Citation-Recommendation Coupling: A 114,596-Citation Study of Whether AI Recommendations Survive Source Turnover

The Finding in 60 Seconds

AI citation volatility is already documented. Recommendation volatility is increasingly being measured too.

This study asks the next question:

When an AI system changes the sources it cites, how much does its commercial recommendation change with them?

Across 114,596 observable AI citation events in two complementary research corpora, we built a longitudinal panel matching the exact same commercial prompt on the exact same AI platform across consecutive months.

The final panel contains 1,451 matched prompt-platform comparisons across 10 commercial verticals.

Among the 690 comparisons where both citation persistence and recommendation persistence were measurable, the relationship between them was:

Citation-Recommendation Coupling: ρ = 0.324

Citation stability was positively associated with recommendation stability.

But the two were far from interchangeable.

Most strikingly, when cited-domain overlap fell to exactly 0%, 80.5% of cases still retained at least one recommended brand.

And when a #1 recommendation existed in both periods, 55.1% retained the same #1 company despite sharing no cited domains from the prior month.

At the opposite extreme, when citation-domain sets were completely stable, 92.2% retained the same #1 recommendation.

Those findings point to a measurable relationship that citation counts and recommendation counts alone do not capture.

We call it Citation-Recommendation Coupling.

Further Reading:

• Review Gemini citation sources for high-intent buying questions for another view of how source choice can shape AI recommendations.
• Compare these findings with Perplexity citation sources for high-intent buying questions and its recommendation patterns.
• See how Grok citation sources for high-intent buying questions differ in the relationship between citations and recommendations.

What Is Actually New About This Research?

Answer Capsule: Citation churn, recommendation volatility, and differences between citations and recommendations are already known. The new contribution is measuring how citation persistence and recommendation persistence move together for the same prompt, platform, and adjacent time periods.

Questions This Section Answers

• What is new about Citation-Recommendation Coupling?
• Has AI citation volatility already been studied?
• Has AI recommendation persistence already been studied?
• What does this research add?

Citation volatility itself is not our novelty claim.

Prior AI Search research has measured citation retention over repeated collection waves and documented substantial citation turnover.

Recommendation persistence is not entirely unexplored either. Other researchers have begun tracking recommendation rankings over time and measuring how often leading recommendations remain stable.

The broader idea that citations, brand visibility, and recommendations are different signals is also increasingly established across AI Search research.

Our question is narrower:

Does the stability of the evidence surfaced by an AI system statistically correspond to the stability of the commercial decision it makes?

As of our review of publicly available research through September 2026, we did not identify a previous published study that combined all four of the following:

1. Matched the same commercial prompt on the same AI platform longitudinally.
2. Independently measured citation-domain persistence and recommendation-set persistence.
3. Calculated the statistical relationship between those variables.
4. Tested recommendation survival when citation overlap reached exactly zero.

We also did not identify prior published research that modeled whether citation persistence remained associated with #1 recommendation persistence after controlling for AI platform and commercial vertical.

Those are the novelty claims in this paper.

They are intentionally narrower than claiming that citation volatility, recommendation volatility, or separate visibility layers are new discoveries.

Finding 1: Citation Persistence and Recommendation Persistence Are Related, but Only Moderately

Answer Capsule: Across 690 matched observations where both variables could be measured, citation-domain persistence and recommendation-set persistence had a Spearman correlation of ρ = 0.324, p < 0.001.

Questions This Section Answers

• Do stable citations correspond to stable AI recommendations?
• How strongly are citation persistence and recommendation persistence related?
• Can citation persistence substitute for recommendation measurement?

We measured two separate longitudinal variables.

Citation Persistence

For the same prompt and platform in consecutive months:

Citation Persistence = Shared Cited Domains / Total Unique Cited Domains Across Both Periods

Recommendation Persistence

For that same observation:

Recommendation Persistence = Shared Valid Recommendations / Total Unique Valid Recommendations Across Both Periods

Both use Jaccard overlap.

A value of 1.0 means the sets were identical.

A value of 0.0 means they shared nothing.

Across 690 observations where both values were measurable:

Spearman ρ = 0.324

The association was statistically significant at p < 0.001.

This is neither a near-zero relationship nor a near-perfect one.

Higher citation stability was associated with greater recommendation stability.

But citation persistence did not provide enough information to infer recommendation persistence reliably on its own.

That is the first important result:

The observable evidence environment and commercial recommendation environment move together to a degree, but they are not the same longitudinal signal.

Finding 2: 80.5% of Commercial Answers Retained a Recommended Brand After Complete Citation Turnover

Answer Capsule: We found 303 matched commercial answers where citation-domain overlap was exactly 0% from one month to the next. Despite complete observable citation turnover, 244 of 303, or 80.5%, retained at least one recommended company.

Questions This Section Answers

• Can AI recommendations survive complete citation turnover?
• What happens when every previously cited domain disappears?
• Does losing all previous citation domains reset the recommendation set?

The most revealing part of the study occurs at the extreme.

We isolated cases where:

Citation Persistence = 0

That means not one cited root domain from the earlier observation appeared in the later observation.

There were 303 such cases where recommendation persistence was also measurable.

The results:

The recommendation list usually changed.

Only 9.6% retained the exact same set.

But the recommendation environment generally did not reset completely.

More than four out of five cases retained at least one previously recommended company.

That is different from merely observing citation churn.

It measures what happens to the commercial answer after the churn occurs.

Further Reading:

• Compare citation-recommendation coupling with frontier AI model citation patterns for high-intent buying questions.
• See how Claude citation sources for high-intent buying questions connect source selection with downstream recommendations.
• Explore OpenAI GPT citation sources for high-intent buying questions to compare how citation behavior aligns with commercial recommendations.

Finding 3: 55.1% Retained the Same #1 Recommendation Despite Zero Shared Citation Domains

Answer Capsule: Among zero-citation-overlap cases where a #1 recommendation existed in both months, 135 of 245, or 55.1%, retained the same #1 company.

Questions This Section Answers

• Can the #1 AI recommendation survive complete citation turnover?
• How often did the top recommended company remain unchanged after every citation domain changed?
• Is the top commercial recommendation more persistent than its observable evidence?

Of the 303 zero-citation-overlap cases, 245 contained a first-ranked recommendation in both measurement periods.

In 135 cases, the identity of the #1 company did not change.

Same #1 Recommendation: 55.1%

The 95% confidence interval is approximately 48.8% to 61.2%.

That means more than half of measurable top recommendations survived despite complete turnover in the publicly observable citation-domain set.

This does not mean the sources were irrelevant.

Displayed citations are not necessarily a complete record of everything involved in generating an AI answer.

It does establish something narrower:

The top commercial recommendation can display materially greater persistence than the set of domains cited alongside it.

Finding 4: Complete Citation Stability Produced Much Greater Recommendation Stability

Answer Capsule: Recommendations did not require stable citation sets to survive, but completely stable citation environments were associated with dramatically more stable recommendations. When citation-domain persistence was 100%, 92.2% of cases with a #1 recommendation in both periods retained the same #1 company.

Questions This Section Answers

• Do stable citations still matter?
• Are recommendations more stable when citation sets remain stable?
• What happens at 0% versus 100% citation persistence?

The zero-overlap result should not be interpreted as:

Citations do not matter.

The other end of the distribution makes that clear.

There were 224 measurable cases with 100% citation-domain persistence.

In other words, the citation-domain set was identical across consecutive observations.

Compare the extremes:

Among the 192 fully citation-stable cases with a #1 recommendation in both periods, 177 retained the same #1 company.

That produces a more precise interpretation:

Stable observable evidence is strongly associated with stable recommendations, but stable recommendations do not require stable observable evidence.

That is a very different conclusion from either "citations determine recommendations" or "citations do not matter."

Finding 5: Citation-Recommendation Coupling Is Not Constant Across Commercial Categories

Answer Capsule: The measured relationship between citation persistence and recommendation persistence varied substantially by vertical. Among adequately sampled categories, Spearman correlations ranged from approximately 0.039 to 0.461.

Questions This Section Answers

• Is Citation-Recommendation Coupling the same in every industry?
• Does citation persistence matter more in some commercial categories?
• Should Citation-Recommendation Coupling be benchmarked by vertical?

The overall ρ = 0.324 does not tell the entire story.

Selected categories produced:

Categories with very small usable samples are not emphasized here.

These results should not be interpreted as permanent characteristics of any industry.

The observation period is short, AI systems continue to change, and sample sizes are unequal.

But they establish an important research question:

Is Citation-Recommendation Coupling itself conditional on the commercial market being evaluated?

The difference between Window Replacement at 0.461 and PR Management Agencies at 0.039 suggests that the answer may not be universal.

Future monthly observations can test whether those differences persist.

Finding 6: Citation Persistence Remained Associated With #1 Recommendation Survival After Controlling for Platform, Vertical, and Time Period

Answer Capsule: Citation persistence remained positively associated with retaining the same #1 company after controlling for AI platform, commercial vertical, and measurement transition. Each 10-percentage-point increase in citation persistence corresponded to approximately 1.265 times the odds of retaining the same #1 recommendation, with a 95% confidence interval of 1.195 to 1.339.

Questions This Section Answers

• Is Citation-Recommendation Coupling merely an industry effect?
• Is it simply caused by differences among ChatGPT, Gemini, Copilot, and Perplexity?
• Does citation persistence remain associated with recommendation persistence after controls?
• Which outcome was used in the controlled model?

Correlation alone leaves an obvious alternative explanation.

Perhaps some industries simply have stable citations and stable recommendations.

Perhaps one AI platform happens to dominate both measurements.

Perhaps July-to-August behavior differs from August-to-September behavior.

We therefore ran a logistic model on matched observations where:

• citation persistence could be calculated;
• a #1 valid recommendation existed in both periods.

The dependent variable was binary:

Did the same company remain #1?

The model included controls for:

• AI platform;
• commercial vertical;
• measurement transition.

Repeated observations of the same prompt-platform combination were accounted for using clustered uncertainty estimates.

The final model contained:

n = 568

For every 0.10 increase in citation-domain persistence, the estimated odds of retaining the same #1 recommendation increased by:

1.265×

95% CI: 1.195× to 1.339×

p < 0.001

This controlled analysis applies specifically to same-#1 recommendation persistence.

It does not statistically control the broader 80.5% result measuring whether at least one recommended brand survived.

That distinction is important.

The controlled model strengthens the narrower conclusion:

The positive relationship between citation stability and top-recommendation stability cannot be explained solely by the platform, commercial vertical, or measurement transition represented in the observation.

It still does not establish causation.

We Propose Citation-Recommendation Coupling as a Distinct AI Search Measurement

Answer Capsule: Citation-Recommendation Coupling measures how closely changes in an AI system's observable citation environment correspond to changes in its commercial recommendation environment.

Questions This Section Answers

• What is Citation-Recommendation Coupling?
• How is Citation-Recommendation Coupling different from citation retention?
• Why measure the relationship instead of measuring citations alone?

Citation Retention asks:

Did the sources survive?

Recommendation Persistence asks:

Did the recommended companies survive?

Top Recommendation Persistence asks:

Did the winner survive?

Citation-Recommendation Coupling asks something different:

How strongly does source persistence move with recommendation persistence across matched commercial AI answers?

For this study:

CRC = Spearman(Citation Persistence, Recommendation Persistence)

The resulting overall coefficient was:

ρ = 0.324

The metric is useful precisely because the same amount of citation drift may have very different commercial implications.

A source disappearing does not automatically mean the brand disappears.

A recommendation persisting does not mean the same sources continue supporting the answer.

Measuring both outcomes and then measuring their relationship allows those cases to be distinguished.

Recommendation Survival Under Zero Citation Overlap Is a Separate Extreme-Case Metric

Answer Capsule: Citation-Recommendation Coupling summarizes the overall relationship. Zero-overlap survival measures what happens when the previously observed citation environment disappears completely.

Questions This Section Answers

• Why measure zero citation overlap separately?
• What does recommendation survival measure?
• Why is zero-overlap survival different from correlation?

Correlation describes the entire distribution.

Zero-overlap survival asks a much simpler question:

When none of last month's cited domains appear this month, what remains of last month's recommendation?

In this dataset:

• 80.5% retained at least one recommended brand.
• 55.1% retained the same #1 company when both periods had a #1.
• 9.6% retained the exact same recommendation set.

That combination is important.

The precise recommendation list usually changed.

But commercial continuity frequently remained.

The evidence environment was less persistent than the decision environment.

The 114,596-Citation Research Base

Answer Capsule: The broader research program contains 114,596 observable citation events across two complementary corpora. The Citation-Recommendation Coupling findings come from the longitudinal commercial corpus. The separate Consensus corpus provides additional standardized cross-model evidence research.

Questions This Section Answers

• Where does the 114,596 number come from?
• Were all 114,596 citations used in the correlation?
• How large is the longitudinal corpus?

The research base consists of:

The longitudinal commercial corpus contains:

• 17 commercial verticals
• 32 vertical-month datasets
• 14,233 AI observations
• 63,396 observable citation events

The underlying records separately preserve:

• cited URLs;
• cited root domains;
• company presence;
• recommendation validity;
• recommendation rank;
• ordered recommendation sets;
• buyer stage;
• prompt text.

The two corpora were not pooled statistically as though they were generated by one experiment.

The novel longitudinal analysis uses the commercial-monitoring corpus.

The standardized Consensus corpus remains a separate research design.

Methodology

Exact Same-Prompt Longitudinal Matching

The analysis matched observations only when they contained:

• the same normalized prompt text;
• the same AI platform;
• consecutive measurement months.

The primary longitudinal analysis included:

• ChatGPT
• Gemini
• Microsoft Copilot
• Perplexity

Google AI Mode and Google AI Overviews remained part of the larger citation corpus but were excluded from the root-domain persistence calculation because redirect and proxy behavior can complicate direct citation-domain matching.

Duplicate Resolution

Pre-publication quality control identified 13 prompt-platform-month keys in which the same normalized prompt and platform appeared more than once within a monthly source dataset.

Rather than select one response based on file order, all ambiguous duplicate keys were excluded from longitudinal matching.

Six would otherwise have entered an adjacent-month matched comparison.

After applying this rule, the final longitudinal panel contained:

1,451 Exact Matched Prompt-Platform Comparisons

This duplicate-exclusion rule was fixed before calculating the final publication statistics.

Persistence Measurement

For sets A and B:

Jaccard Persistence = |A ∩ B| / |A ∪ B|

Mention Persistence

Mean Jaccard overlap of companies marked present:

80.5% across 1,451 matched comparisons

Recommendation Persistence

Mean Jaccard overlap of valid recommended companies:

59.7% across 892 measurable comparisons

Citation Persistence

Mean Jaccard overlap of cited root domains:

37.9% across 1,078 measurable comparisons

Citation-Recommendation Coupling

Both measurements were available for:

690 comparisons

Spearman correlation:

ρ = 0.324

p < 0.001

Zero Citation Overlap Analysis

The analysis separately isolated cases where:

Citation Persistence = 0

There were:

303 measurable cases

Of those:

• 244 retained at least one recommended brand
• 29 retained the exact recommendation set
• 245 had a #1 recommendation in both periods
• 135 retained the same #1 recommendation

Controlled #1 Recommendation Analysis

The logistic model was restricted to cases with:

• measurable citation persistence;
• a #1 valid recommendation in both periods.

Outcome:

1 = same company remained #1

0 = different company became #1

Controls:

• platform;
• commercial vertical;
• measurement transition.

Clustered uncertainty was used for repeated prompt-platform observations.

n = 568

OR per +0.10 citation persistence = 1.265

95% CI = 1.195 to 1.339

p < 0.001

What This Study Does Not Prove

This research measures output relationships.

It does not expose model internals.

The study does not prove that:

• displayed citations caused a recommendation;
• changing a citation will cause the recommended brand to change;
• losing a citation has no commercial consequence;
• displayed sources are every source or signal involved in the response;
• the measured coupling coefficients are permanent platform characteristics;
• or short-term persistence predicts long-term persistence.

The appropriate conclusion is narrower:

Observable citation stability contains information about recommendation stability, but it does not fully determine recommendation stability.

Frequently Asked Questions About Citation-Recommendation Coupling

What is Citation-Recommendation Coupling?

Citation-Recommendation Coupling measures how closely citation-domain persistence and commercial recommendation-set persistence move together for matched AI answers over time.

What was Citation-Recommendation Coupling in this study?

Across 690 measurable matched observations, Spearman correlation was approximately 0.324.

Can AI recommendations survive complete citation turnover?

Yes. Among 303 cases with zero citation-domain overlap, 244, or 80.5%, retained at least one recommended brand.

Can the #1 AI recommendation survive when every cited domain changes?

Yes. Among 245 zero-overlap cases with a #1 recommendation in both periods, 135, or 55.1%, retained the same #1 company.

Are recommendations more stable when citations remain stable?

Yes in this dataset.

Among cases with completely stable citation domains and a #1 recommendation in both periods, 92.2% retained the same #1 company.

Does that prove citations determine recommendations?

No.

The relationship is observational.

Citation persistence remained associated with #1 persistence after controls, but the analysis does not establish causation.

What did the controlled analysis find?

Among 568 eligible observations, every 10-percentage-point increase in citation persistence was associated with 1.265 times the odds of retaining the same #1 company, with a 95% confidence interval of 1.195 to 1.339.

Does Citation-Recommendation Coupling vary by industry?

It did in this dataset.

Among adequately sampled verticals, measured Spearman correlations ranged from approximately 0.039 to 0.461.

Is this the first study of AI citation volatility?

No.

Citation volatility has already been studied.

Is this the first study of AI recommendation persistence?

No.

Recommendation persistence is also now being measured longitudinally.

What is the first-of-its-kind claim?

To our knowledge, this is the first published matched-prompt longitudinal study to directly measure the statistical relationship between citation persistence and commercial recommendation persistence, quantify recommendation survival at 0% citation-domain overlap, and test #1 recommendation persistence while controlling for platform and commercial vertical.

How many AI citation events are included in the broader research program?

The two complementary research corpora contain 114,596 observable AI citation events.

Were all 114,596 citations used to calculate Citation-Recommendation Coupling?

No.

The Citation-Recommendation Coupling analysis comes from the 63,396-citation longitudinal commercial-monitoring corpus.

The separate 51,200-citation standardized Consensus corpus is part of the broader LLM Authority Index research program but was not statistically pooled into the longitudinal calculation.

Conclusion: Citation Turnover and Recommendation Turnover Are Not the Same Event

We already know AI citations change.

We increasingly know AI recommendations change.

The unanswered question was:

Do they change together?

This study provides an initial measurement.

Across matched commercial AI answers:

Citation-Recommendation Coupling Was ρ = 0.324

Greater citation stability was associated with greater recommendation stability.

But recommendation continuity frequently survived dramatic changes in observable evidence.

When citation-domain overlap fell to zero:

80.5% retained at least one previously recommended company.

55.1% retained the same #1 company when both periods produced a #1 recommendation.

When the citation-domain set remained completely unchanged:

92.2% retained the same #1 company.

The evidence and recommendation environments therefore appear neither independent nor interchangeable.

The next question in AI Search measurement should not simply be:

Who cited the brand?

And it should not simply be:

Who recommended the brand?

It should also be:

How Tightly Are the Evidence and Recommendation Outcomes Coupled Over Time?

That relationship can now be measured.

We call it Citation-Recommendation Coupling.

Disclosure

LLM Authority Index and CiteWorks Studio share common ownership.

The longitudinal commercial-monitoring corpus originated from CiteWorks Studio measurement systems and was analyzed separately from the standardized Consensus corpus.

The two research designs were not statistically pooled as one experiment.

Novelty Statement

The first-of-its-kind characterization is intentionally narrow.

It refers specifically to:

• longitudinal measurement of Citation-Recommendation Coupling;
• recommendation survival under zero citation-domain overlap;
• and the controlled association between citation persistence and #1 recommendation persistence.

It does not claim first discovery of citation volatility, recommendation volatility, citation fragmentation, or the distinction between citations and recommendations.