Goldman Sachs and Marcus AI Search Visibility: Recommendation Coverage Declined Across All Six Platforms

Research status: Exploratory longitudinal research. Not validated as a predictor of deposits, consumer banking growth, investment banking revenue, asset management flows, trading revenue, earnings, analyst revisions, valuation, or stock returns.

Observation window: July through September 2026

Public parent: The Goldman Sachs Group, Inc. (NYSE: GS)

Tracked entities: Goldman Sachs Group and Marcus by Goldman Sachs

Exposure type: Parent + consumer banking brand

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

Current methodology version: V0

Answer Capsule

Goldman Sachs and Marcus recommendation coverage declined from 36.5% in July to 25.2% in September 2026, a decrease of 11.3 percentage points across 318 clean matched prompt-platform cells and 256 normalized prompt clusters.

The exploratory 95% interval for the primary clean recommendation change was approximately -15.8 to -6.9 percentage points, remaining below zero. All six measured AI platform families declined.

The decline also appeared in adjacent AI-side measures. Presence coverage fell from 49.4% to 37.7%, down 11.6 points, while average recommendation rank worsened from 3.14 to 3.40.

Under the current V0 framework, Goldman Sachs receives a Medium AI-measurement confidence classification and a Negative AI divergence candidate research label.

Sensitivity analysis matters more for this row than for several prior company articles. The primary clean estimate is -11.32 points, while the no-dedupe estimate is approximately -9.32 points and the capture-average estimate is approximately -9.08 points. The alternative estimates remain negative, but they are roughly two percentage points less negative than the primary clean result.

The row also combines Goldman Sachs Group with Marcus by Goldman Sachs. That makes the measured entity broader than a single consumer brand but less clean than a pure direct/core brand exposure. The AI signal should not be interpreted as measuring every Goldman Sachs business equally.

This result does not establish that Marcus deposits, consumer banking activity, investment banking revenue, trading revenue, asset management flows, earnings, valuation, or GS stock returns will decline.

Key Findings

Additional V0 signal properties:

The broader Bank Stocks and AI Search and Fintech, Brokerage and Crypto Stocks analyses provide the appropriate sector context for this row.

Questions This Section Answers

• How large was the Goldman Sachs / Marcus recommendation-coverage decline?
• Was the decline broad across AI platforms?
• Why do the sensitivity checks and combined entity mapping matter?

How Large Was the Goldman Sachs / Marcus Recommendation-Coverage Decline?

The primary clean matched-panel estimate shows recommendation coverage declining 11.3 percentage points, from 36.5% in July to 25.2% in September 2026.

The clean comparison is based on 318 matched prompt-platform cells representing 256 normalized prompt clusters.

The exploratory interval extends from approximately -15.8 to -6.9 percentage points. Because the interval remains below zero, the current V0 framework classifies the recommendation direction as negative.

The row receives a Medium AI-measurement confidence classification.

That classification should not be read as a financial forecast.

It describes the current strength of the AI-side measurement after accounting for matched-panel size, interval direction, platform representation, and sensitivity behavior.

The signal is directionally clearer than Mixed / neutral rows such as Citi or Labcorp, but it also shows more sensitivity to cleaning and capture treatment than several other negative rows.

That is why the primary estimate and the alternative estimates should remain visible together rather than reducing the result to a single percentage-point number.

The distinction between AI-side measurement and financial prediction is central to the AI Commercial Momentum Hypothesis, the initial 25-company findings, and the methodology.

Was the Decline Broad Across AI Platforms?

Yes.

Recommendation coverage declined on all six measured platform families.

Gemini showed the largest measured decline at -32.26 percentage points, followed by ChatGPT at -21.21 points.

Microsoft Copilot declined 15.00 points, Google AI Mode declined 12.20 points, and Perplexity declined 11.76 points.

Google AI Overviews showed the smallest platform-level change at -1.69 points, but it still moved in the same negative direction.

This six-platform breadth distinguishes the Goldman Sachs / Marcus row from cases where a negative aggregate result is driven by only part of the platform set.

The cross-platform AI visibility analysis treats platform breadth separately from aggregate recommendation change for exactly this reason.

A six-platform decline is descriptively broader than a one- or two-platform decline.

It still does not establish causality, persistence, or financial significance.

The next periods matter because a broad move can reverse, narrow, or persist.

Why the Sensitivity Checks Matter More Here

The Goldman Sachs / Marcus row is one of the cases where the alternative specifications move the estimate noticeably.

The primary clean estimate is:

-11.32 percentage points

The no-dedupe alternative is:

-9.32 percentage points

The capture-average alternative is:

-9.08 percentage points

That produces a no-dedupe sensitivity difference of approximately +2.00 points and a capture-average sensitivity difference of approximately +2.24 points relative to the primary estimate.

The direction remains negative under all three specifications.

The capture-average exploratory interval is also negative, approximately -13.87 to -4.28 percentage points.

This matters because robustness is not binary.

A result can remain directionally negative while still changing materially in magnitude under different cleaning or aggregation choices.

For Goldman Sachs / Marcus, the evidence supports a broad negative AI-side observation, but the exact size of the decline depends more on specification than it does for rows with near-zero sensitivity differences.

That is one reason the row remains Medium rather than High confidence under the current V0 framework.

The correct interpretation is not that the result is invalid.

The correct interpretation is that direction is more stable than magnitude.

Why the Goldman Sachs and Marcus Mapping Requires Care

The tracked entities are Goldman Sachs Group and Marcus by Goldman Sachs.

This is a broader and more complicated mapping than a direct consumer brand such as Chime.

Goldman Sachs is a global financial institution with major businesses in Global Banking & Markets and Asset & Wealth Management. Goldman Sachs also identifies Marcus as a consumer-facing offering within its broader business ecosystem.

That means the public-company row combines:

• a parent-level Goldman Sachs entity;
• a consumer banking brand;
• prompts that can touch different customer journeys and financial products; and
• a public parent whose revenue mix extends far beyond consumer banking.

The mapping is therefore useful for public-company research but should not be treated as a single homogeneous commercial exposure.

A decline in Marcus-related consumer recommendations could eventually prove relevant to consumer deposit acquisition or retail financial-product consideration.

A decline in parent-level Goldman Sachs recommendations could reflect something entirely different, including wealth management, institutional finance, advisory, trading, or broader brand consideration.

The current row combines those exposures.

That makes downstream attribution more difficult.

Future versions of the research should test whether parent-level and Marcus-specific signals can be separated without sacrificing matched-panel continuity.

Official Goldman Sachs context:

• Goldman Sachs Investor Relations
• Goldman Sachs Businesses
• Marcus by Goldman Sachs

Recommendation Coverage, Presence, and Rank All Weakened

The three primary AI-side measures moved in the same broad direction.

From July to September:

• recommendation coverage declined from 36.5% to 25.2%;
• presence coverage declined from 49.4% to 37.7%; and
• average recommendation rank worsened from 3.14 to 3.40.

Recommendation coverage asks how often the tracked entities were actually recommended across eligible matched responses.

Presence asks whether the tracked entities appeared at all, even when they were not framed as recommendations.

Average recommendation rank asks where the tracked entities appeared among recommendations in the subset of responses where they were recommended.

These measures should remain separate.

The AI Recommendations vs. Mentions vs. Citations framework keeps recommendation frequency, presence, citations, rank, sentiment, and downstream outcomes conceptually distinct.

For Goldman Sachs / Marcus, all three current measures weakened.

That alignment makes the AI-side observation easier to describe than a mixed case such as Labcorp.

It still does not show that the decline is economically material to Goldman Sachs.

Why This Matters for Investor Research

Goldman Sachs / Marcus is one of the stronger negative-direction test cases in the V0 public-company panel because several AI-side properties move together:

• recommendation coverage declined 11.3 percentage points;
• the primary exploratory interval remains fully below zero;
• all six platforms moved lower;
• presence coverage declined 11.6 points;
• average recommendation rank worsened; and
• the alternative sensitivity estimates remain negative.

At the same time, two counterweights should remain visible:

• the estimated magnitude shifts by roughly two points under alternative cleaning and capture specifications; and
• the public-company row combines a parent entity with a consumer banking brand.

Those characteristics make the row valuable for prospective validation.

If later consumer or firm-level outcomes move in the same direction, the next question will be whether the AI signal contributed information beyond conventional data.

If those outcomes remain strong while the AI signal remains negative, that divergence would be equally informative.

The AI Investor Signal Tracker freezes the observation before future outcomes are known.

Potential Marcus-proximate variables include:

• Marcus branded search demand;
• deposit-account consideration;
• consumer savings product demand;
• consumer banking site traffic;
• digital account-opening activity where observable; and
• changes in consumer deposit balances where the relationship can be isolated credibly.

Potential broader Goldman Sachs variables include:

• investment banking fees;
• advisory activity;
• underwriting activity;
• FICC and equities revenue;
• asset management flows;
• assets under supervision;
• wealth-management client activity;
• net revenues;
• earnings and returns on equity; and
• analyst estimate revisions.

The research should test the closest entity-level outcomes first before extending the signal to the entire company.

Goldman Sachs Compared With Other Banking and Fintech Signals

The public-company panel provides several useful financial-services comparisons.

Axos Financial

Axos Financial moved strongly in the opposite direction.

Its recommendation coverage increased approximately 19.7 percentage points in the current V0 panel and it is classified as a Positive AI divergence candidate.

That makes Axos a useful positive-direction contrast.

Webull

Webull increased modestly in the initial fintech slice but remains Mixed / neutral because its interval crosses zero.

Its consumer financial-platform model differs materially from Goldman Sachs, but it provides another comparison between directional and non-directional signals.

Citi

Citigroup declined only modestly in the initial V0 panel and remains Mixed / neutral.

The contrast with Goldman Sachs / Marcus shows why the magnitude of a point estimate alone is not enough. Interval direction and platform breadth matter.

Ally Financial

Ally Financial declined approximately 5.2 percentage points and worsened on five of six platform families.

Goldman Sachs / Marcus shows a larger aggregate decline and all-six-platform breadth.

American Express

American Express provides another consumer-financial-services comparison with broad negative platform movement.

The customer journey differs from Marcus and from Goldman Sachs' institutional businesses.

Bank of America

Bank of America also declined across all six measured platform families.

This gives the research program a direct comparison between two large financial institutions with broad negative platform breadth but different business mixes and entity structures.

Chime

Chime declined 9.4 percentage points and worsened on five of six platforms.

Chime is a cleaner direct/core brand mapping, while Goldman Sachs / Marcus combines parent and consumer-brand exposures.

The Bank Stocks and AI Search and Fintech, Brokerage and Crypto Stocks articles provide the broader context.

Information Timing and Prospective Validation

Prospective testing requires a strict information cutoff.

Goldman Sachs reported second-quarter 2026 results on July 14, 2026, during the July-to-September AI observation window.

The firm reported:

• $20.34 billion in second-quarter net revenues;
• $6.63 billion in net earnings;
• diluted EPS of $20.98; and
• annualized return on average common shareholders' equity of 23.5%.

Goldman Sachs filed its second-quarter 2026 Form 10-Q on August 3, 2026.

Those financial results and filings were already public before the September AI observation was frozen.

They therefore belong in the baseline information set for future testing.

This timing distinction is critical.

The strong Q2 results cannot be treated as an outcome predicted by a July-to-September AI decline because the results were already public during the observation window.

Instead, prospective validation begins after the signal is frozen.

Future analysis should ask whether the September AI-side observation adds information beyond already-known financial performance, deal pipelines, market activity, asset and wealth management trends, Marcus consumer data, analyst estimates, market prices, and other conventional information.

The answer is currently unknown.

Official Goldman Sachs sources:

• Goldman Sachs Investor Relations
• Goldman Sachs Q2 2026 earnings release
• Goldman Sachs Quarterly Earnings
• Goldman Sachs Financials

What This Does Not Mean

The Goldman Sachs / Marcus result does not establish that:

• Marcus deposits will decline;
• Marcus account acquisition will weaken;
• consumer banking growth will slow;
• investment banking fees will decline;
• advisory activity will weaken;
• FICC or equities revenue will decline;
• asset management flows will turn negative;
• assets under supervision will decline;
• Goldman Sachs net revenues or earnings will decline;
• analyst estimates will fall;
• Goldman Sachs' valuation is too high or too low;
• GS shares will underperform; or
• AI systems caused any future commercial or financial outcome.

The term Negative AI divergence candidate is an internal V0 research classification for unusual AI-side recommendation movement under predefined thresholds.

It is not an investment recommendation.

The Medium confidence label applies to the AI-side measurement only.

It does not describe confidence in future financial performance.

Methodology

The V0 Goldman Sachs / Marcus signal uses the same core framework documented in How We Measure AI Commercial Momentum.

Matched-panel denominator

The primary clean estimate compares outcomes across 318 matched prompt-platform cells available in both the July baseline and September observation after the current cleaning rules are applied.

The raw matched panel contains 347 cells.

The difference between the raw and clean matched-cell counts is one reason this row deserves explicit sensitivity reporting.

Prompt-level uncertainty

The primary clean exploratory interval is approximately -15.8 to -6.9 percentage points.

The interval remains below zero.

The capture-average alternative interval is also negative, approximately -13.87 to -4.28 percentage points.

These are exploratory measurement intervals.

They are not causal confidence intervals and they are not forecast intervals for deposits, revenue, earnings, valuation, or stock returns.

Platform breadth

All six platform families declined in recommendation coverage.

That is recorded separately from the aggregate estimate because platform breadth and aggregate magnitude answer different questions.

Entity normalization and parent mapping

The tracked row combines Goldman Sachs Group and Marcus by Goldman Sachs and maps them to The Goldman Sachs Group, Inc.

This parent-plus-consumer-brand structure is broader than a single product or brand exposure.

It also creates attribution ambiguity because the parent has major institutional, markets, advisory, asset management, wealth management, and consumer exposures.

The current AI signal should not be assumed to measure those businesses equally.

Sensitivity checks

The primary clean recommendation change is approximately -11.32 percentage points.

The no-dedupe change is approximately -9.32 percentage points, a difference of approximately +2.00 points relative to the primary estimate.

The capture-average change is approximately -9.08 percentage points, a difference of approximately +2.24 points relative to the primary estimate.

The direction remains negative under all three specifications.

The sensitivity results show that the directional conclusion is more stable than the exact magnitude.

Confidence rules

Goldman Sachs receives a Medium V0 AI-measurement confidence classification.

The row has:

• 318 clean matched cells;
• 256 prompt clusters;
• all six platform families represented;
• a primary exploratory interval fully below zero;
• all six platforms moving in the same negative direction; and
• negative alternative sensitivity estimates.

However, the cleaning and capture-average alternatives differ from the primary estimate by roughly two points.

That sensitivity is material enough to keep visible in the interpretation.

Again, Medium confidence applies to the AI-side measurement, not to future financial performance.

Limitations

The Goldman Sachs / Marcus observation has several important limitations.

1. The row combines parent and consumer-brand entities

Goldman Sachs Group and Marcus by Goldman Sachs are not one homogeneous commercial exposure.

2. The observation window is short

The primary comparison uses July and September 2026 endpoints. The signal may reverse, narrow, persist, or change materially in later periods.

3. The prompt universe is not a sample of Goldman Sachs clients

Commercially oriented prompts can reveal AI recommendation behavior without representing the distribution of institutional clients, wealth-management clients, Marcus consumers, or investors.

4. Sensitivity changes the estimated magnitude

The alternative no-dedupe and capture-average estimates are approximately two points less negative than the primary clean estimate.

5. Platform agreement does not prove causality

All six platforms moved lower, but that does not show that AI recommendation changes caused or forecasted any later commercial outcome.

6. AI systems can change independently of Goldman Sachs fundamentals

Model updates, retrieval systems, ranking behavior, index changes, answer policies, and product changes can alter recommendation coverage.

7. Recommendation coverage is not market share or financial performance

The metric does not directly measure deposits, investment banking wallet share, trading activity, assets under supervision, flows, revenue, earnings, or return on equity.

8. Already-public information may explain part of the movement

Q2 results, filings, market conditions, product changes, transactions, and other public information released during the observation window may influence both AI outputs and investor expectations.

Future validation must treat those data as baseline information.

What We Will Test Next

The Goldman Sachs / Marcus observation is now a frozen prospective AI-side signal.

Future testing should proceed in layers.

First, separate and test consumer-facing Marcus variables where possible:

1. Marcus branded search demand;
2. savings-account consideration;
3. site traffic and digital engagement;
4. account-opening proxies where reliable;
5. consumer deposit trends; and
6. product-level recommendation behavior.

Second, test parent-level Goldman Sachs commercial and financial outcomes:

7. investment banking fees;
8. advisory and underwriting activity;
9. FICC and equities revenue;
10. asset management flows;
11. assets under supervision;
12. wealth-management activity;
13. total net revenues;
14. earnings and return on equity;
15. analyst estimate revisions; and only then
16. later excess stock returns.

The preferred design remains a walk-forward framework in which conventional commercial, market, financial, and product variables enter the baseline model first.

AI variables should then be added to test whether they contribute incremental information.

A null result is a valid outcome.

If Goldman Sachs / Marcus recommendation coverage remains weak but conventional business measures stay strong, that divergence would help define what the AI metric does and does not capture.

If both move together over multiple periods, the next question would be whether the AI-side variable adds information beyond the conventional baseline.

Related LLM Authority Index Research

• Can AI Search Signal Future Revenue Growth? The AI Commercial Momentum Hypothesis
• Initial Findings From 25 Public Companies
• How We Measure AI Commercial Momentum
• AI Investor Signal Tracker
• AI Recommendations vs. Mentions vs. Citations
• Cross-Platform AI Recommendation Portability
• Bank Stocks and AI Search
• Fintech, Brokerage and Crypto Stocks
• Axos Financial AI Search Visibility
• Webull AI Search Visibility
• Citigroup AI Search Visibility
• Ally Financial AI Search Visibility
• American Express AI Search Visibility
• Bank of America AI Search Visibility
• Chime AI Search Visibility

The live public-company series is maintained in the AI Investor Signal Tracker.