How Stable Are AI Citations for High-Stakes Consumer Decisions? Measuring Citation Drift From July to September 2026

LLM Authority Index Research | High-Stakes Consumer Decisions | July-September 2026

A publisher can remain near the top of the AI citation leaderboard while the actual sources attached to an individual consumer question change substantially from one month to another.

That is the central finding of this longitudinal analysis.

Across the LLM Authority Index high-stakes consumer decision panel, the aggregate source ecosystem was unusually persistent. Eighty-eight domains appeared in the overall Top 100 in July, August, and September 2026. Pairwise Top 100 overlap ranged from 83.5% to 88.7% by Jaccard similarity.

But when we narrowed the analysis to the same prompt on the same source surface in July and September, the picture changed. Among 4,309 matched prompt/platform pairs with at least one cited domain in either month, only 11.3% returned the exact same cited-domain set. The mean Jaccard similarity of those source sets was 26.9%, equivalent to 73.1% observed citation-set turnover under the turnover definition used in this article.

In other words:

The leading publishers were relatively stable as a portfolio, while the route from an individual buyer question to a particular set of cited sources was much more volatile.

That distinction matters to publishers selling AI-search visibility, CMOs deciding where to invest in earned media, agencies building citation strategies, and brands evaluating whether a one-time AI visibility audit is still useful several months later.

This study is part of the 2026 AI Citation Authority Study, which analyzes 278,499 visible citation events from 60,881 deduplicated AI responses across six AI platform families and 53 high-stakes consumer vertical datasets.

Answer Capsule

How stable are AI citations over time for high-stakes consumer decisions? The aggregate source leaderboard was highly stable from July through September 2026: 88 of the overall Top 100 domains appeared in all three monthly Top 100 lists, and 91 of 100 July leaders were still in the Top 100 in September. However, citation stability was much lower at the individual-query level. In 4,309 non-empty matched July-to-September prompt/platform pairs, only 11.3% returned the exact same cited-domain set, the mean cited-domain Jaccard similarity was 26.9%, and 32.6% shared no cited domain at all. The data suggests that publisher authority can be persistent at the portfolio level while the citations attached to a specific buyer question remain highly dynamic.

Key Findings

• 88 of the Top 100 domains appeared in the Top 100 in July, August, and September 2026.
• July and September shared 91 of 100 Top 100 domains, producing 83.5% Jaccard similarity.
• August and September were even more similar, sharing 94 of 100 domains, or 88.7% Jaccard similarity.
• The result remained strong after restricting the analysis to the 45 source verticals represented in all three months. In that balanced sensitivity panel, 88 domains still appeared in all three Top 100 lists.
• Among the 88 persistent Top 100 domains, the median absolute July-to-September rank movement was six positions.
• 65 of the 88 persistent domains remained within 10 ranking positions of their July rank by September.
• The July-to-September rank correlation among shared Top 100 domains was 0.883 by Spearman correlation, indicating strong aggregate ordering persistence.
• The Top 10 was especially stable. Nine domains appeared in the Top 10 in all three months, and the July and August Top 10 sets were identical.
• Citation stability was much lower when the same query was examined directly over time. The strict July-to-September matched analysis contained 6,601 same-prompt, same-source-surface pairs.
• Of those matched pairs, 4,309 had at least one cited domain in July or September. Only 11.3% of those non-empty pairs returned the exact same cited-domain set.
• The mean July-to-September same-prompt Jaccard similarity was 26.9%, corresponding to 73.1% citation-set turnover.
• 32.6% of the non-empty matched July-to-September pairs shared no cited domain at all.
• Same-prompt citation turnover was substantial in all four high-stakes consumer decision families, ranging from 68.8% in Health and Medical to 76.1% in Investing, Retirement, Mortgage and Financial Decisions.
• In this measured panel, the two Google search surfaces showed more same-prompt source overlap from July to September than the chat-oriented platforms. This is descriptive of this collection system and should not be interpreted as a universal property of the platforms.

Why This Research Belongs in Daily AI Search Marketing Decisions

This study should be part of normal AI search marketing decision-making because it separates durable source authority from temporary answer-level visibility.

At the portfolio level, the citation market looked stable. The median Jaccard similarity across the three monthly Top 100 source lists was 85.2%, and 88 domains remained in the Top 100 across July, August and September.

At the same-prompt level, the market was much more volatile. Among 4,309 non-empty July-to-September matched prompt/platform pairs, only 11.3% returned the exact same cited-domain set. Mean Jaccard similarity was 26.9%, equivalent to 73.1% observed citation-set turnover under this study's definition.

That distinction has direct commercial value. A publisher can be part of a durable authority layer while the specific buyer questions citing that publisher change substantially. A brand can correctly identify important publishers in its category and still need recurring prompt-level monitoring to see whether those sources remain attached to its most valuable commercial questions.

This article supplies the time dimension behind the Persistence-Portability Gap. Article 03 measures how source authority differs across platforms. Article 04 shows why persistence must also be interpreted at more than one level: aggregate publisher portfolios can remain stable while individual query-level source sets rotate.

Commercial Action Matrix: What Publishers and CMOs Can Do With the Findings

The commercial operating model is therefore two-layered:

1. Strategic source layer: maintain the publishers that repeatedly matter across months, relevant platforms and the brand's decision category.
2. Operational query layer: repeatedly test the high-intent prompts that drive actual buyer decisions and watch which sources enter, leave or persist.

This research supports recurring measurement. It does not support the claim that a citation, publisher relationship or content placement will cause an AI system to recommend a brand.

How This Study Fits the High-Stakes Consumer Decisions Research Corpus

This article is the primary longitudinal citation-drift study in the LLM Authority Index High-Stakes Consumer Decisions research series.

• The 2026 AI Citation Authority Study establishes the overall source layer, concentration patterns and the combined persistence-versus-portability framework.
• Do AI Platforms Cite the Same Websites? measures the cross-platform portability side of the framework.
• The Persistence-Portability Gap defines how time persistence and cross-platform source-set agreement should be reported separately.
• Reddit's Decline in AI Citations shows how one major source can decline materially even while the broader aggregate source portfolio remains persistent.
• Articles 05 through 08 break persistence, concentration and source structure into the four high-stakes decision families.
• Articles 09 through 14 provide platform-specific source maps that should be used alongside this longitudinal analysis.

The studies answer different operating questions. Article 03 asks where source authority transfers across platforms. Article 04 asks how much source authority persists over time. The category and platform studies show where that authority is concentrated. Together, they provide a more useful commercial planning system than one blended AI visibility score.

Questions This Study Answers

• How long can a CMO rely on an AI citation baseline before it needs to be measured again?
• Can a publisher use one strong AI citation snapshot as evidence of durable source visibility?
• Which publishers remain visible in AI answers month after month?
• Does a stable Top 100 publisher leaderboard mean individual AI answers cite the same websites every month?
• How quickly do the source sets behind high-intent consumer questions change?
• Should brands continuously monitor the publishers cited for their most commercially valuable prompts?
• Does citation drift differ across Google AI Overviews, Google AI Mode, ChatGPT, Gemini, Perplexity, and Microsoft Copilot?
• Are finance, insurance, and health citation ecosystems all affected by source turnover?
• Can a publisher demonstrate persistent AI visibility to advertisers without claiming that every individual query will keep citing it?
• How should agencies distinguish durable publisher authority from temporary query-level citation visibility?

The Key Distinction: Portfolio Stability Is Not Query Stability

Citation stability can be measured at several levels, and they should not be treated as interchangeable.

A domain can remain one of the most frequently cited websites in an entire research panel while disappearing from one prompt and appearing on another. Likewise, an AI platform can maintain a similar overall concentration of sources while substantially changing the websites attached to individual questions.

For this study, we therefore separate two different questions.

Portfolio-level stability

Are the same domains still among the most frequently cited sources across the entire high-stakes consumer panel?

For this question, the answer was largely yes.

Same-prompt stability

When the same source surface encounters the same normalized prompt in a later month, does it cite the same set of domains?

For this question, the answer was much less stable.

This distinction is commercially important. A publisher may have durable value as part of the broader AI source ecosystem without being guaranteed visibility for any single question at any single point in time.

Overall Top 100 Citation Authority Was Highly Persistent

The monthly Top 100 rankings were constructed using response-level domain presence. A domain counts once per eligible AI response even if several URLs from that domain appear in the same answer.

Top 100 overlap by month

Across all three months, 88 domains remained in the Top 100.

That level of persistence is notable because the broader study includes dozens of commercial verticals and six AI platform families. It means that, at the ecosystem level, the identity of the highest-visibility citation sources did not reset every month.

The Stability Result Survived a Balanced-Vertical Test

The full monthly source files are not perfectly balanced. Some vertical datasets were not produced in July, so September contains broader source-vertical coverage than July.

To test whether the high Top 100 persistence was simply an artifact of changing vertical coverage, we repeated the monthly Top 100 comparison using only the 45 source verticals represented in July, August, and September.

Balanced 45-vertical sensitivity analysis

The balanced panel again produced 88 domains present in the Top 100 in all three months.

That does not eliminate every possible sampling difference, but it makes the central portfolio-stability finding harder to explain as merely a consequence of adding verticals later in the study.

Stability Was Strongest Near the Top of the Citation Ecosystem

The persistent-source pattern is visible at several ranking depths.

The Top 10 was particularly durable. July and August contained the same ten domains, and nine of those remained in the September Top 10.

This does not mean rank positions never changed. It means the group of dominant citation sources remained relatively recognizable across the study period.

Several Citation Leaders Barely Moved

Among the domains that remained in the Top 100 throughout the study, many of the largest publishers and financial information sites showed limited rank movement.

NerdWallet remained #1 in all three months, while Forbes remained #2. CNBC, Bankrate, Money.com, Investopedia, U.S. News, YouTube, The Wall Street Journal, Yahoo, Experian, and LendingTree also stayed near their original positions.

Among the 88 domains that remained in the Top 100 for all three months:

• the median absolute July-to-September rank movement was 6 positions;
• the mean absolute movement was 9.4 positions;
• 41 of 88 moved no more than five positions;
• 65 of 88 moved no more than ten positions; and
• 10 of 88 moved at least 20 positions.

The July-to-September Spearman rank correlation among the shared Top 100 domains was 0.883. August-to-September was even higher at 0.954.

These rank correlations describe ordering persistence among domains that were present in both lists. They do not imply that all individual prompts preserved their citation sources.

The Same Prompt Told a Much More Volatile Story

Aggregate rankings can hide source turnover inside individual answers.

To test this, LLM Authority Index built a strict repeated-query comparison. A July observation and a later observation were paired only when they shared:

1. the same raw source surface;
2. the same normalized prompt text; and
3. exactly one eligible analytical observation in each compared month.

Ambiguous keys with multiple distinct analytical records in either month were excluded from the strict pair set.

For July to September, that produced 6,601 matched prompt/platform pairs. Among those, 4,309 pairs had at least one cited domain in July or September.

July-to-September same-prompt citation stability

For this article, citation-set turnover = 1 - mean Jaccard similarity among matched pairs with at least one cited domain in either period.

That means the aggregate source universe can remain stable even when most of the specific source set attached to a repeated buyer question changes.

Citation Drift Increased Over the Wider July-to-September Interval

The same-prompt comparisons were also calculated for the adjacent month pairs.

The periods are not perfectly interchangeable. Model versions, retrieval indexes, query availability, and collection behavior may have changed between months. The different pair counts also mean these are not three measurements of one identical balanced population.

Still, the July-to-September comparison provides the widest interval in this three-month study and shows substantial source-set turnover despite high aggregate Top 100 persistence.

Citation Drift Appeared Across All Four High-Stakes Consumer Decision Families

The repeated-prompt result was not limited to one broad category.

For the following comparison, a matched prompt was assigned to a family only when its July and September source records carried the same proposed family classification. Empty-to-empty pairs were excluded.

Every family showed substantial same-prompt turnover.

The narrowest average turnover was still 68.8% in Health and Medical. The highest was 76.1% in Investing, Retirement, Mortgage and Financial Decisions.

This section answers

• Can a publisher's AI citation visibility be assumed to persist simply because the category is regulated or high stakes?
• Are finance citations more stable than health citations?
• Should publishers in insurance, lending, health, and investing all monitor AI citation movement over time?

The data supports the same operational conclusion across all four families: a historical citation map should not be treated as a permanent map of the category.

Citation Drift Differed Materially by AI Platform

The July-to-September matched analysis also varied by platform family.

In this corpus, the two captured Google search surfaces were much more likely to retain at least one domain between July and September than the other platform families.

That should be interpreted cautiously.

The platforms do not expose citations in identical ways, and the source collection paths are not technically identical. Google keyword and non-keyword collection lanes are also consolidated into public platform families for editorial reporting. The table therefore describes observed source-set persistence in this research pipeline, not an intrinsic or permanent stability ranking of the AI products themselves.

The broader lesson is safer: citation drift is platform-specific, just as citation authority itself is platform-specific.

For the cross-platform source comparison, see Do AI Platforms Cite the Same Websites for High-Stakes Consumer Decisions?.

What Citation Drift Means for Publishers

A publisher that ranks highly in AI citations can make a stronger commercial case when that visibility persists over time.

For example, there is an important difference between saying:

"We were cited frequently in one AI snapshot."

and saying:

"Our domain remained among the most frequently cited sources across multiple monthly measurements of high-stakes consumer decisions."

The second statement describes persistence rather than a single observation.

But publishers should also be careful not to overextend the claim.

A persistent aggregate ranking does not mean a publisher will continue appearing for every individual query. The same-prompt analysis shows why.

Questions publishers can use this study to answer

• Has our AI citation visibility persisted for multiple months?
• Are we consistently visible across a category, or are a small number of prompts responsible for our position?
• Which AI platforms show the strongest persistence for our domain?
• Can we show advertisers sustained AI citation visibility rather than a one-time screenshot?
• Are we losing visibility even though our historical AI citation rank still looks strong?

A defensible publisher sales claim is based on measured persistence, platform breadth, and category coverage. It should not promise that publishing with the site will cause an AI system to cite or recommend a client.

What Citation Drift Means for CMOs and Brands

For CMOs, the practical consequence is that a one-time publisher map can become stale even when the overall market leaders remain recognizable.

A brand might correctly identify NerdWallet, Forbes, Bankrate, Money.com, or another durable publisher as an important source in its broader category. But the websites cited for the brand's exact high-intent questions can still change substantially.

That suggests two different planning layers:

Strategic source layer

Track publishers that repeatedly appear across months, platforms, and related verticals. These sources may represent durable citation authority in the broader market.

Operational query layer

Re-run the brand's most commercially important prompt clusters regularly enough to identify source turnover, newly visible publishers, and disappearing citation opportunities.

For a high-priority AI-search program, monthly monitoring is a more defensible operating baseline than treating a quarterly or annual snapshot as permanent. Brands with fast-moving categories or active campaigns may reasonably monitor more frequently. This cadence is an operational recommendation based on the observed drift, not a claim that one universal interval is optimal for every company.

A Stable Publisher Ecosystem Can Still Produce Unstable Buyer Journeys

This is the part of citation drift that is easy to miss.

Imagine a market where the same 100 publishers receive most of the attention every month. That market can look stable from the top down.

But if the specific five or six publishers attached to each individual buyer question are constantly rotating inside that broader pool, a consumer can still encounter a substantially different source path from one month to another.

That is approximately what the LLM Authority Index data shows.

The portfolio of leading domains is persistent.

The query-level source composition is much more dynamic.

Both statements can be true at the same time.

This distinction is especially important for agencies and publishers that sell AI-search visibility. A broad claim such as "AI cites these publishers" is less useful than a measurement system that can answer:

• which publishers are persistently visible;
• which publishers are visible for a particular commercial prompt cluster;
• which platforms are citing them;
• whether that visibility is rising or falling; and
• whether the same source relationships persist when the prompt is repeated later.

How This Study Relates to Prior Citation-Persistence and Citation-Drift Research

Citation persistence, citation decay and source drift are active areas of AI-search research. This study should not be interpreted as the first attempt to measure whether citations persist over time.

The important comparison is the unit being measured. A page-level persistence rate, a source cohort half-life, a same-question Jaccard score, a monthly leaderboard and a Top 100 portfolio overlap statistic can all be valid while producing very different percentages.

AirOps: Citation Persistence as Repeated-Run Durability

AirOps defines citation persistence as the degree to which an AI answer engine keeps citing the same brand or page when the identical query is run repeatedly over time. Its recommended measurement approach fixes the query set, reruns prompts across days, records cited pages and calculates how often a page remains cited.

Source: AirOps, "What is citation persistence?"

That definition is closely related to this study's same-prompt layer, but the unit is not identical. AirOps focuses on page or brand persistence across repeated runs. LLM Authority Index separately reports aggregate domain-portfolio persistence and matched cited-domain-set similarity.

The shared commercial implication is important: a one-time citation is not enough evidence to describe durable AI visibility.

Foglift: Frozen Buyer-Intent Prompts Across Quarters

Foglift's controlled 2026 citation-drift benchmark repeats the same 75 brand-neutral buyer-intent questions across 25 verticals and five engine lanes in Q2 and Q3. Foglift reports mean same-question citation-set turnover ranging from 77.0% to 86.7% depending on engine.

Source: Foglift, "AI Engine Citation Drift 2026"

Foglift defines turnover as one minus mean Jaccard similarity after domain normalization and within-response deduplication.

The LLM Authority Index July-to-September matched panel reports 73.1% aggregate turnover among 4,309 non-empty matched prompt/platform pairs. The figures should not be treated as estimates of one universal volatility rate. The time interval, prompt set, engine implementation and collection surfaces differ.

The stronger comparison is structural: both studies show that stable-looking aggregate citation markets can coexist with substantial same-question source turnover.

Scrunch and Stacker: Citation Half-Life and Cohort Survival

Scrunch and Stacker analyzed 3.5 million citation events from September 2025 through March 2026 using source cohorts and survival curves. Their reported average citation activity fell by half in roughly 4 to 5 weeks, with meaningful differences by platform, industry and source type.

Source: Scrunch, "The half-life of AI citations"

That is a different measurement from this study. A half-life asks how quickly citation activity decays across a source cohort. LLMAI asks how similar monthly leading-domain portfolios remain and how much the cited-domain set attached to an identical prompt changes.

The commercial relevance is complementary: portfolio authority can persist while individual citations decay, which is another reason publishers and brands need both strategic source maps and recurring monitoring.

Writesonic: Page-Level Citation Decay

Writesonic's Citation Decay framework tracks page-level citation share over a rolling 14-day window and classifies pages as rising, holding, declining, faded, recovering, dormant or new. It defines observed half-life as the number of days from a page's peak citation share until that share falls below half and remains there within the measurement window.

Source: Writesonic, "Introducing Citation Decay"

Writesonic also separates trends by platform and topic. That reinforces a key conclusion of this series: a combined visibility number can hide platform-specific gains and losses.

The LLMAI study differs by using a frozen high-stakes consumer-decision corpus, monthly portfolio analysis and strict matched-prompt domain-set comparisons rather than a page-level rolling share metric.

Ahrefs: Monthly Citation Leaderboards and Rank Movement

Ahrefs updates platform-specific most-cited-domain studies and publishes month-over-month rank movement for major sources across AI surfaces.

Source: Ahrefs, "The Most-Cited Websites in ChatGPT"

Ahrefs is useful prior art for showing that source rankings are time-specific. Its broad U.S. query population and leaderboard design answer a different question from the strict matched-prompt analysis here. A domain can move in a monthly ranking without revealing whether an identical buyer question changed its exact source set.

Tinuiti and Profound: Citation Trends Across Platforms and Categories

Tinuiti's Q3 2026 AI Citation Trends Report, using Profound data, tracks citation changes across ChatGPT, Perplexity, Google AI Mode, Google AI Overviews, Gemini and Microsoft Copilot. It also breaks trends down by commercial category and source type.

Source: Tinuiti, "AI Citation Trends Report Q3 2026"

Tinuiti's findings reinforce the need to separate engines and categories rather than treating AI citation behavior as one market-wide metric.

Why These Studies Do Not Produce One Universal Persistence Number

These studies measure related but different phenomena:

The answer in this high-stakes consumer panel is yes.

How the LLM Authority Index Citation-Drift Study Is Different

This article is not differentiated simply because it measures change over time. Other researchers already do that.

The combination of the following design elements is what makes this analysis useful:

High-stakes consumer decision focus

The panel is concentrated on financial, insurance, mortgage, retirement, health, medical, and related commercial decisions where publisher authority can have meaningful consumer and business implications.

Six public AI platform families

The research covers Google AI Overviews, Google AI Mode, ChatGPT, Gemini, Perplexity, and Microsoft Copilot.

Large repeated-query layer

The strict July-to-September comparison contains 6,601 same-prompt, same-source-surface pairs, including 4,309 pairs with at least one cited domain in either period.

Aggregate and same-prompt stability are reported separately

This avoids an important analytical mistake. A stable Top 100 is not presented as proof that the same answers use the same sources.

Balanced-vertical sensitivity test

Because some July verticals were never produced, the study repeats the aggregate Top 100 test using the 45 verticals present in all three months rather than silently treating missing data as zeros.

Commercial category breakdown

The matched analysis is also reported across all four high-stakes consumer decision families rather than only as one cross-web average.

Why Citation-Drift Studies Can Produce Different Percentages

There is no universal "citation volatility rate."

A drift percentage depends on what is being compared.

Important differences include:

• hours, days, months, or quarters between observations;
• exact repeated prompts versus changing prompt populations;
• buyer-intent versus informational questions;
• consumer UI capture versus API or grounded-model monitoring lanes;
• whether domains are deduplicated within a response;
• URL-level versus registrable-domain comparison;
• treatment of zero-citation answers;
• treatment of platform-owned surfaces and infrastructure links;
• geographic market;
• model version and retrieval-index changes;
• source extraction rules; and
• whether the metric is Jaccard similarity, retention, rank correlation, raw citation share, or another measure.

For this reason, a 73.1% LLMAI turnover figure and an 80% turnover figure from another study are not competing estimates of one universal parameter. They describe different measured populations under different collection contracts.

Verticals Covered in the High-Stakes Consumer Decision Study

The source corpus spans 53 vertical labels grouped into four high-stakes consumer decision families. Closely related verticals can share prompts, so the labels should not be interpreted as 53 statistically independent industries.

Credit, Debt, Banking and Lending

Auto Refinance Loans; Bad Credit Loans; Best Banks; Certificates of Deposits; Credit Cards; Credit Cards for Building Credit; Credit Monitoring; Credit Repair; Debt Relief & Consolidation; Home Equity Loans; Money Market Accounts; Personal Loans and Online Lenders; Savings Account; Student Loan Refinance; Student Loans; Tax Relief.

Insurance

Car Insurance; Dental Insurance; Disability Insurance; Health Insurance; Life Insurance Companies; Long-Term Care Insurance; Medicare Supplement Insurance; Pet Insurance; Renters Insurance; Short Term Health Insurance; Travel Insurance; Vision Insurance.

Investing, Retirement, Mortgage and Financial Decisions

Annuities; Crypto Exchanges; Crypto Wallets; Gold IRAs and Precious Metals Dealers; Mortgage; Mortgage Refinance Lenders; Online Financial Advisors; Online Stock Brokers; Reverse Mortgage; Robo-Advisors; Structured Settlements.

Health and Medical

Addiction Treatment Centers; Assisted Living Facilities; Dental Implants; Fertility Clinics; Hearing Aids; Home Health Care; IVF Clinics; Medical Alert Systems; Mental Health Treatment Centers; Online Doctors; Online Pharmacies; Online Therapy; STD Tests; Weight Loss and Metabolic Health.

Category-level research is published separately so readers can evaluate source authority without requiring hundreds of thin individual-vertical pages:

• Most-Cited Websites in AI for High-Stakes Credit, Debt, Banking and Lending Consumer Decisions
• Most-Cited Websites in AI for High-Stakes Insurance Consumer Decisions
• Most-Cited Websites in AI for High-Stakes Investing, Retirement, Mortgage and Financial Consumer Decisions
• Most-Cited Websites in AI for High-Stakes Health and Medical Consumer Decisions

Methodology

Research source

This analysis uses the same frozen source archive as the 2026 AI Citation Authority Study.

The broader primary structural dataset contains 60,881 deduplicated AI responses and 278,499 visible citation events collected across July, August, and September 2026.

Eligible months

The drift study uses July, August, and September 2026.

A May dataset exists in the archive but is excluded from the primary three-month research window.

Explicit extraction failures

The delivered July-to-September archive contains 1,278 observations explicitly marked as extraction failures. Those records are excluded from valid-response analysis rather than treated as ordinary zero-citation responses.

Corpus-level exact-repeat handling

For aggregate rankings, identical analytical records are counted once when month, exact raw source surface, normalized prompt, and exact original citation URL set all match.

This prevents the same extracted answer from being counted repeatedly merely because it appears in overlapping vertical exports.

Prompt normalization

Prompt matching uses Unicode NFKC normalization, case folding, whitespace collapse, and trimming. Prompt wording is otherwise preserved.

Registrable-domain normalization

Citation URLs are parsed to hostnames, normalized to lowercase, stripped of leading www., and consolidated to the registrable domain using the Public Suffix List.

This means subdomains such as money.usnews.com are analyzed under usnews.com for the public domain-level stability measures.

Infrastructure exclusions

Obvious non-substantive navigation or infrastructure artifacts are excluded from the public domain comparison, including Bing image hosts, certain Bing map or ad wrappers, OpenAI image hosts, Google account-activity pages, and similar technical assets.

Platform-owned substantive surfaces are not automatically deleted solely because the platform owns the domain.

Response-level domain ranking

Monthly Top 100 rankings use response-level domain presence. A registrable domain counts no more than once within an AI response, even if the response cites multiple URLs from that site.

Monthly Top 100 overlap

For each month, the 100 domains with the highest response-level presence are selected. Pairwise list similarity is measured with Jaccard similarity:

Jaccard similarity = shared domains / unique domains across both lists

Rank correlation

Spearman correlation is calculated among domains appearing in both compared Top 100 sets. It measures whether shared domains retain a similar ordering, not whether every domain remains in the list.

Strict matched-prompt analysis

For same-prompt drift, records are grouped by exact raw source surface and normalized prompt.

A month pair is eligible only when exactly one analytical record exists for that key in each compared month. Keys with multiple distinct records in either month are excluded from the strict comparison.

This produces:

• 8,685 strict July-August matched pairs;
• 6,601 strict July-September matched pairs; and
• 10,986 strict August-September matched pairs.

For source-set similarity, pairs where both months contain no cited domain are excluded from the Jaccard and turnover calculation.

Same-prompt Jaccard similarity

For each eligible repeated prompt:

Jaccard similarity = cited domains in both months / cited domains in either month

An exact match scores 1.0. A pair with no shared domain scores 0.

Citation-set turnover

For this article:

Observed citation-set turnover = 1 - mean Jaccard similarity

This makes the measure comparable in concept to other same-prompt citation-drift studies, while the different datasets and collection systems still prevent direct percentage equivalence.

Balanced-vertical sensitivity analysis

The full monthly corpus contains different numbers of source verticals because some July datasets were never produced.

A sensitivity analysis therefore restricts the aggregate monthly Top 100 calculations to the 45 source verticals represented in all three months.

No missing July dataset is imputed or zero-filled.

Limitations

1. Three months is a short longitudinal window. The study measures July through September 2026, not long-term source stability across years.
2. Observed drift does not identify a cause. Model updates, retrieval-index changes, platform experiments, publisher changes, query-routing differences, and collection effects can all contribute.
3. The platforms expose citations differently. Cross-platform turnover percentages should not be interpreted as a pure ranking of product stability.
4. The source archive is not a perfectly balanced panel of verticals. The 45-vertical sensitivity test addresses part of this issue but cannot remove every sampling difference.
5. Strict matched-prompt pairs are a subset of the corpus. Ambiguous prompt/surface keys with multiple distinct analytical records are excluded from the paired analysis.
6. A normalized prompt string does not prove the surrounding product context was identical. Platform interfaces and retrieval systems may have changed between collection dates.
7. Visible citations are not the same as hidden retrieval, training data, or model reasoning. The study measures surfaced source relationships only.
8. Citation persistence does not prove recommendation persistence. A domain can remain cited even when the recommended companies or framing change.
9. Domain-level analysis can hide page-level turnover. Two months can cite the same domain while citing different URLs from that publisher.
10. The high-stakes commercial panel is intentionally non-representative of all AI usage. The findings should not be generalized to every informational, entertainment, coding, or casual prompt.

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 that appear in the dataset.

Commercial relationships are not inputs to the citation rankings, matched-prompt calculations, or turnover metrics. No publisher's presence in this study should be interpreted as an endorsement, and no commercial relationship should be interpreted as evidence that it caused citation visibility.

What We Would Measure Next

The next phase of longitudinal citation research should extend the same frozen measurement contract beyond September.

That would allow LLM Authority Index to test:

• whether the 88 persistent Top 100 domains remain persistent over six or twelve months;
• whether same-prompt turnover compounds over longer intervals;
• whether platform-specific turnover converges or diverges;
• whether source types such as publishers, company-owned sites, government sources, nonprofits, Reddit, and YouTube have different persistence profiles;
• whether citation turnover predicts later changes in company mentions or recommendations; and
• whether press coverage or publisher outreach is followed by measurable changes in citation visibility.

The purpose of those future analyses would be measurement, not causal attribution unless the research design supports it.

References

1. LLM Authority Index. The 2026 AI Citation Authority Study.
2. LLM Authority Index. The Persistence-Portability Gap.
3. LLM Authority Index. Do AI Platforms Cite the Same Websites for High-Stakes Consumer Decisions?.
4. LLM Authority Index. Reddit's Decline in AI Citations.
5. AirOps. What is citation persistence?.
6. Foglift Research. AI Engine Citation Drift 2026.
7. Scrunch and Stacker. The half-life of AI citations.
8. Writesonic. Introducing Citation Decay.
9. Ahrefs. The Most-Cited Websites in ChatGPT.
10. Tinuiti. AI Citation Trends Report Q3 2026.

Related LLM Authority Index Research

Foundation and measurement framework

• The 2026 AI Citation Authority Study
• The Persistence-Portability Gap
• Do AI Platforms Cite the Same Websites?
• Reddit's Decline in AI Citations

Decision-family source studies

• The Most-Cited Websites in AI for High-Stakes Credit, Debt, Banking and Lending Consumer Decisions
• The Most-Cited Websites in AI for High-Stakes Insurance Consumer Decisions
• The Most-Cited Websites in AI for High-Stakes Investing, Retirement, Mortgage and Financial Consumer Decisions
• The Most-Cited Websites in AI for High-Stakes Health and Medical Consumer Decisions

Platform-specific Top 100 source studies

• The 100 Most-Cited Websites in Google AI Overviews for High-Stakes Consumer Decisions
• The 100 Most-Cited Websites in Google AI Mode for High-Stakes Consumer Decisions
• The 100 Most-Cited Websites in ChatGPT for High-Stakes Consumer Decisions
• The 100 Most-Cited Websites in Gemini for High-Stakes Consumer Decisions
• The 100 Most-Cited Websites in Perplexity for High-Stakes Consumer Decisions
• The 100 Most-Cited Websites in Microsoft Copilot for High-Stakes Consumer Decisions