-confidence labels for the AI measurement. They are not confidence levels for stock-price or financial predictions.

What Could Make These Initial Findings Disappear

There are many plausible reasons the current pattern may fail to become financially useful.

The relationship could disappear if:

• recommendation changes are mostly model noise,
• AI platforms refresh too irregularly for monthly momentum to be stable,
• the measured prompt set does not correspond closely enough to real purchase demand,
• consumers ignore AI recommendations,
• recommendations influence consideration but not enough transactions to affect public-company financials,
• company revenue is too diversified for the measured consumer segment to matter,
• existing market expectations already absorb the same information faster than the AI signal,
• traditional variables such as search demand or web traffic explain the apparent relationship entirely,
• the signal works only in selected sectors and fails elsewhere,
• the apparent effect disappears in out-of-sample testing.

We are documenting those failure paths in advance. See What Would Prove the AI Commercial Momentum Hypothesis Wrong?.

What We Will Test Next

The next validation layer is not another descriptive AI score. It is time-ordered comparison with downstream outcomes.

The research program will test whether frozen AI recommendation signals precede changes in:

Consumer and digital behavior

• branded search demand
• direct and organic website traffic
• app engagement where observable
• AI-referred traffic

Financial expectations

• analyst revenue estimate revisions
• analyst EPS estimate revisions
• consensus growth expectations

Reported company outcomes

• quarterly revenue growth
• revenue surprise
• earnings surprise

Market outcomes

• sector-relative stock returns
• valuation-multiple changes
• excess returns after basic risk controls

The dedicated study Can AI Search Visibility Predict Revenue Growth? How We Plan to Test the Relationship will define those tests before the outcomes are known.

What Investors Should and Should Not Take From the Initial Findings

Reasonable interpretation today

An investor, analyst, private-equity team, or corporate strategist can reasonably say:

LLM Authority Index observed measurable changes in AI recommendation coverage across a 25-public-company panel between the base month and September 2026. Some changes were broad across platforms and stable to sensitivity tests. The economic significance of those movements has not yet been validated.

Interpretation the data does not support today

The data does not support statements such as:

• Axos is undervalued because its AI recommendation coverage increased.
• MetLife will grow revenue because it improved across five platforms.
• Bank of America will lose revenue because its AI recommendation coverage declined.
• Goldman Sachs is overvalued because Marcus lost recommendation coverage.
• Coinbase stock will underperform because Coinbase Wallet declined in the measured prompt set.

Those would go beyond the evidence.

The research is designed specifically so that later articles can test whether any of those stronger relationships exist.

Why Publishing the Initial Results Before Validation Matters

Alternative-data research is vulnerable to hindsight.

After a company reports unexpectedly strong growth, it is easy to search historical data for a variable that happened to move first. After a stock falls, it is easy to identify an earlier decline that appears predictive.

That is not the standard we want for this project.

The stronger design is to publish the signal before the relevant future outcomes are known, preserve the full panel, preserve the methodology, and then evaluate what happens.

If the signal works, the publication history shows that the finding was prospective.

If it fails, the same publication history shows the failure.

If it works only in certain sectors, at certain horizons, or only when recommendation changes persist across several months, the archive should make that visible too.

This article is therefore less a conclusion than a timestamp.

Related AI Investor Signals 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
• What Would Prove the AI Commercial Momentum Hypothesis Wrong?
• AI Investor Signal Tracker: Public Company AI Recommendation Momentum
• What Is an AI Investor Signal? How AI Search Visibility Could Become Financial Alternative Data
• AI Recommendations vs. Mentions vs. Citations: Which AI Visibility Metrics Could Matter Most to Investors?
• Can AI Search Visibility Predict Revenue Growth? How We Plan to Test the Relationship
• AI Visibility Market Divergence: Can AI Recommendation Momentum Reveal Information Not Yet Reflected in Investor Expectations?
• AI Recommendation Share vs. Market Share: Could the Gap Reveal Emerging Commercial Strength or Weakness?
• Does Cross-Platform AI Visibility Matter? Measuring Recommendation Portability and Platform Concentration Risk
• How Investors Could Backtest AI Search Signals Against Revenue, Analyst Estimates and Stock Performance
• The Persistence-Portability Gap: Why AI Citation Authority Persists Over Time but Fragments Across Platforms

External Research and Prior Art

1. NielsenIQ. Majority of U.S. Consumers Now Use AI to Shop, NIQ Finds. September 24, 2026. https://investors.nielseniq.com/news/news-details/2026/Majority-of-U-S--Consumers-Now-Use-AI-to-Shop-NIQ-Finds/default.aspx
2. Kaiser, Maximilian and Christian Schulze. Frontiers: ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels? Marketing Science. Published online April 21, 2026. https://pubsonline.informs.org/doi/10.1287/mksc.2025.0489
3. AIVO Journal. Correcting LLM Equity Valuation: why we withdrew our Grüns figures and what replaces them. September 23, 2026. https://www.aivojournal.org/correcting-llm-equity-valuation-why-we-withdrew-our-gruns-figures-and-what-replaces-them/
4. 5W. AI Visibility Index / IPO AI Visibility Index. 2026. https://www.5wpr.com/ai-visibility-index/
5. Q4 Inc. Q4 Launches AEO for IR Web to Help Public Companies Stand Out in AI-Generated Answers. March 2026, updated July 2026. https://www.q4inc.com/resource-center/newsroom/q4-launches-aeo-for-ir-web-to-help-public-companies-stand-out-in-ai-generated-answers

Research Disclosure

This study is published by LLM Authority Index as exploratory research into AI recommendation behavior and possible future commercial applications. It is not investment advice, a securities recommendation, a valuation opinion, or a forecast of company revenue or stock performance.

Company inclusion reflects the available mapped public-company observations in the underlying research corpus and should not be interpreted as a representative sample of all listed companies.

The research methodology, classifications, and future revisions should remain versioned and historically accessible. Later evidence may strengthen, weaken, or overturn the interpretations presented here.