How Investors Could Backtest AI Search Signals Against Revenue, Analyst Estimates and Stock Performance

Research status: Prospective backtesting framework. AI recommendation momentum has not been validated as a predictor of revenue growth, analyst revisions, earnings, valuation, or stock returns.

Initial AI 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

Current methodology version: V0

Answer Capsule

A credible backtest of AI search signals should ask whether information measured at time T improves the prediction of outcomes that occur after time T, compared with a reasonable model that excludes the AI variables.

For LLM Authority Index, that means the backtest should not begin with stock returns. It should begin closer to the proposed commercial mechanism:

AI recommendation momentum -> branded search and digital engagement -> analyst revenue revisions -> reported revenue growth and revenue surprise -> earnings outcomes -> sector-relative stock performance

The current July-September 2026 dataset is not long enough to claim a valid stock-performance backtest. Its role is to establish a dated, frozen starting record before later outcomes are known. The initial public-company panel contains 25 mapped public parents, and the broader company/entity momentum table contains 406 signals. Those observations can become the first test cases as future months, financial reports, analyst revisions, and market outcomes accumulate.

The core backtesting rule is simple:

Every variable used to form the AI signal, baseline model, company ranking, or simulated portfolio must have been knowable on the historical signal date.

That requirement prevents look-ahead bias. The design must also address survivorship bias, multiple testing, backtest overfitting, changing prompt populations, entity mapping, missing AI observations, transaction costs, turnover, and the fact that public-company stock returns are much noisier than the commercial outcomes the hypothesis is designed to explain first.

The prospective revenue-validation design is published in Can AI Search Visibility Predict Revenue Growth?. The AI-side signal construction is documented in How We Measure AI Commercial Momentum. The failure conditions are precommitted in What Would Prove the AI Commercial Momentum Hypothesis Wrong?.

The Backtest Must Reconstruct the Historical Information Set

Questions This Section Answers

• What makes an AI search backtest credible?
• How do we prevent look-ahead bias?
• Why is a simple correlation with later stock returns not enough?

The most important requirement in a financial backtest is not model complexity. It is historical honesty.

A backtest should recreate the information set that would actually have been available when a signal was observed.

Suppose an AI recommendation signal is frozen on September 30, 2026. A valid test can use:

• AI recommendation observations collected through September 30;
• financial statements already released by September 30;
• analyst estimates published by September 30;
• market prices and returns observed by September 30;
• web traffic, search demand, app usage, or other alternative data available by September 30;
• sector classifications known at the time;
• public-company entity mappings documented by that date.

It cannot use:

• a quarterly revenue result released in November;
• an analyst revision published in October;
• a corrected ticker or parent-company mapping discovered after the outcome unless the correction is treated as a methodology revision rather than silently inserted into the historical signal;
• a prompt universe redesigned after researchers learn which companies later performed well;
• a company universe that excludes firms that later failed, delisted, merged, or disappeared simply because they are no longer convenient to analyze.

This is why the AI Investor Signal Tracker preserves dated historical signals rather than rewriting them when later evidence changes the interpretation.

CFA Institute's 2026 backtesting curriculum describes rolling-window, or walk-forward, backtesting as a way to approximate the real investment process and specifically warns about look-ahead and survivorship bias. Its active-equity curriculum also identifies overfitting, data mining, unrealistic turnover assumptions, and transaction costs as common pitfalls in quantitative investing. Those principles apply directly to any attempt to convert AI-search measurements into a financial signal.

See:

• CFA Institute, Backtesting & Simulation
• CFA Institute, Active Equity Investing: Strategies

A simple retrospective statement such as "companies with rising AI visibility later outperformed" would therefore be inadequate.

The required question is more demanding:

Could a researcher using only information available at each historical signal date have formed the same signal, applied the same rules, and obtained useful out-of-sample information?

Backtesting Architecture at a Glance

The horizons above are research targets, not universal rules. Reporting calendars do not line up neatly with fixed day counts, so financial outcome tests should also use event-based windows such as the next quarterly report after the signal date.

Start With the Baseline Model, Then Add AI

Questions This Section Answers

• How do we know whether AI adds information rather than restating known trends?
• What should the baseline model contain?
• What is the correct comparison between a traditional model and an AI-enhanced model?

The central backtest should compare two model families.

Model A: baseline information available without AI-search variables

The baseline should include variables that a reasonable analyst could already observe before the outcome date.

Depending on the target, these could include:

• lagged revenue growth;
• prior revenue surprise;
• prior EPS surprise;
• consensus revenue expectations;
• recent analyst estimate revisions;
• company size;
• valuation variables;
• sector and industry;
• recent stock-price momentum;
• seasonality;
• branded search trend;
• website or app traffic trend where legally and operationally available;
• prior market-share indicators;
• company fixed effects when the panel becomes long enough;
• time fixed effects or macro controls.

Not every model should use every variable. The point is to establish a defensible non-AI benchmark appropriate to the outcome being predicted.

Model B: baseline plus predeclared AI variables

The AI-enhanced model should then add a controlled set of AI variables measured at the signal date.

Candidate variables include:

• recommendation coverage;
• change in recommendation coverage;
• presence coverage;
• change in presence coverage;
• average recommendation rank when observable;
• rank change;
• sentiment or framing variables;
• cross-platform breadth;
• Directional Portability Ratio;
• persistence across monthly observations;
• competitive recommendation share once comparable sector definitions are standardized;
• recommendation-share gap once comparable real-world market-share denominators are available.

These variables should not be collapsed into one opaque score before their individual behavior is understood.

The distinction between the metrics is explained in AI Recommendations vs. Mentions vs. Citations, while cross-platform breadth is developed in Does Cross-Platform AI Visibility Matter?.

The key test is incremental value

The relevant comparison is:

Baseline model performance vs. Baseline + AI model performance

If adding AI variables does not improve out-of-sample performance, the AI signal may be descriptively interesting but financially redundant.

If AI variables improve prediction only in-sample, the result is also insufficient.

If they improve out-of-sample performance in a stable and economically meaningful way, that would be evidence that AI-search behavior contains information beyond the conventional baseline.

This is why Article 09 defines the core question as incremental rather than correlational.

Walk-Forward Testing Instead of Random Train-Test Splits

Financial time series should not be treated as if observations were randomly interchangeable.

A random train-test split can let future regimes leak into the training sample and can make performance appear more stable than it would have been in practice.

The preferred design is an expanding or rolling walk-forward test.

Example expanding-window design

Assume enough monthly data eventually exists.

1. Train the model on months 1 through 12.
2. Generate forecasts for month 13 using only data available through month 12.
3. Add month 13 to the training set after its outcomes are known.
4. Refit using months 1 through 13.
5. Forecast month 14.
6. Continue forward one period at a time.

This design answers the operational question: what would the model have predicted at that date using only information available then?

Example rolling-window design

If older relationships may become stale because AI platforms change rapidly, a fixed rolling window may be preferable.

For example:

• train on the previous 12 months;
• predict the next month;
• roll the window forward one month;
• repeat.

The expanding-window and rolling-window results should be compared because AI systems may experience structural breaks when models, retrieval systems, search indexes, or product interfaces change.

CFA Institute explicitly describes rolling-window backtesting as a standard way to proxy the real investment process and notes that financial data can experience structural breaks. That is particularly relevant to AI-search data, where platform behavior can change faster than many conventional financial variables.

The Model Complexity Ladder

The research should begin with interpretable models before moving to more flexible machine-learning systems.

A useful sequence is:

1. Univariate tests. Does recommendation momentum by itself relate to later outcomes?
2. Simple multivariate regression. Does the AI variable retain information after obvious controls?
3. Panel regression. Add company, sector, and time structure where enough repeated observations exist.
4. Regularized linear models. Use ridge, lasso, or elastic net when the feature set grows and predictors become correlated.
5. Tree-based models. Random forests or gradient boosting can test nonlinear effects after the sample size is large enough.
6. More complex architectures. Only after simpler models establish a reason to expect nonlinear interactions and the dataset can support the additional degrees of freedom.

The goal is not to maximize historical fit.

The goal is to determine whether the AI variables contain a stable relationship that survives forward testing.

A complex model that produces a slightly better in-sample result but unstable out-of-sample performance should not be preferred over a simpler model.

Backtest Overfitting Is a Major Risk

Questions This Section Answers

• Why can a strategy look excellent in historical data and fail later?
• How should multiple AI metrics and thresholds be handled?
• What should happen if many model specifications are tried?

AI-search research creates an unusually large space of possible specifications.

Researchers can vary:

• prompt wording;
• prompt clusters;
• platform subsets;
• company universes;
• recommendation thresholds;
• rank thresholds;
• sentiment transformations;
• time windows;
• persistence rules;
• platform-breadth rules;
• sector definitions;
• parent-company rollups;
• lag lengths;
• outcome horizons;
• baseline variables;
• model types;
• hyperparameters.

If enough combinations are tested, some will look successful by chance.

Bailey, Borwein, Lopez de Prado, and Zhu formalized this problem as the probability of backtest overfitting, showing that selecting from many historical strategy configurations can produce apparent winners that fail out of sample.

See The Probability of Backtest Overfitting.

For AI Investor Signals, the practical response is to reduce researcher degrees of freedom.

Predeclare the main specification

Before evaluating each new outcome window, freeze:

• the company universe;
• signal date;
• prompt universe;
• AI platforms;
• data-cleaning rules;
• entity mappings;
• primary AI variables;
• baseline variables;
• target outcome;
• target horizon;
• model family;
• evaluation metrics;
• missing-data policy;
• outlier policy;
• portfolio construction rules if returns are tested.

Report specification sensitivity

Reasonable alternative specifications should be shown as sensitivity tests rather than silently searched until one works.

Examples include:

• clean-dedupe vs. no-dedupe AI signal;
• primary cell-state vs. capture-average sensitivity;
• four-of-six vs. five-of-six platform breadth;
• fixed-prompt subset vs. broader prompt panel;
• equal-weight vs. exposure-weighted parent-company mapping;
• alternative sector controls.

Preserve failed tests

A research program becomes less credible if only successful specifications are published.

Failed and null tests are part of the evidence.

That principle is built into What Would Prove the AI Commercial Momentum Hypothesis Wrong?.

Outcome 1: Branded Search, Web Traffic and App Engagement

The closest downstream outcomes are behavioral.

If AI recommendations place a company into more consideration sets, one plausible intermediate effect is an increase in brand-directed activity.

Potential outcomes include:

• branded search volume;
• direct website traffic;
• organic branded traffic;
• app visits;
• app downloads;
• retailer-page visits;
• branded marketplace searches.

A roughly 30-day horizon can be useful for these variables because consumer behavior may respond faster than quarterly revenue.

The test should compare the future change in the behavioral metric with information available at the AI signal date.

Example specification:

Future branded-search change = baseline variables + AI recommendation momentum + sector/time controls

The AI coefficient alone is not the final result.

Researchers should also compare:

• out-of-sample MAE;
• out-of-sample R-squared or delta R-squared;
• direction accuracy;
• rank correlation between predicted and realized change.

If the AI-enhanced model does not improve these metrics, the commercial mechanism weakens before the research ever reaches stock returns.

Outcome 2: Analyst Revenue Revisions

Analyst estimate revisions are a particularly important target because they sit between commercial activity and reported results.

The question is not whether analysts are right or wrong. It is whether AI recommendation momentum contains information that appears before professional revenue expectations change.

Possible targets include:

• 30-day change in next-quarter revenue consensus;
• 30-day change in next-12-month revenue consensus;
• fraction of analysts revising upward or downward;
• direction of net estimate revisions;
• magnitude of standardized revision relative to prior estimate dispersion.

The baseline should include the consensus estimate level, prior revision trend, prior company results, sector trends, price momentum, and other information already known at T.

Then add the AI variables.

For direction classification, useful evaluation metrics include:

• accuracy;
• balanced accuracy when classes are uneven;
• area under the ROC curve;
• precision and recall for large revisions;
• calibration if predicted probabilities are produced.

For continuous revision magnitude, useful metrics include:

• MAE;
• RMSE;
• delta R-squared;
• Spearman rank information coefficient across companies.

If AI momentum reliably precedes revenue-estimate revisions out of sample, that would be important evidence for the future AI Visibility Market Divergence framework.

Outcome 3: Revenue Growth and Revenue Surprise

Revenue is closer to the core economic hypothesis than stock price.

There are several distinct revenue outcomes, and they should not be combined casually.

Reported revenue growth

Examples:

• year-over-year quarterly revenue growth;
• sequential revenue growth where seasonality is controlled;
• segment revenue growth when the AI-tracked brand maps only to one business segment.

Revenue surprise

Revenue surprise compares reported revenue with the consensus expectation that existed before the report.

That distinction matters because a company can grow rapidly and still disappoint if expectations were higher.

Revenue revision before the report

Analysts may update expectations before the company reports. If AI recommendation momentum predicts those revisions, it may be informative even if the final surprise is small.

The backtest should therefore distinguish:

1. future growth;
2. future change in expectations;
3. final surprise relative to expectations.

These are separate dependent variables.

A fixed 90-day horizon can be used as one benchmark, but event-based testing around the next quarterly report is often more appropriate because reporting calendars differ.

Outcome 4: EPS Surprise and Earnings Trajectory

The AI Commercial Momentum Hypothesis is primarily about consumer consideration and commercial demand, so EPS should be treated as a later-stage outcome.

Earnings can diverge from revenue because of:

• margins;
• operating leverage;
• interest expense;
• credit costs;
• reserve changes;
• taxes;
• stock-based compensation;
• restructuring;
• acquisitions;
• share count;
• one-time items.

An AI-search signal could be commercially correct and still fail to predict EPS.

For that reason, a null EPS result should not automatically invalidate an earlier validated revenue relationship.

But if the signal is eventually presented as an earnings or valuation tool, it must separately demonstrate value against earnings outcomes.

Outcome 5: Stock Performance

Questions This Section Answers

• How should stock returns be tested without confusing commercial momentum with market performance?
• What does sector-neutral performance mean?
• Why should transaction costs and turnover be included?

Stock returns are the hardest test and should come later.

A stock can underperform despite improving commercial momentum if the improvement was already expected, valuation was high, macro conditions changed, margins deteriorated, or investors rotated away from the sector.

Conversely, a company can outperform despite weakening AI recommendation momentum for reasons unrelated to consumer discovery.

The stock test should therefore focus on incremental, risk-adjusted and expectation-aware performance, not raw returns alone.

Start with total returns

Use total returns, including dividends and relevant corporate actions, rather than price-only returns.

The company universe should preserve delisted, acquired, bankrupt, or otherwise discontinued firms to the extent the historical data allow. Removing them after the fact creates survivorship bias.

Compare with sector-relative returns

For a first investor-oriented test:

Sector-relative return = company total return - appropriate sector benchmark return

This helps reduce the risk that a signal simply captures a sector-wide move.

Add factor controls

A more complete test can estimate returns after controlling for broad market and style factors.

The Kenneth R. French Data Library provides current research factors including market excess return, size, value, profitability, and investment factors, with historical archives that can help preserve vintage-consistent tests. See the Kenneth R. French Data Library.

Depending on the design, later tests can also control for momentum and sector exposure.

The goal is not to claim a particular factor model is perfect. The goal is to avoid labeling ordinary beta, sector, size, or style exposure as AI-generated alpha.

Use information coefficients before portfolio claims

Before simulating a trading strategy, compute the cross-sectional relationship between the AI signal and later sector-relative returns.

A common measure is the Spearman rank information coefficient:

IC = rank correlation between signal score at T and future return

The IC should be calculated repeatedly across independent forward periods, not once on the full history.

Researchers should report:

• average IC;
• median IC;
• fraction of periods with the expected sign;
• dispersion across periods;
• sector-specific ICs;
• sensitivity to signal definition.

Portfolio spreads require a larger universe

The initial 25-company public panel is too small to support strong decile-portfolio conclusions.

When the universe expands enough, a later test could rank companies by a predeclared AI signal, form sector-neutral groups, and compare higher-signal with lower-signal groups.

Potential outputs include:

• top-minus-bottom decile return;
• equal-weight and value-weight results;
• sector-neutral spreads;
• gross and net returns;
• volatility;
• maximum drawdown;
• turnover;
• factor-adjusted alpha;
• hit rate across rebalance periods.

A portfolio result should not be published without the costs required to implement it.

Transaction Costs, Turnover and Capacity

A paper backtest can look attractive because it assumes frictionless trading.

Real strategies face:

• bid-ask spreads;
• commissions where applicable;
• market impact;
• slippage;
• taxes depending on account type and jurisdiction;
• borrowing costs for short positions;
• stock availability constraints;
• turnover;
• delays between signal measurement and execution.

Even a statistically real return relationship may be economically useless if turnover is too high.

Therefore later portfolio simulations should report both:

gross performance and performance after predeclared implementation costs.

Turnover should also be analyzed as a signal property. If monthly AI rankings reverse constantly, the implementation burden itself becomes evidence that the signal may not be durable enough for investment use.

Sector Neutrality and Company Heterogeneity

The current public-company panel spans banks, insurance, fintech, brokerage, crypto, mortgage, lending, and healthcare.

Those sectors have different economics.

They differ in:

• revenue models;
• regulation;
• customer-acquisition cycles;
• frequency of consumer decisions;
• sensitivity to interest rates;
• exposure to AI-mediated discovery;
• brand-to-parent mapping;
• reporting cadence;
• market-share data quality.

A pooled model that ignores those differences can create misleading results.

Backtests should therefore use one or more of the following where sample size permits:

• sector fixed effects;
• sector-by-time effects;
• within-sector ranks;
• company fixed effects;
• exposure weighting for tracked brands or segments;
• sector-specific models;
• hierarchical models that estimate an overall relationship while allowing sector variation.

The sector articles beginning with Bank Stocks and AI Search will preserve the sector-specific starting observations for later comparison.

Parent-Company Exposure Must Be Modeled, Not Assumed

Several current public-company rows are measured through a product, division, operating brand, or legacy entity rather than the entire public parent.

Examples in the current panel include:

• Axos Financial through Axos Bank and UFB Direct;
• Coinbase Global through Coinbase Wallet;
• CVS Health through CVS Pharmacy;
• UnitedHealth Group through UnitedHealthcare-related consumer brands;
• Goldman Sachs through Goldman Sachs and Marcus;
• Labcorp through Labcorp OnDemand;
• Happen through the legacy LendingClub entity during a rebrand transition.

A strong signal for a small subsidiary should not automatically be treated as a strong signal for consolidated revenue.

Future financial backtests should therefore estimate economic exposure where possible.

Potential exposure measures include:

• segment revenue share;
• segment operating-income share;
• customer count;
• geographic exposure;
• product revenue share;
• management-disclosed segment importance.

A parent-level AI variable could then be weighted by the economic relevance of the tracked entity.

This issue is especially important for the future AI Recommendation Share vs. Market Share framework.

Statistical Inference for the Panel

A growing monthly panel will contain repeated observations for the same companies and related prompts.

Ordinary regression standard errors may therefore be too optimistic if observations are correlated within company, prompt, sector, or time.

Depending on the model and sample size, later analyses should consider:

• company-clustered standard errors;
• two-way clustering by company and time where appropriate;
• prompt-clustered uncertainty for AI-side measurements;
• block bootstrap procedures that preserve temporal dependence;
• sector clustering in sensitivity tests;
• robust standard errors for heteroskedasticity.

The current V0 AI-side interval is already prompt-clustered at the signal-construction stage. That interval should not be confused with the uncertainty of the later financial model. The two analyses answer different questions and require separate inference.

Outliers and Winsorization

Financial variables can contain extreme values, especially percentage growth and surprise measures when the denominator is small.

Outlier treatment should be decided before researchers inspect whether it improves the desired result.

If winsorization is used, the rules should be predeclared, for example:

• 1st and 99th percentile within a training window;
• sector-specific limits where economically justified;
• no winsorization of the AI signal unless a measurement artifact is documented.

Results should be shown with and without reasonable outlier treatment.

The purpose is not to make the model look smoother. It is to prevent a few extreme observations from dominating the relationship while preserving transparency about the choice.

Missing Data and Data Availability

Missing data must remain missing unless there is a defensible imputation rule.

This applies to both AI and financial variables.

An explicit AI extraction failure is not a zero recommendation. That rule is already built into How We Measure AI Commercial Momentum.

Likewise:

• an unavailable analyst estimate is not a zero estimate;
• a missing web-traffic series is not zero traffic;
• a missing July AI dataset is not zero visibility;
• a company without a comparable market-share denominator should not receive an invented market-share gap.

The backtest should report sample size for every model, not only the headline universe size.

A Practical Evaluation Scorecard

No single performance metric should decide whether the AI signal works.

Continuous outcomes

For revenue growth, estimate revisions, traffic, or other continuous variables:

• mean absolute error (MAE);
• root mean squared error (RMSE);
• out-of-sample R-squared;
• delta R-squared versus baseline;
• mean absolute percentage error only when denominators are stable enough for it to be meaningful.

Directional outcomes

For up/down classification:

• direction accuracy;
• balanced accuracy;
• AUC;
• precision and recall for large positive or negative outcomes.

Cross-sectional ranking

For comparing companies at the same date:

• Spearman information coefficient;
• top-vs-bottom group separation;
• monotonicity across quantiles;
• sector-neutral rank correlation.

Portfolio outcomes

Only after the signal passes earlier stages:

• gross return;
• net return after costs;
• volatility;
• drawdown;
• turnover;
• sector and factor exposure;
• factor-adjusted alpha;
• stability across rebalance periods.

A result should be considered more credible when several metrics point in the same direction rather than when one optimized metric looks unusually strong.

The Backtesting Decision Matrix

What Would Count as a Successful Backtest?

A successful backtest should not be defined as one significant coefficient or one profitable historical portfolio.

A stronger standard is cumulative evidence.

For example, the AI signal would become more credible if it demonstrates most of the following:

1. Prospective time ordering. Signals are frozen before outcomes.
2. Incremental value. Baseline + AI beats baseline alone.
3. Out-of-sample stability. Improvement persists in walk-forward periods.
4. Cross-sector or predeclared sector validity. The relationship is not created by one retrospectively chosen niche.
5. Metric robustness. Results are not dependent on one cleaning choice or one AI platform.
6. Economic relevance. Effect sizes are large enough to matter.
7. Reasonable turnover. Any investment implementation survives realistic costs.
8. Risk adjustment. Return results are not merely market, sector, size, value, or momentum exposure.
9. Replication. Later data and, ideally, independently collected data show similar behavior.
10. Transparent failures. Negative tests are published rather than hidden.

No single item is enough by itself.

What Would Count Against the Signal?

The backtest should weaken the hypothesis if:

• AI variables add no out-of-sample value beyond the baseline;
• coefficients change sign repeatedly;
• results depend on one platform;
• results depend on one prompt subset selected after the fact;
• performance vanishes after cleaning, exposure weighting, or sector controls;
• the relationship appears only after the downstream outcome is already visible;
• stock-performance results disappear after factor adjustment;
• turnover and costs eliminate the simulated return;
• results fail in new months or new companies;
• the signal is economically trivial even when statistically detectable.

These conditions are consistent with the precommitted falsification framework rather than being new rules invented after the fact.

Why the Current Dataset Is a Starting Point, Not a Finished Backtest

The current public-company prototype contains 25 mapped public parents and an initial July-September 2026 matched observation window.

That is enough to freeze examples of AI recommendation movement.

It is not enough to make a credible claim about long-run revenue prediction or stock-performance predictiveness.

The present dataset lacks the temporal depth required to observe enough independent cycles of:

• signal formation;
• analyst revision;
• quarterly reporting;
• earnings response;
• market repricing.

The current 25-company panel is also too small for robust decile portfolio construction.

The correct use of the current data is prospective:

1. preserve the signals now;
2. keep collecting future AI observations under governed methodology;
3. append financial outcomes only after they become public;
4. run the same predeclared models on successive forward windows;
5. publish both positive and negative results.

The initial company classifications already provide a testable historical record. For example, Axos Financial and MetLife are positive AI divergence candidates under the V0 AI-side rules, while eleven companies are negative candidates and twelve are mixed or neutral. Those labels should remain frozen as descriptions of the original AI measurement, regardless of what later financial results show.

That is the point of prospective alternative-data research.

Questions This Section Answers

• Can ChatGPT recommendations be backtested against stock performance?
• Can AI search data be used as alternative data for investors?
• What evidence would justify using AI visibility in an investment model?

Can ChatGPT Recommendations Be Backtested Against Stock Performance?

Yes, AI recommendation signals can be backtested against later stock performance, but a valid design requires more than collecting historical ChatGPT outputs and comparing them with returns.

The signal must be timestamped, prompt-governed, free of look-ahead leakage, mapped correctly to the public parent, evaluated out of sample, and compared against market, sector, style, and conventional information available on the signal date.

The backtest should also include other AI platforms because Article 12 shows that company recommendation movement can differ materially across ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, and Perplexity.

The key empirical question is not whether a platform sometimes names companies that later perform well. It is whether a predeclared AI signal improves forward prediction on repeated unseen periods.

Can AI Search Data Be Used as Alternative Data for Investors?

It can be studied as alternative data today because it is a measurable nontraditional information source.

But financial usefulness must be earned through validation.

The research sequence is:

AI visibility measurement -> AI Commercial Momentum -> commercial leading-indicator validation -> financial expectations validation -> market-divergence testing -> stock-return testing

That progression is also described in What Is an AI Investor Signal?.

What Evidence Would Justify Using AI Visibility in an Investment Model?

The strongest evidence would be repeated walk-forward results showing that predeclared AI variables improve the prediction of commercial or financial outcomes beyond a reasonable baseline, survive sector and risk controls, remain useful in later periods, and produce economically meaningful results after implementation costs where a portfolio is simulated.

Until that evidence exists, AI recommendation momentum should be treated as an exploratory research variable rather than an investment decision rule.

Relationship to AI Visibility Market Divergence

Backtesting is the bridge between AI-side measurement and any future market-divergence framework.

AI Visibility Market Divergence asks whether a validated AI-derived commercial signal differs from analyst or market expectations.

That question cannot be answered credibly until the AI signal first demonstrates predictive information about commercial outcomes.

Article 13 therefore supplies the test architecture needed before the research can move from:

AI momentum

to

AI-implied commercial expectation.

The same logic applies to AI Recommendation Share vs. Market Share. A recommendation-share gap becomes financially interesting only if changes in that gap predict later changes in measurable business outcomes.

What This Does Not Mean

This article does not establish that AI search predicts revenue, earnings, analyst revisions, market share, valuation, or stock returns.

It does not establish that companies with positive AI momentum should outperform companies with negative AI momentum.

It does not establish that any current V0 company signal is a buy, sell, hold, undervalued, or overvalued conclusion.

It does not assume that every sector is equally exposed to AI-mediated consumer discovery.

It does not assume that recommendation coverage is the only AI metric that could eventually matter.

It defines a framework for finding out which, if any, of those relationships survive disciplined forward testing.

Methodology Integration With the V0 Signal

The starting AI signal follows the published V0 methodology:

• same normalized prompt and same platform family are matched across periods;
• July 2026 is the base month when available, with Travelers using August because July is unavailable in its matched panel;
• explicit extraction failures are excluded rather than counted as zero visibility;
• response-identical cross-vertical duplicates are collapsed in the primary panel;
• company-array fields provide explicit presence, recommendation, rank, sentiment, and not-mentioned states;
• related brands and entity variants can be rolled to a public parent;
• recommendation change is measured in percentage points;
• platform direction is preserved separately;
• the exploratory 95% AI-side interval is prompt-clustered;
• no-dedupe and capture-average sensitivities are retained;
• watch categories depend on magnitude, interval direction, and platform breadth.

The backtest must use the historical version of those rules that existed at the signal date or explicitly mark later methodology versions.

If methodology changes, V0 should not be overwritten. A new version should be tested side by side so researchers can distinguish genuine improvement from retrospective rule changes.

Limitations of the Proposed Backtesting Framework

1. Short initial history. July-September 2026 is not enough for a long-horizon investment backtest.
2. Small initial public universe. Twenty-five public parents do not support robust decile portfolios or broad cross-sector inference.
3. Prompt-population risk. The research prompts are not search-volume-weighted samples of all AI usage.
4. Platform instability. AI systems, retrieval layers, model versions, and interfaces can change rapidly.
5. Entity mapping. Some measured brands represent only part of a public parent's economics.
6. Outcome availability. Analyst estimates, segment revenue, traffic, app data, and market-share data vary in quality and history.
7. Asynchronous timing. AI observations, analyst updates, financial reports, and market prices occur on different schedules.
8. Data-vintage risk. Revised historical datasets can inadvertently introduce information unavailable at the original signal date.
9. Multiple testing. The number of plausible AI variables, prompts, horizons, and models creates substantial overfitting risk.
10. Nonstationarity. A relationship that works during one generation of AI products may weaken after model or product changes.
11. Economic exposure. Consumer AI search may be material for some businesses and largely irrelevant for others.
12. Causality. Predictive performance would not by itself prove that AI recommendations cause the financial outcome.
13. Implementation frictions. A statistically significant stock-return relationship may be too costly or unstable to trade.

What We Will Test Next

The research program will extend the frozen AI observations forward rather than reconstruct the signal after the outcome.

Priority tests include:

1. 30-day commercial behavior: branded search, website traffic, app engagement, and related demand indicators where comparable data are available.
2. 30-day analyst expectations: direction and magnitude of revenue-estimate revisions.
3. Next reported quarter: revenue growth, revenue surprise, and later EPS surprise where the tracked business has meaningful parent-company exposure.
4. Persistence: whether one-month AI movements that persist across multiple observations are more informative than transient changes.
5. Portability: whether signals moving across more AI platforms outperform similarly sized but platform-concentrated signals.
6. Recommendation share: whether properly normalized AI recommendation-share gaps lead changes in market share or company growth.
7. Later stock-performance tests: 180-day sector-relative and factor-adjusted total returns after the commercial relationship has been tested.

The historical record will remain available through the AI Investor Signal Tracker.

Related LLM Authority Index Research

• Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis
• 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
• 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

External Methodology References

1. CFA Institute. Backtesting & Simulation. 2026 CFA Program Level II Portfolio Management. https://www.cfainstitute.org/insights/professional-learning/refresher-readings/2026/backtesting-and-simulation
2. CFA Institute. Active Equity Investing: Strategies. 2026 CFA Program. https://www.cfainstitute.org/insights/professional-learning/refresher-readings/2026/active-equity-investing-strategies
3. Bailey, David H., Jonathan Borwein, Marcos Lopez de Prado, and Qiji Jim Zhu. The Probability of Backtest Overfitting. Journal of Computational Finance / SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2326253
4. Kenneth R. French Data Library. Fama/French Research Factors and Historical Archives. https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html
5. CFA Institute. Capital Market Expectations, Part I: Framework and Macro Considerations. 2026 CFA Program. https://www.cfainstitute.org/insights/professional-learning/refresher-readings/2026/capital-market-expectations-part-i