Citigroup AI Search Visibility: Initial AI Recommendation Momentum Findings for Citi

Research status: Exploratory longitudinal research. Citi's AI recommendation momentum has not been validated as a predictor of deposit growth, card spending, account growth, wealth flows, revenue, earnings, analyst revisions, valuation, or stock returns.

Observation window: July through September 2026

Ticker: C

Public parent: Citigroup Inc.

Tracked entities: Citi and Citigroup Inc.

Exposure type: Direct/core brand; entity variants merged

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

Current methodology version: V0

Answer Capsule

Citigroup recorded a small negative but clearly inconclusive change in AI recommendation coverage in the initial LLM Authority Index public-company panel.

Across 111 matched prompt-platform cells, Citi recommendation coverage declined from 23.4% in July to 21.6% in September 2026, a change of -1.8 percentage points.

The exploratory 95% interval was wide, ranging from approximately -8.9 to +5.3 percentage points, and therefore crossed zero by a substantial margin. The platform pattern was balanced rather than directional:

• ChatGPT: 0.00 pp
• Gemini: 0.00 pp
• Google AI Mode: +4.76 pp
• Google AI Overviews: -5.88 pp
• Microsoft Copilot: +12.50 pp
• Perplexity: -25.00 pp

Two platform families improved, two worsened, and two were unchanged.

Citi therefore receives a Medium AI-measurement confidence classification and remains Mixed / neutral under the current V0 watch rules.

Simple presence was already extremely high and changed very little, declining from 95.5% to 94.6%, a difference of only -0.9 percentage points. Average recommendation rank moved more noticeably, worsening from approximately 2.74 to 4.38 among responses where Citi was recommended.

That combination is analytically important. Citi continued to appear in nearly all eligible responses, recommendation frequency changed only modestly, yet average recommendation position weakened substantially within the subset of responses where Citi was recommended.

The AI-side conclusion is therefore narrow: Citi remained broadly visible across the matched panel, showed little aggregate movement in recommendation frequency, and experienced a weaker average recommendation position despite sharply different platform-level movements.

This does not establish weaker deposit growth, card activity, wealth flows, customer acquisition, revenue, earnings, or stock performance.

The prospective question remains the one defined by the AI Commercial Momentum Hypothesis: do persistent changes in unbranded AI recommendation behavior contain incremental information about future commercial outcomes after accounting for information already available when the signal was measured?

The company-level signal construction is documented in How We Measure AI Commercial Momentum, while the original 25-company panel is preserved in Initial Findings From 25 Public Companies.

Citi AI Recommendation Momentum at a Glance

Additional V0 signal properties:

Citi sits near the middle of the initial Bank Stocks and AI Search sector distribution. Axos Financial showed a large positive aggregate change, while Ally Financial, Bank of America, Chime, and Goldman Sachs / Marcus showed larger negative changes under the same V0 framework.

Citi's result is useful precisely because it does not fit either extreme.

Questions This Section Answers

• How much did Citi's AI recommendation coverage change?
• Why is Citi still classified as Mixed / neutral?
• What changed beneath the small aggregate movement?

How Much Did Citi's AI Recommendation Coverage Change?

Citi recommendation coverage declined 1.8 percentage points, from 23.4% to 21.6% across the matched July-to-September panel.

The magnitude is small relative to many of the larger movers in the initial public-company dataset.

The exploratory interval also provides no directional support. It ranges from approximately -8.9 to +5.3 percentage points, so the observed point estimate is compatible with a range of underlying movements that includes zero and positive values.

Under the V0 framework, this means the result should remain explicitly inconclusive.

It would be inappropriate to convert a -1.8-point aggregate change into a negative commercial narrative when the uncertainty band is this wide and the platform evidence is balanced.

This is exactly why the project preserves mixed observations rather than ranking every company as a winner or loser.

Why Is Citi Still Mixed / Neutral?

Citi remains Mixed / neutral for three independent reasons.

First, the aggregate change is small.

A decline of 1.8 percentage points is below the current magnitude thresholds used for stronger watch classifications.

Second, the exploratory interval crosses zero substantially.

The interval extends from -8.9 to +5.3 percentage points, which means the aggregate direction is not statistically directional under the current exploratory approach.

Third, the platform pattern is balanced.

Two platforms improved, two worsened, and two were unchanged.

That is very different from cases such as Bank of America, where all six measured platform families declined, or Goldman Sachs / Marcus, which also showed six-of-six negative direction.

The distinction between aggregate movement and cross-platform breadth is discussed in Does Cross-Platform AI Visibility Matter?.

Citi provides a useful opposite case: modest aggregate movement with substantial platform disagreement.

What Changed Beneath the Small Aggregate Movement?

The most important underlying change was not recommendation coverage. It was average recommendation rank.

Citi's average recommendation rank worsened from approximately 2.74 to 4.38.

Because lower numerical rank is better, the September value indicates that Citi was positioned lower on average within the responses where it was recommended.

At the same time:

• recommendation coverage declined only 1.8 percentage points;
• presence declined only 0.9 points; and
• Citi remained present in approximately 94.6% of eligible matched responses in September.

This is a strong example of why a single visibility score would be misleading.

A composite metric could conceal the fact that:

1. Citi remained almost universally present;
2. recommendation frequency was broadly stable to modestly lower; and
3. recommendation position weakened materially when Citi was recommended.

The measurement distinctions are defined in AI Recommendations vs. Mentions vs. Citations.

Presence, recommendation frequency, rank, citations, sentiment, and downstream business outcomes answer different questions. They should not be collapsed into a single score before longitudinal validation establishes which dimensions matter economically.

Platform Differences: Copilot Improved While Perplexity Declined Sharply

Citi's six-platform pattern was highly fragmented.

The largest positive movement came from Microsoft Copilot, at +12.50 percentage points.

The largest negative movement came from Perplexity, at -25.00 points.

Those two platform changes move in opposite directions by nearly 38 percentage points.

That divergence helps explain why Citi's aggregate change is not especially informative by itself.

A blended average can make the company look stable while masking platform-specific movements that may reflect different retrieval systems, source ecosystems, product-query mixes, or recommendation behaviors.

The correct interpretation is not that one platform is right and another is wrong. The correct interpretation is that Citi's recommendation momentum was not portable across platforms during this observation window.

Future validation should test whether platform-specific changes have different relationships with later commercial outcomes, or whether the disagreement is mostly measurement noise.

Citi Was Already Highly Visible Before the Recommendation Change

Citi's presence coverage was 95.5% in July and 94.6% in September.

That is an unusually high baseline level of simple presence.

When a brand is already present in nearly every eligible response, presence has limited room to increase further. In that situation, recommendation frequency and recommendation position may provide more useful differentiation than simple appearance rate.

Citi illustrates this clearly.

The company did not disappear from AI responses. It remained highly visible.

The more nuanced question is how often Citi moved from being merely present to being actively recommended, and where it ranked when that happened.

The initial data suggest that recommendation frequency was broadly stable to modestly lower, while average position weakened.

That is a very different statement from saying Citi "lost AI visibility."

How Citi Compares With Other Bank-Related Companies in the Initial Panel

Citi's mixed result sits between a large positive outlier and several broader negative cases in the initial bank research slice.

These comparisons are descriptive only.

They do not establish that Citi's commercial outlook is stronger or weaker than any peer. The companies also have different customer journeys, product mixes, and economic exposures.

The comparison is useful because it shows how the same measurement framework preserves different signal shapes rather than forcing one sector-wide conclusion.

Citi and Citigroup Entity Normalization

The V0 parent row merges Citi and Citigroup Inc. entity variants.

That normalization is necessary because AI systems may refer to a public company by its consumer brand, formal corporate name, or both.

Without entity normalization, the same economic organization could be split into multiple rows and undercounted.

The merged row is appropriate for parent-level measurement, but it does not mean every Citi business line is equally represented by the underlying prompt set.

Citi's own current investor materials describe a broad franchise spanning institutional banking, wealth management, and U.S. personal banking. The company also operates across more than 180 countries and jurisdictions. These business lines have very different relationships with consumer AI recommendation behavior.

Official Citi investor materials: https://www.citigroup.com/global/investors

That matters for future validation.

An AI recommendation signal generated from consumer-oriented prompts may map more directly to personal banking, cards, deposits, or wealth consideration than to institutional services, markets, or cross-border corporate banking.

The economic-exposure problem should therefore be modeled, not assumed away.

Why the Timing of Citi's Financial Disclosures Matters

Prospective validation depends on strict information timing.

Citi released its second-quarter 2026 results on July 14, 2026, early in the observation window. Those results were therefore already public and belong in any baseline model that later tests whether AI variables add incremental predictive value.

Official Q2 2026 earnings archive: https://www.citigroup.com/global/investors/quarterly-earnings

Citi has scheduled its third-quarter 2026 earnings release for October 13, 2026, after the September AI signal was frozen.

Official Q3 2026 announcement: https://www.citigroup.com/global/news/press-release/2026/citi-third-quarter-2026-earnings-call

That creates a clean prospective checkpoint.

The project should preserve the September AI observation exactly as published, then later compare it with Q3 results, analyst-estimate changes, and other commercial indicators without rewriting the original signal.

The proper test is not whether the -1.8-point AI movement can be retroactively narrated to fit whatever happens next.

The proper test is whether this frozen, mixed signal provides incremental information relative to the financial and market information that was already available by September 30.

Questions This Section Answers

• Does Citi's high AI presence mean the signal is already strong?
• Which future Citi outcomes are most relevant for validation?
• What would falsify the idea that Citi's AI recommendation data has commercial value?

Does Citi's High Presence Mean the Signal Is Already Strong?

No.

A 94.6% September presence rate means Citi appeared in almost every eligible matched response. It does not mean Citi was recommended in almost every response, ranked first, or gained commercial share.

Recommendation coverage was only 21.6% in September, far below simple presence.

The gap between those two measurements is exactly why the framework treats them separately.

High presence can indicate broad brand salience without equivalent recommendation strength.

Which Future Citi Outcomes Are Most Relevant for Validation?

The most relevant downstream variables depend on which consumer journeys dominate the underlying prompt population.

Potential Citi-aligned outcomes include:

• branded search demand;
• visits to consumer banking or card product pages;
• checking and savings account acquisition;
• deposit growth in relevant consumer businesses;
• credit-card account growth and card spending where prompt alignment supports it;
• wealth-management client acquisition or flows where relevant;
• analyst revenue-estimate revisions;
• reported segment revenue and customer metrics;
• later earnings surprises; and
• sector-relative stock returns only after commercial validity is established.

These should be tested against appropriate baseline variables such as rates, deposit pricing, credit conditions, card spending trends, macroeconomic conditions, previously disclosed financial results, and market expectations.

The broader validation protocol is defined in Can AI Search Visibility Predict Revenue Growth? and How Investors Could Backtest AI Search Signals.

What Would Falsify the Citi Signal Hypothesis?

The current Citi observation is already mixed, so future testing should be especially resistant to narrative drift.

The idea that Citi AI recommendation data has useful commercial information would be weakened if:

1. recommendation changes fail to precede any relevant consumer-demand or segment-level outcomes;
2. traditional variables fully explain the same future outcomes;
3. apparent relationships disappear in walk-forward testing;
4. only one platform repeatedly contributes useful movement while the aggregate signal does not;
5. rank changes and coverage changes point in different directions without a stable relationship to outcomes;
6. results depend heavily on retrospective prompt changes; or
7. apparent financial relationships emerge only after selecting favorable horizons or excluding failed tests.

A mixed observation that remains mixed later is a valid research result.

The project does not require every company to produce a usable financial signal.

What This Does Not Mean

The Citi V0 observation does not mean:

• Citi is losing consumer banking share;
• deposit growth will weaken;
• card spending or card acquisition will decline;
• wealth-management growth will weaken;
• revenue or earnings will miss expectations;
• Citi's stock will underperform;
• Perplexity is more informative than Copilot; or
• the worsening average recommendation rank has been validated as financially important.

The current finding is only that Citi's aggregate recommendation coverage moved slightly lower, while simple presence stayed very high, average recommendation rank worsened, and platform directions offset one another.

That is an AI measurement result, not an investment conclusion.

Methodology

Citi is measured using the same V0 matched-panel framework applied across the initial public-company research.

Eligible denominator

Recommendation coverage is calculated across matched prompt-platform cells where the same normalized prompt and AI platform family are observed in the comparison periods.

Matched-panel construction

The current Citi row contains 111 matched cells and 93 normalized prompt clusters across six AI platform families.

Recommendation classification

A company must be identified as a valid recommendation in an eligible response. Mere appearance does not count as recommendation coverage.

Presence

Presence is measured separately and records whether the normalized Citi entity appeared in the response at all.

Rank

Average recommendation rank is calculated conditionally within observations where Citi was recommended. Rank therefore answers a different question from coverage.

Entity normalization

Citi and Citigroup Inc. variants are merged to the Citigroup Inc. public parent to avoid artificial fragmentation of the same economic entity.

Failure handling

Extraction failures are not converted into zeros. Failed observations remain distinct from valid observations in which Citi was not recommended.

Duplicate handling

The source corpus underwent exact-file and observation-level duplicate controls. Citi's no-dedupe sensitivity difference is 0.0 percentage points.

Capture-average sensitivity

The capture-average sensitivity difference is also 0.0 points, indicating no observed change in the Citi point estimate under that sensitivity construction.

Exploratory interval

The V0 exploratory 95% interval is based on variation in prompt-level matched changes rather than a formal causal model. For Citi, that interval is approximately -8.9 to +5.3 percentage points.

Confidence classification

Citi receives Medium AI-measurement confidence because the matched sample and platform coverage meet the current Medium threshold, but the interval is not directional.

The complete methodological framework is documented in How We Measure AI Commercial Momentum.

Limitations

Several limitations are material to the Citi observation.

Three months is a short observation window

July through September cannot establish persistence, seasonality, or predictive validity.

The prompt population is not the entire Citi business

Consumer-oriented AI prompts may map more directly to cards, deposits, consumer banking, and selected wealth decisions than to institutional services, markets, or corporate banking.

Platform disagreement is substantial

Copilot improved while Perplexity declined sharply. Aggregate movement may obscure platform-specific behavior.

Rank is conditional

The average rank change applies only to observations where Citi was recommended. It should not be interpreted as an unconditional visibility metric.

Entity normalization solves one problem but not every exposure problem

Merging Citi and Citigroup variants prevents duplicate fragmentation, but parent-level financial exposure remains broader than the measured consumer-facing prompt set.

No downstream outcome has yet been validated

The current signal has not been shown to predict account growth, deposits, card spending, wealth flows, revenue, earnings, analyst revisions, valuation, or stock returns.

What We Will Test Next

The September Citi observation should remain frozen as a prospective record.

Future analysis should test whether changes in recommendation coverage, rank, presence, and cross-platform breadth add incremental out-of-sample information about:

1. branded search and site demand;
2. consumer account acquisition;
3. deposit and card metrics where prompt alignment is appropriate;
4. wealth client or asset flows where the prompt universe supports that mapping;
5. analyst revenue-estimate revisions;
6. reported segment-level financial outcomes;
7. earnings surprises; and
8. later sector-relative returns only after a commercial pathway is established.

The test should compare baseline models with baseline-plus-AI models and preserve failed, null, and contradictory results.

Related LLM Authority Index Research

• Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis
• Can AI Recommendation Momentum Identify Emerging Company Trends? Initial Findings From 25 Public Companies
• How We Measure AI Commercial Momentum
• AI Investor Signal Tracker
• AI Recommendations vs. Mentions vs. Citations
• Does Cross-Platform AI Visibility Matter?
• Bank Stocks and AI Search
• Axos Financial AI Search Visibility
• Ally Financial AI Search Visibility
• American Express AI Search Visibility
• Bank of America AI Search Visibility
• Chime AI Search Visibility
• Goldman Sachs and Marcus AI Search Visibility