The 2026 AI Citation Authority Study: Which Websites Does AI Cite for High-Stakes Decisions?

278,499 visible citation events | 60,881 deduplicated AI responses | Six AI platform families | July-September 2026

AI citation authority is not one universal leaderboard.

In the LLM Authority Index high-stakes consumer-decision panel, the Top 100 cited domains accounted for 74.8% of domain-response appearances in credit, debt, banking and lending, compared with 44.9% in health and medical decisions. A source strategy built for financial services would therefore describe a substantially different citation environment from one built for healthcare products and providers.

The differences extend beyond concentration.

A synthesis of the existing category studies indicates that insurance had the largest approximate difference between monthly source persistence and cross-platform source agreement. Credit had the strongest cross-platform agreement among the four families. Health had the lowest persistence and portability, but not the largest gap.

Across the overall panel, just 15 domains reached the Top 100 on all six platforms, while 88 remained in the aggregate monthly Top 100 across all three months. The leading sources were persistent over time, but their prominence was not consistent across platforms.

LLM Authority Index examined 278,499 visible citation events from 60,881 deduplicated responses across 53 source vertical datasets to investigate these relationships. The study measures observable citation behavior, not hidden retrieval, clinical quality, financial suitability or the causal influence of a publisher on a recommendation.

The family-level synthesis draws on the credit, debt, banking and lending study, insurance study, investing, retirement, mortgage and financial study, and health and medical study.

Answer Capsule

What does AI citation authority look like for high-stakes consumer decisions?

In this July-September 2026 panel, AI citation sources differed substantially by decision category and platform. The Top 100 captured 74.8% of domain-response appearances in credit, debt, banking and lending, but 44.9% in health and medical decisions. Only 15 domains appeared in all six platform-specific Top 100 lists. At the aggregate level, median monthly Top 100 Jaccard similarity was 85.2%, compared with 36.1% between platform Top 100 lists, producing an approximately 49.1 percentage-point Persistence-Portability Gap for AI citation source sets.

An exploratory synthesis of the category studies places insurance's approximate gap at 51.2 percentage points, compared with 32.1 points for credit, debt, banking and lending. These family gaps are calculated from reported summary statistics, not a newly reconciled raw-data analysis. They suggest that concentration, persistence and portability should be reported separately rather than compressed into one measure of "AI authority."

Which High-Stakes Consumer Decisions Does This Study Cover?

This study focuses on consumer decisions involving money, insurance, financial services, healthcare products and providers. Its four analytical families are:

These are the study's decision-family groupings, not a claim to represent every financial or medical question. The source datasets emphasize consequential consumer evaluation and selection, while retaining some adjacent questions and overlapping vertical memberships.

The study uses YMYL, or Your Money or Your Life, as a description of this high-stakes context. It does not establish that every individual prompt belongs to the same purchase stage or carries the same degree of risk.

That distinction matters. Evaluating a hearing-aid product, selecting an online therapy service and seeking an explanation of a medical condition are related forms of information seeking, but they are not interchangeable research populations.

The six public platform families are Google AI Overviews, Google AI Mode, ChatGPT, Gemini, Perplexity and Microsoft Copilot. The archive preserves eight raw source-surface labels, with Google keyword and non-keyword lanes consolidated into the two Google AI Search families for public reporting.

How Do Concentration, Persistence and Portability Differ Across the Four Families?

Questions This Section Answers

• Which decision family has the most concentrated citation sources?
• Which family has the most persistent and most portable Top 100 source sets?
• Which family has the largest Persistence-Portability Gap?

Three measurements describe different properties of the citation environment.

Concentration measures how much of the observed source activity is captured by the leading domains.

Persistence measures how similar the leading source sets remain across months.

Portability measures how similar the leading source sets are across platforms. Here, portability is a symmetric source-set comparison, not a probability that a placement or ranking will transfer from one platform to another.

Family-level comparison

Analytical status: The concentration figures come from the existing studies. Persistence is summarized from their reported monthly overlaps; portability uses their reported median platform overlaps. The approximate gaps below subtract the displayed one-decimal component values. This is a synthesis of existing results, not a new raw-data replication.

Concentration uses each slice's own Top 100 and domain-response denominator. Persistence and portability use median pairwise Jaccard similarity. Family values are supported by Article 05: Credit, Article 06: Insurance, Article 07: Investing and related financial decisions, and Article 08: Health and medical. The overall values come from the flagship's original structural analysis.

Credit has the most concentrated and most portable source set

Credit, debt, banking and lending had the highest Top 100 concentration, at 74.8%, and the highest median platform overlap among the four families, at 44.9%.

Its approximate gap was the smallest, at 32.1 percentage points, because its relatively strong monthly persistence was accompanied by greater cross-platform agreement. A concentrated source environment did not automatically produce the largest persistence-portability difference.

The practical interpretation is that a recurring financial source portfolio can provide a useful starting point across platforms, while still requiring platform-specific validation.

Insurance has the largest reconstructed gap

Insurance combined 80.2% monthly persistence, the highest among the four families, with 29.0% platform portability. Subtracting those reported components produces an approximate 51.2-point gap.

The category also retained 87 domains in its monthly Top 100 across all three months, but only 16 domains reached the insurance Top 100 on all six platforms.

This is the clearest category example of a persistent aggregate source portfolio that should not be mistaken for a universal platform source list.

Investing and related financial decisions show lower persistence than insurance

The investing, retirement, mortgage and financial family recorded approximately 66.7% monthly persistence and 28.2% portability, producing a 38.5-point gap from the displayed values.

Seventy domains remained in its Top 100 across all three months, while 19 reached all six platform-specific Top 100 lists.

This family should not be treated as interchangeable with credit and banking merely because both concern financial decisions.

Health has the lowest persistence and portability, not the largest gap

Health and medical decisions had 44.9% Top 100 concentration, approximately 61.3% monthly persistence and 25.8% platform portability.

Its approximate 35.5-point gap was smaller than insurance's because both components were lower. Health's source environment was less concentrated, less persistent across the monthly portfolios and less similar across platforms in these summaries.

A smaller gap does not necessarily mean broader or more dependable visibility. Two low component values can produce a smaller difference than a high persistence value paired with modest portability.

What the family comparison establishes

The existing summaries show persistence exceeding portability in all four families. They also show meaningfully different combinations of the two components.

These are descriptive differences within this panel, not significance-tested estimates of permanent differences between industries. The family analysis does not establish that category concentration causes portability, that insurance will always have the largest gap, or that the aggregate gap should equal an average of the family gaps.

Commercial Action Matrix: What the Research Means for Publishers and CMOs

This research is designed for commercial use. The measurements describe observable citation behavior, but the purpose of the framework is to help publishers understand the commercial value of their AI visibility and help brands decide where to focus PR, earned media, content, partnerships and AI search optimization.

The commercial implication is not that every highly cited publisher should receive the same investment. The opportunity is to identify the sources that are persistent in the relevant decision category, portable across the AI platforms that matter, and closely connected to the brand mentions and recommendations a company is trying to influence.

Which Domains Reached the Top 100 on All Six Platforms?

Only 15 registrable domains appeared in every platform-specific Top 100 across Google AI Overviews, Google AI Mode, ChatGPT, Gemini, Perplexity and Microsoft Copilot. Their source roles were varied.

Across the six final platform Top 100 lists, the combined universe contained 246 distinct domains. The 15 appearing in every list represented 6.1% of that union. Meanwhile, 117 appeared in only one platform's Top 100, and another 36 appeared in exactly two.

These results require two qualifications.

First, the 15 are not the only domains cited at least once by every platform. They are the domains that cleared the Top 100 threshold on every platform. Being cited somewhere and ranking among a platform's leading sources are different conditions.

Second, the group is not composed exclusively of independent review publishers. It includes marketplaces, a credit bureau, broad media and specialist resources. Cross-platform prominence does not establish a common business model, editorial independence or equal usefulness for every advertiser.

Related analysis: Do AI Platforms Cite the Same Websites for High-Stakes Consumer Decisions?

What Is the Persistence-Portability Gap for AI Citation Source Sets?

Questions This Section Answers

• How is the Persistence-Portability Gap calculated?
• What does the aggregate 49.1-point result measure?
• Does using Jaccard on both dimensions make this a controlled comparison?

The Persistence-Portability Gap for AI citation source sets is the difference between temporal source-set similarity and cross-platform source-set similarity under a stated measurement protocol.

This study uses Top 100 registrable-domain sets and Jaccard similarity:

Jaccard similarity = shared domains / unique domains across both sets

Persistence = median pairwise Jaccard across monthly Top 100 sets

Portability = median pairwise Jaccard across platform Top 100 sets

Gap = persistence minus portability, in percentage points

The aggregate monthly Top 100 was highly persistent

The median monthly similarity was 85.2%, and 88 domains appeared in the monthly Top 100 throughout all three snapshots. This measures persistence of membership, not frozen rank positions.

Platform Top 100 lists were less similar

The following are selected comparisons reported in the flagship. The overall platform median summarizes all 15 platform pairs.

The reported median platform similarity was 36.1%.

The resulting aggregate comparison is:

85.2% persistence - 36.1% portability = approximately 49.1 percentage points

This is a study-specific pooled-portfolio gap, not a universal constant or an individual publisher's probability of retaining a citation.

The fact that the gap is approximately 49.1 points and overall Top 100 concentration is 49.1% is a numerical coincidence. They measure different properties and use different denominators.

The pooling distinction matters

Monthly Top 100 sets pool observations across platforms. Platform Top 100 sets pool observations across months.

Using the same overlap statistic makes the two components interpretable on the same scale. It does not eliminate differences in sample size, prompt composition or pooling.

For example, changes within individual platforms could partially offset one another in a stable aggregate monthly list. The pooled result therefore supports the statement:

The aggregate monthly source portfolios were more similar to one another than the pooled platform portfolios were to one another.

It does not independently establish how much each individual platform changed over time after holding prompts and sampling conditions constant.

Likewise, the 88-domain three-month intersection and 15-domain six-platform intersection should not be subtracted to create a gap score. They involve different numbers of sets and answer different questions.

Can the Leading Publishers Stay Stable While Individual Answers Change?

Yes. Aggregate source persistence and same-prompt citation stability operate at different levels.

The companion citation-drift analysis matched the same prompt and platform between July and September. Among 4,309 matched pairs with at least one cited domain in either observation, only 11.3% returned exactly the same cited-domain set. Mean Jaccard similarity was 26.9%.

A publisher can remain prominent across a large portfolio of questions while appearing in different individual answers each month.

This distinction separates three questions:

Domain-level persistence also does not establish URL-level persistence. Two answers can cite different pages from the same domain and still overlap at the registrable-domain level.

For marketing teams, a persistent publisher portfolio is a reason to investigate recurring sources. It is not a guarantee that a particular article, product comparison or brand recommendation will remain visible.

Which 25 Websites Were Most Frequently Cited in the Measured Panel?

Questions This Section Answers

• Which domains had the highest overall response citation coverage?
• What kinds of sources appear in the ranking?
• Does this table represent every high-stakes category or platform equally?

Composition warning: finance-oriented and Google-heavy pooled sample. This ranking reflects the study's category composition and unequal platform collection volumes. It is not an equally weighted average of decision families or platforms, a market-share estimate, or a universal ranking of financial and medical authority. The category studies illustrate the imbalance: the health analysis alone reports 4,160 Google AI Overviews responses and 4,158 Google AI Mode responses, compared with 1,116 ChatGPT responses and 858 Perplexity responses. See the Health and Medical source study.

Response citation coverage is the percentage of the 60,881 eligible deduplicated responses citing a domain at least once. It is not that domain's share of all citation events. Multiple domains can appear in one response, so coverage percentages are non-exclusive.

The source-role labels below are coarse editorial descriptors. They are not a completed URL-level classification or an audit of editorial independence.

Counts, ranks and reported percentages are retained from the flagship's original ranking table. Its source-role discussion distinguishes comparison resources, business media, specialized health resources, communities and platforms.

The source reports Credit Karma coverage as 1.95%. Its displayed count of 1,184 divided by the stated denominator of 60,881 rounds to 1.94%. The discrepancy is flagged rather than silently reconciled.

Platform breadth, vertical breadth and monthly movement

The following companion table preserves the breadth and monthly rank measurements for the same 25 domains.

"Platforms citing" means the number of platform families that cited the domain somewhere in the measured panel. It does not mean the number of platform Top 100 lists containing the domain.

These values describe the original flagship panel, including overlapping source-vertical memberships.

Reddit illustrates why membership and rank movement should remain separate. It remained highly visible in the pooled ranking, but moved from #3 in July to #5 in August and #7 in September. That movement does not, by itself, quantify a change in absolute citation frequency. See the separate Reddit longitudinal analysis.

Why source role matters

A comparison publisher, community platform and provider-owned website are not interchangeable outreach opportunities.

Source role also differs from a citation's relationship to the evaluated company. A provider-owned domain can supply first-party evidence about itself and external evidence in an answer discussing another company.

The ranking should therefore be used as a source map, not a media-buy list. Citation frequency alone does not establish accuracy, independence, endorsement or a causal effect on brand recommendations.

Why Does This Ranking Differ From Broad Ahrefs Citation Rankings?

Questions This Section Answers

• Why do broad ChatGPT rankings and this high-stakes panel have different leaders?
• Can their percentages be compared directly?
• Do those differences prove that high-stakes intent causes a different source mix?

Ahrefs' September 2026 ChatGPT ranking reports Reddit first and Wikipedia second across a broad U.S. query population. The LLMAI panel ranks NerdWallet first and Forbes second across its high-stakes, six-platform population. These are different measurement designs, not conflicting answers to an identical question.

Ahrefs defines its percentage as mention share among the top sources. LLMAI uses response citation coverage. The percentages are not interchangeable.

The different leaders support a practical warning: a broad citation leaderboard should not be assumed to describe a specific commercial decision environment.

They do not isolate intent as the cause. Platform mix, dates, sampling, source normalization and denominators differ as well. A more controlled comparison would align platform and observation period before evaluating the remaining differences.

How Does Commercial Health Research Differ From Clinical-Information Research?

BrightEdge's health-and-finance research found that ChatGPT's leading healthcare sources were government and hospital domains, including NIH, MedlinePlus, Mayo Clinic, CDC and Cleveland Clinic. Its healthcare population included diseases, treatments, symptoms and mental health.

LLMAI's health category emphasizes consumer evaluation of products, providers and services. In its ChatGPT slice, the reported leading sources were Healthline at 21.1% response coverage, Forbes at 14.8% and Reddit at 14.0%, as reported in the Health and Medical category study.

The contrast is consistent with different information needs.

A clinical explanation and a comparison of healthcare providers may require different evidence. However, the studies also differ in collection and sampling, so their source lists do not independently prove an intent effect.

The appropriate conclusion is narrower:

A source map for medical information seeking should not automatically be used as a source map for healthcare product and provider selection.

Neither comparison ranks clinical truth or establishes that frequently cited commercial sources are medically superior to institutional sources.

Is the Citation Distribution More Concentrated Than Other Benchmarks?

Within the flagship, the Top 100 accounted for 49.1% of domain-response appearances. Separately, when domains were ranked by citation-event volume, the Top 100 accounted for 53.0% of citation events after the stated public-table exclusions. These are different concentration measures.

For context, the Machine Relations concentration study reported a Top 100 share of 18.56% across its domain-run citation distribution, using 15,782 observed answer runs and 22,179 cited domains over 125 days. It counted a domain once within an observed run for that measure.

LLMAI's 49.1% figure is numerically about 2.6 times that benchmark. It is not evidence that high-stakes intent alone produces 2.6 times the concentration. The query baskets, platform combinations, periods and sampling designs differ.

The strongest within-study evidence is the family comparison: the same broad research program reports Top 100 concentration ranging from 44.9% to 74.8%. That variation makes a universal concentration assumption unsuitable for planning.

A concentrated Top 100 also does not make the long tail irrelevant. The flagship contained 17,477 registrable cited domains before publisher-specific source-type filtering.

What Has Earlier Research Already Established?

Citation persistence, source decay and cross-platform disagreement are established areas of research. This study builds on them rather than claiming those subjects began with the Persistence-Portability Gap.

Short citation half-lives and persistent domain leaderboards are not necessarily contradictory. A cohort's activity, a page's citation share and the membership of the leading domain portfolio are different quantities.

The contribution here is the combination of a defined high-stakes consumer panel, explicit response-level accounting, separate temporal and platform source-set measurements, and an exploratory comparison across four decision families.

The framework names a measurable contrast. The category results determine what that contrast means in this dataset.

What Are the Implications for CMOs and AI Search Teams?

Questions This Section Answers

• Should brands use one publisher list across every AI platform?
• How should the four family patterns affect source research?
• Does citation persistence justify assuming future visibility?

The results support category-specific and platform-specific source research rather than one universal publisher list.

The following are planning implications derived from the descriptive patterns, not tested interventions or promises of improved recommendations.

Persistence provides historical evidence, not a guarantee. A source can remain prominent while its cited pages, associated companies or recommendation context change.

The next analytical step is to connect source visibility with company mentions, recommendation frequency, ranking, sentiment and source-company co-occurrence. Those relationships must be measured directly rather than inferred from citation rank.

What Can Publishers Responsibly Claim From This Research?

A publisher can use the study to document observable visibility in a defined panel.

A defensible claim identifies the metric and scope: response coverage, platform breadth, category strength or persistence across the measured snapshots.

A broad publisher might emphasize multi-category reach. A specialist might emphasize a narrow decision environment. A publisher that reaches every platform Top 100 can describe that achievement without claiming equal importance on every platform or for every brand.

What the research does not support is a promise that advertising, sponsorship or editorial inclusion on a cited website will cause an AI system to recommend an advertiser. The study is observational and does not measure that intervention.

How Was the Study Conducted?

Source archive and analytical population

The research uses stage-0 citation extraction datasets from the LLM Authority Index pipeline. The preserved records include citation arrays, prompt metadata, platform-surface labels, company extraction fields and provenance.

The primary counts and the separately applied public-table infrastructure exclusions should be distinguished when reproducing a particular table.

Observation and citation units

An observation is an extracted AI response associated with a prompt, month and raw source surface.

A citation event is a URL entry in that observation's citation array after the corpus-level observation rules.

A domain-response appearance counts a registrable domain at most once in an observation. Three URLs from one publisher therefore contribute three citation events but one domain-response appearance.

Failure handling

Observations were flagged as explicit extraction failures when their extraction notes or company-level exclusion reasons contained "extraction failed."

The 1,278 flagged observations were excluded from valid-response analysis rather than treated as ordinary responses with no citations. The source reports that those failed observations contained no extracted citations.

Repeated-record handling

The corpus includes overlapping vertical exports. An observation was treated as an exact repeated analytical record when its report month, exact raw source surface, normalized prompt and exact set of original citation URLs matched.

One analytical instance was retained for the overall structural analysis. This is an anti-double-counting rule, not proof that the upstream collector executed the prompt only once.

Prompt normalization used Unicode NFKC normalization, case folding, whitespace collapse and trimming, while otherwise preserving wording.

Domain normalization and source exclusions

URLs were consolidated to registrable domains using the Public Suffix List. For example, finance.yahoo.com was consolidated to yahoo.com. Original URLs and hosts remained available in the archive.

Obvious navigation and infrastructure artifacts were excluded from the public ranking. Substantive platform-owned surfaces were not automatically removed solely because a platform owned the domain. Their interpretation remains distinct from independent editorial sources.

Ranking construction

The primary ranking uses domain-response appearances. With a common response denominator inside a slice, this produces the same ordering as response citation coverage.

Monthly and platform Top 100 sets were constructed independently. The flagship did not introduce an additional minimum number of appearances, dates or platforms before a domain could enter the Top 100. The threshold was relative rank within the measured slice.

Family-gap synthesis

The family table summarizes the existing category analyses, using the median of three reported monthly Jaccard values and the reported median of 15 platform-pair values.

For consistent display, persistence and portability are presented to one decimal place, and the approximate gap subtracts those displayed values. An unrounded recomputation may change the final decimal place.

The synthesis does not rebuild source sets, impose a new prompt-matching rule, reconcile every parent/child denominator or estimate confidence intervals. It should therefore be treated as an exploratory extension of the existing reports.

Missing periods and overlapping categories

Some verticals have no July dataset. Missing periods are unavailable observations, not zero-citation responses or extraction failures. The monthly portfolios are therefore not a perfectly balanced longitudinal panel.

Source labels also contain topical overlap and adjacent prompts. The four families should not be interpreted as four statistically independent replications.

What Would a More Controlled Follow-Up Test?

A controlled extension would construct one Top 100 source set for each platform x month x decision-family combination.

With six platforms and three months, that creates 18 sets per family.

Within-platform persistence would compare months while keeping the platform fixed. Using all three month pairs across six platforms produces 18 temporal comparisons per family.

Within-month portability would compare platforms while keeping the month fixed. Fifteen platform pairs across three months produce 45 platform comparisons per family.

A stronger protocol would also align prompt populations, document sample sizes and missingness, test multiple ranking depths, and preserve prompt relationships when estimating uncertainty.

Those controlled results are not reported here. They would test whether the family patterns remain after reducing the pooling and composition effects present in the current descriptive comparison.

What Are the Main Limitations?

The population is selected. This is a deliberately high-stakes consumer-decision panel, not a representative sample of all AI use, all financial questions or all medical information seeking.

The sampling is unequal and partly unbalanced over time. Platform volumes differ, some verticals are missing July data, and related source datasets overlap.

The measurement is visible citation behavior. It does not capture every retrieval step or identify the full evidence used to generate an answer. The original stage-0 files do not contain a complete raw-response snapshot for every external interaction, which limits extraction auditing without additional collector logs.

Source-set overlap is not a complete authority measure. Jaccard does not measure changes in rank, citation frequency or the particular pages cited. Domain-level aggregation can hide different content and business roles within one website.

The family gaps are exploratory. They are based on reported summaries and have not been validated here with a reconciled raw-data rebuild or uncertainty analysis.

The study is not causal. It does not show that citations produce recommendations, that concentration produces portability, or that high-stakes intent alone explains differences from external benchmarks.

Commercial Relationship Disclosure

LLM Authority Index is part of a business ecosystem that provides AI visibility, research and marketing services. Affiliated businesses may have current or historical commercial relationships with companies or publishers appearing in the dataset.

Commercial relationships are not inputs to the ranking methodology. Domains are included based on observed citation data under the stated rules. Neither a ranking position nor a commercial relationship establishes endorsement or evidence that the relationship caused citation visibility.

Data Availability and Reproducibility

The supplied flagship identifies downloadable CSV and JSON assets as planned, but does not provide finalized public download locations. This edition should therefore not be described as a fully public, independently rerunnable data release.

Reproducing the concentration and gap measurements requires more than the headline table. It requires the relevant source-set memberships, response denominators, slice definitions, normalization and exclusion rules, and a versioned record of how each ranking was generated.

The family results should remain labeled as a summary-based synthesis until that reconstruction and the outstanding consistency checks are complete.

Conclusion: Category, Platform and Time Describe Different Dimensions of Citation Authority

The central finding is not simply that one publisher ranks above another.

High-stakes consumer decisions have different citation-source distributions, and those distributions differ across platforms and time.

Credit, debt, banking and lending had the highest Top 100 concentration, at 74.8%. Health and medical decisions had the lowest, at 44.9%. The existing category summaries place insurance at the largest approximate persistence-portability gap, while health is lower on both components.

Only 15 domains reached every platform Top 100, despite 88 remaining in the aggregate monthly Top 100 across all three snapshots. That contrast makes persistence and portability worth measuring separately.

For publishers, the useful output is a profile of category strength, coverage, breadth and persistence.

For CMOs, the useful starting point is the exact consumer decision and target platform, not a general list of the internet's most prominent websites.

For researchers, the next test is whether the observed family patterns survive a controlled, reproducible platform-by-month analysis.

A source can remain prominent over time without being equally prominent across platforms. Understanding where that happens is more useful than assuming there is one universal form of AI citation authority.

References and Related Research

LLM Authority Index source studies

• Article 05: Credit, Debt, Banking and Lending Source Study
• Article 06: Insurance Source Study
• Article 07: Investing, Retirement, Mortgage and Financial Source Study
• Article 08: Health and Medical Source Study
• Do AI Platforms Cite the Same Websites?
• How Stable Are AI Citations?
• Reddit's Decline in AI Citations
• Article 37: Perplexity Investing and Related Financial Source Study
• Top 100 Websites Cited by Google AI Overviews
• Top 100 Websites Cited by Google AI Mode
• Top 100 Websites Cited by ChatGPT

External research

• Ahrefs: Most-Cited Websites in ChatGPT
• BrightEdge: What Health and Finance Citations Reveal About ChatGPT and Google AI Overviews
• Machine Relations: Citation Concentration
• Tinuiti: AI Citation Trends Report
• Foglift: AI Search Citation Benchmark Q2 2026
• Wellows: AI Citation Overlap Study
• AirOps: Citation Persistence
• Stacker and Scrunch: Half-Life of AI Citations
• Writesonic: Citation Decay

---