Upstart AI Search Visibility: Recommendation Coverage Declined From July to September 2026

Research status: Exploratory longitudinal research. Upstart's AI recommendation momentum has not been validated as a predictor of loan originations, funded volume, borrower demand, partner activity, revenue, earnings, analyst revisions, valuation, or stock returns.

Observation window: July through September 2026

Ticker: UPST

Public parent: Upstart Holdings, Inc.

Tracked entity: Upstart

Exposure type: Direct/core brand

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

Current methodology version: V0

Answer Capsule

Upstart recorded a directional decline in AI recommendation coverage from July to September 2026 in the initial LLM Authority Index public-company panel.

Across 273 matched prompt-platform cells, Upstart recommendation coverage declined from 81.0% in July to 73.6% in September 2026, a change of -7.3 percentage points.

The exploratory 95% interval ranged from approximately -12.9 to -1.8 percentage points, remaining below zero. Four of six measured AI platform families declined:

• ChatGPT: +3.12 pp
• Gemini: -16.13 pp
• Google AI Mode: -14.71 pp
• Google AI Overviews: -4.55 pp
• Microsoft Copilot: +12.50 pp
• Perplexity: -27.27 pp

Upstart therefore meets the V0 Negative AI divergence candidate rules and receives a High AI-measurement confidence classification.

That classification refers only to the measured AI-side movement. It does not mean that Upstart loan originations, revenue, partner activity, earnings, or stock performance will decline.

The most important counterpoint is that simple presence barely changed. Upstart appeared in 91.6% of eligible matched responses in July and 90.8% in September, a decline of only 0.7 percentage points. Average recommendation rank also changed only modestly, moving from approximately 2.18 to 2.28.

The resulting measurement pattern is specific:

1. Upstart remained present in nearly all eligible matched responses;
2. it was recommended in a smaller share of those responses; and
3. when recommended, its average position was only slightly worse.

This makes Upstart another strong example of why AI recommendations, mentions, citations, presence, and rank should remain separate metrics.

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

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

Upstart AI Recommendation Momentum at a Glance

Additional V0 signal properties:

Upstart appears in both the Fintech, Brokerage and Crypto Stocks and Mortgage and Lending Stocks in AI Search sector analyses. That overlap reflects Upstart's role as a technology-enabled lending marketplace rather than a duplicate signal.

Questions This Section Answers

• How much did Upstart's AI recommendation coverage decline?
• Was the decline broad across AI platforms?
• Why can presence stay nearly flat while recommendation coverage falls?

How Much Did Upstart's AI Recommendation Coverage Decline?

Upstart recommendation coverage declined 7.3 percentage points, from 81.0% in July to 73.6% in September 2026.

The current V0 framework treats a decline of at least 5 percentage points as one component of the negative-candidate rule. Upstart also clears the uncertainty requirement because the exploratory interval remains below zero, from approximately -12.9 to -1.8 points.

This distinguishes Upstart from companies such as Travelers or Allstate, whose negative point estimates exceeded 5 percentage points but whose exploratory intervals crossed zero.

Upstart's point estimate is therefore directionally clearer within the current AI-side measurement system.

That does not make it a financial forecast. The interval describes uncertainty around the AI recommendation-change estimate, not uncertainty around future revenue, loan volume, earnings, or stock returns.

Was the Decline Broad Across AI Platforms?

The decline was broad enough to satisfy the V0 negative-candidate framework, but it was not universal.

Four platform families declined:

The largest negative movement came from Perplexity, while Microsoft Copilot moved strongly in the opposite direction.

That matters because aggregate recommendation momentum and cross-platform portability are separate variables. The project addresses this directly in Does Cross-Platform AI Visibility Matter?.

Upstart is not an all-platform decline like Happen / LendingClub. Instead, it is a directional aggregate decline supported by four of six platforms, with two meaningful counter-movements.

This difference should remain visible in any future predictive model. A -7.3-point aggregate change with four-platform support may contain different information from the same aggregate change produced by six-platform agreement.

Why Can Presence Stay Nearly Flat While Recommendation Coverage Falls?

Because presence and recommendation are not the same event.

Upstart's presence declined only 0.7 percentage points, from 91.6% to 90.8%, while recommendation coverage declined 7.3 points.

An AI answer can mention or discuss Upstart without recommending it as one of the products or companies that satisfies the user's commercial question.

That means the data can support all three of the following statements at once:

• Upstart remained highly visible in eligible responses.
• Upstart was recommended less frequently.
• Upstart's position when recommended was only slightly worse.

This distinction is important for investor-oriented measurement because a system based only on presence or mentions would miss much of the change observed in Upstart's recommendation behavior.

It also prevents the research from treating one number as a universal AI visibility score.

Upstart Compared With the Initial Lending Panel

The Mortgage and Lending Stocks research slice contains four mapped public parents, all of which currently meet the V0 negative-candidate rules.

Upstart sits in the middle of that initial group by aggregate recommendation change.

The similar direction does not make the companies economically interchangeable. PennyMac and UWM are mortgage businesses. Happen / LendingClub carries a legacy-entity mapping complication. Upstart operates a lending marketplace that connects consumers with bank and credit-union partners across personal loans, automotive loans, home equity lines of credit, and other credit products.

Upstart's own investor materials describe the company as an AI lending marketplace connecting consumers with banks and credit unions. Its Q2 2026 release also reported product-level activity across unsecured and secured lending. These details reinforce that future validation should focus on commercial variables appropriate to Upstart's marketplace model rather than borrowing outcome definitions from mortgage originators.

Official source: https://ir.upstart.com/news-releases/news-release-details/upstart-announces-second-quarter-2026-results

Upstart Compared With Fintech Peers

Upstart also appears in the broader Fintech, Brokerage and Crypto Stocks analysis.

That slice includes economically different companies such as Webull, Ally Financial, Chime, Coinbase, Happen / LendingClub, and Goldman Sachs / Marcus.

The relevant comparison is therefore not whether all of those companies share one business model. They do not.

The useful comparison is whether similar AI-side measurement patterns later have similar or different commercial implications across business models.

For example:

• Webull had a positive aggregate recommendation change but remained mixed.
• PennyMac had a directional negative recommendation change while average rank improved.
• Happen / LendingClub declined across all six platforms under a legacy-brand mapping.
• Chime is another future-linked negative-candidate peer in the fintech slice.

Those differences are exactly why the project is preserving company-level detail rather than replacing the panel with a single sector score.

Why the Upstart Result Could Matter for Future Investor Research

Upstart is useful for the broader AI Commercial Momentum research program because its commercial funnel is relatively measurable.

Potential future validation variables include:

• monthly originations;
• loan count;
• funded loan volume;
• branded search demand;
• direct web traffic;
• application starts;
• partner-bank and credit-union activity;
• unsecured versus secured product mix;
• fee revenue;
• contribution profit;
• revenue revisions by analysts;
• earnings surprise; and
• later excess stock returns, tested only after commercial relationships are evaluated.

Upstart also publishes monthly origination information. The company published August 2026 origination volume on September 3, 2026. Because that information was available before the September observation window closed, it must be treated as contemporaneously available baseline information in any future predictive test.

Official source: https://ir.upstart.com/news-and-events/news-releases

Likewise, Upstart released its second-quarter 2026 results on August 4, 2026. Those results are already inside the July-to-September signal period and cannot legitimately be treated as future outcomes predicted by the signal.

This timing discipline matters. A future backtest should ask whether the frozen September AI signal adds information after controlling for data that investors could already observe by September 30.

Questions This Section Answers

• Does a High AI-measurement confidence label mean Upstart fundamentals will weaken?
• Does the negative-candidate label mean the stock is overvalued or should be sold?
• What would falsify the usefulness of the Upstart signal?

What Does High Confidence Mean Here?

High confidence means the measured AI-side change meets the current V0 reliability rules.

For Upstart, those rules are satisfied because:

• the row has 273 matched cells, above the current 200-cell High-confidence threshold;
• all six platform families are represented;
• the exploratory interval remains below zero;
• no-dedupe sensitivity is approximately -0.58 percentage points; and
• capture-average sensitivity is approximately -0.90 points.

Both sensitivity values remain well inside the current 2-point High-confidence tolerance.

This does not mean there is high confidence about any financial outcome.

It means there is comparatively strong evidence, under the current V0 measurement process, that Upstart recommendation coverage was lower in September than in the matched July baseline.

Does Negative Candidate Mean Sell, Overvalued, or Financially Weak?

No.

The V0 Negative AI divergence candidate label is a research classification for unusual negative movement in AI recommendation coverage under predefined measurement rules.

It is not a buy, sell, hold, valuation, or credit recommendation.

The current research has not established that a negative AI recommendation change predicts:

• lower originations;
• lower revenue;
• weaker contribution profit;
• fewer bank partners;
• negative earnings revisions;
• valuation compression; or
• negative stock returns.

Those relationships must be tested prospectively.

What Would Falsify the Upstart Signal's Predictive Usefulness?

Several future results could weaken or falsify the hypothesis that this type of AI-side movement contains investor-useful information.

Examples include:

1. Upstart's AI recommendation decline fails to persist in later months.
2. The signal disappears when prompt mix or platform composition changes.
3. Recommendation momentum shows no incremental relationship with later originations or other commercial variables after controlling for already-public information.
4. Presence, rank, branded search, traffic, or direct commercial metrics explain later outcomes while recommendation coverage adds nothing.
5. Similar negative-candidate observations across companies fail to show reproducible commercial relationships out of sample.
6. Cross-platform fragmentation proves more important than the aggregate change itself, making Upstart's four-of-six pattern less informative than fully portable signals.

That is why the project also maintains a formal falsification framework.

Methodology

The Upstart result is derived from the V0 matched-panel framework described in How We Measure AI Commercial Momentum.

Matched denominator

The row contains 273 matched prompt-platform cells spanning 219 normalized prompt clusters.

A cell enters the comparison only when the same normalized commercial question and AI platform family can be compared across the baseline and September observation.

Recommendation coverage is calculated as the share of eligible matched cells in which Upstart is classified as a recommendation.

Presence is separate

Presence measures whether Upstart appears in the eligible response, not whether it is recommended.

That distinction is essential in this article because presence changed only -0.7 points, far less than the -7.3-point recommendation-coverage change.

Rank is conditional

Average rank is calculated only among responses where Upstart is recommended. The average moved from approximately 2.18 to 2.28.

Rank therefore answers a different question from coverage.

Platform breadth

The aggregate row spans ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, and Perplexity.

Four declined and two improved.

Exploratory interval

The current interval is based on variation in prompt-level matched changes. Prompt-level means are used to estimate the standard error, and a normal 1.96 interval is reported.

For Upstart, the resulting exploratory interval is approximately -12.9 to -1.8 percentage points.

This is not a causal confidence interval and should not be interpreted as one.

Sensitivity analysis

The no-dedupe sensitivity difference is approximately -0.58 percentage points. The capture-average sensitivity difference is approximately -0.90 points.

Both remain inside the V0 High-confidence tolerance.

Failure and duplicate handling

Known extraction failures, duplicate observations, and capture-level issues are handled under the standing V0 data-quality rules. The project preserves those QA issues rather than silently treating all source observations as equally reliable.

Limitations

The Upstart observation has several important limitations.

First, the observation window is short. Three months are not enough to establish persistence or predictive usefulness.

Second, the prompt universe is research-defined. It is designed to capture unbranded commercially relevant questions, but it is not a complete representation of every borrower, loan product, or customer journey.

Third, platform behavior differs materially. Microsoft Copilot improved 12.5 points while Perplexity declined 27.3 points. Aggregation can obscure that fragmentation.

Fourth, AI recommendation coverage does not directly measure loan demand, application conversion, funded volume, credit quality, pricing, or partner-bank economics.

Fifth, Upstart publishes relatively frequent commercial information, including monthly origination volume. Future studies must use strict information cutoffs so already-public operating data is not accidentally counted as a later outcome.

Sixth, the current V0 confidence system is a measurement-quality classification, not a probability that the signal predicts anything financially important.

Seventh, macroeconomic credit conditions, interest rates, lender risk appetite, underwriting changes, product expansion, partner additions, and company-specific announcements may affect both AI outputs and commercial outcomes.

What We Will Test Next

The frozen September 2026 Upstart observation should be compared prospectively with later data rather than rewritten after the fact.

The preferred validation ladder is:

1. AI persistence: Does the negative recommendation change persist in October and later months?
2. Demand indicators: Does branded search, direct traffic, or application interest change after the frozen signal?
3. Operating outcomes: Do later monthly originations, loan counts, or funded volume move in a way associated with earlier AI recommendation momentum?
4. Financial outcomes: Does the signal add incremental information about future revenue, contribution profit, or analyst estimate revisions after controlling for already-public operating data?
5. Market outcomes: Only after earlier links are evaluated should the research test later excess stock returns.

The formal prospective design is described in Can AI Search Visibility Predict Revenue Growth? A Longitudinal Validation Study and How to Backtest AI Search Signals.

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
• Does Cross-Platform AI Visibility Matter?
• Fintech, Brokerage and Crypto Stocks
• Mortgage and Lending Stocks in AI Search
• Webull AI Search Visibility
• PennyMac AI Search Visibility
• LendingClub and Happen AI Search Visibility
• Chime AI Search Visibility
• United Wholesale Mortgage AI Search Visibility