AI Citation Decay Measured Across Five Engines Shows One in Four Cited Domains Still Cited Two Months Later

Answer first: Citation decay is the measured decline in how often AI engines keep citing the same source for the same question. In the Machine Relations Index, a domain cited in an answer was cited again by the same engine for the same question 49.6% of the time within 7 days, 32.2% of the time 29 to 56 days later, and 26.2% of the time 57 to 91 days later (4,560 question-engine pairs across ChatGPT, Claude, Gemini, Google AI Overviews and Perplexity, May 10 to October 15, 2026). About half of cited domains already fail to repeat a week later, so a single missed citation is noise; decay is the decline below that one-week floor. Freshness, competitive publication, engine changes, retrieval availability, entity ambiguity and source changes are candidate explanations to diagnose after the decline is observed.

Last updated: October 15, 2026. First published May 14, 2026.

What citation decay means #

Citation decay is the measurable loss of AI citation presence over time. A brand shows citation decay when the same tracked query set, run against the same engine set under comparable conditions, cites the brand less often than it did in earlier observation windows.

The key constraint is comparability. A one-off missing citation is not enough. The observation should hold across a frozen prompt panel, a defined engine set, consistent geography and account state where possible, and repeated measurement windows. Without that longitudinal baseline, the safer label is citation absence, answer volatility, or engine divergence — not decay.

Citation decay is distinct from SEO ranking loss. A page can retain an organic search position while appearing less often in generated answers because AI engines choose sources at answer time. The metric records what the engine cited, named, or omitted; it does not by itself explain why the engine changed.

In the Machine Relations framework, citation decay belongs in Machine Relations Layer 5: Measurement. It sits opposite citation velocity, but the relationship is observational rather than a simple equation: velocity records new citation gains, while decay records citation loss across a defined panel.

How fast AI citations decay: the Index measurement #

The Machine Relations Index reruns its library of buyer questions in each engine on a recurring schedule, so the same question can be compared with itself over time. For every pair of answers to one question in one engine, the measure below asks: of the domains cited in the earlier answer, what share is cited again in the later one?

Three readings follow from the table.

1. The one-week floor is about half. Even a few days apart, an engine re-cites only about half of the domains it cited for the same question. Any decay measurement that treats one missing citation as a loss is measuring this churn, not decay.
2. Decay is the slope below the floor. Retention keeps falling with time: from 49.6% at one week to 26.2% at two to three months for domains, and from 43.2% to 19.4% for exact pages. Exact pages fall faster than domains, so a publication can stay in an answer while the specific article it was cited for drops out.
3. Engines decay at different speeds. The curve is not one number.

ChatGPT and Perplexity start with the highest short-term repeat rate and lose the most over two months, roughly 27 to 29 points. Claude starts lower and loses about 10 points. Google AI Overviews has the fewest comparisons (71 answer pairs in the 29 to 56 day band), so its row is the least stable.

Method note. Source: Machine Relations Index answer runs for its public-scoring question library, observation window 2026-05-10 to 2026-10-15, release mri_score_v2.1+2026-10-15+d0dabb90d4d0 (release manifest). One answer per question, engine and date; answers with at least one citation; 18,699 answers to 1,227 questions, citing 27,596 distinct domains. Retention is pooled over every ordered pair of answers to the same question in the same engine: cited domains in the earlier answer that reappear in the later one, divided by cited domains in the earlier answer. Pages are compared as exact URLs without query strings. Google AI Mode is excluded because its citations per answer fell from about 15 to about 3 in September 2026, so its retention cannot be separated from that interface change. Two checks hold the shape: measured in reverse (domains in the later answer that were present in the earlier one, all six engines) retention runs from 50.4% at one week to 19.2% after 92 days, so the decline is not an artifact of answers citing fewer sources over time; and in a fixed cohort of 798 question-engine pairs observed at every interval, retention falls from 54.3% at one week to 26.4% after 57 days. Engine models changed during the window, and that change is part of what a brand experiences as decay. These are observed repeat rates, not causes.

What the current evidence can and cannot prove #

Several adjacent literatures help explain why citation presence may be unstable, but they do not directly prove longitudinal brand decay. This page uses them as boundary evidence, not as a causal guarantee.

• Academic citation age is not brand retrieval decay. Wahle et al. studied citation-age recession across scholarly fields and found that many fields cite older work less often over time (Wahle et al., 2024). That supports caution about recency effects in information environments. It does not prove that AI answer engines demote older commercial brand pages.
• Cross-sectional citation association is not time-series survival. Kumar & Palkhouski's GEO-16 study analyzed English-language B2B SaaS pages, 70 product-intent prompts, and citations from Brave Summary, Google AI Overviews, and Perplexity; it found associations between page quality pillars and citation in that corpus (Kumar & Palkhouski, 2025). Because the design is observational and cross-sectional, it should not be read as proof that treatment effects persist or erode over time.
• Citation repair is not a guarantee of durable presence. Tian et al. introduced AgentGEO, a diagnostic benchmark for repairing citation failures and reported relative citation-rate improvements in their test setting (Tian et al., 2026). That supports the idea that citation failures can have diagnosable causes. It does not show that refreshed commercial content restores or extends citation presence in live engines.
• Academic citation validity is not retrieval freshness. Xu et al. measured invalid and fabricated citations in academic writing and LLM citation-generation tasks (Xu et al., 2026). That is evidence about scholarly citation integrity, not direct evidence that AI engines apply stricter freshness filters to brands.
• Retrieval-contamination simulation is not brand decay measurement. Yu et al. modeled retrieval collapse under AI-generated web contamination (Yu et al., 2026). It is relevant to source-quality risk, but it is a simulation of contaminated retrieval pools, not a longitudinal study of brand citation loss.
• Scientific LLM adoption is not commercial content lifespan. Trišović measured shrinking adoption lifespans for LLMs in scientific papers (Trišović, 2026). That finding concerns model adoption trajectories, not how long a brand page stays cited by answer engines.
• Selection and absorption differ from survival. Zhang, He & Yao proposed a measurement framework distinguishing citation selection from citation absorption across AI search platforms (Zhang, He & Yao, 2026). The framework is useful for measuring what happened in a run, but it is not a direct decay-rate study.
• Citation failure benchmarks do not prove corroboration as a freshness signal. Buchmann & Gurevych studied failures in RAG citation completeness and mitigation methods (Buchmann & Gurevych, 2025). That helps separate answer quality from citation quality; it does not establish independent-domain corroboration as a live-engine freshness mechanism.
• GEO optimization is visibility intervention evidence, not decay evidence. Aggarwal et al. introduced Generative Engine Optimization and showed that content changes can improve visibility in benchmarked generative-engine settings (Aggarwal et al., 2024). That supports measuring visibility as an outcome. It does not prove that the improvement decays on a fixed timeline when content is not refreshed.

The practical conclusion is narrower and stronger: citation decay should be measured directly, as the Index measurement above does. Adjacent studies can inform hypotheses, but the decay claim comes from repeated comparable engine observations.

Candidate explanations to diagnose after decay is observed #

Once a longitudinal decline is visible, the next job is diagnosis. Common candidate explanations include:

1. Freshness changes. Newer pages, revised pages, dated rankings, or recent news may become more attractive sources for a time-sensitive query. This is a candidate cause, not a universal rule.
2. Competitive publication. Competitors may publish clearer, more extractable, or more directly query-aligned evidence that enters the answer engine's source pool.
3. Engine changes. The engine may change retrieval providers, browsing behavior, citation UI, ranking heuristics, or answer composition. A decay pattern in one engine but not another often points here.
4. Retrieval availability. The brand page may be harder to crawl, fetch, parse, or cite because of technical changes, rendering problems, blocked resources, canonical conflicts, or content structure.
5. Entity ambiguity. The engine may no longer resolve the brand, product, founder, or category claim as cleanly as before, especially when names are shared or entity signals conflict.
6. Source changes. The cited page, supporting third-party sources, or surrounding site architecture may have changed in ways that reduce extractability, specificity, or evidence density.

A measured decline can have more than one explanation. Treat the list as a diagnostic map, not as a deterministic hierarchy.

How to detect citation decay #

Detection requires systematic measurement, not spot checks. The following five-step workflow separates signal from noise.

Step 1: Establish a query baseline #

Select 20 to 50 queries that represent the brand's core buyer conversations. Run each query across a defined engine set on a defined cadence. Keep prompts, geography, account state, device context, and date handling as stable as possible. Record whether the brand is cited, which URL is cited, the visible citation position where the interface exposes it, and which other sources appear beside it.

Step 2: Track citation frequency over time #

Plot citation frequency as appearances per query per engine per observation window. A downward pattern across repeated windows is the decay signal. Measure each engine's own one-week repeat rate first: in the Index it ranges from about 35% to 55%, so a brand cited half the time week to week is at the floor, not decaying. The number of windows required depends on answer volatility and sample size, so label early movement as a candidate decline until enough repeated observations exist.

Step 3: Identify displacement sources #

When the brand drops from a citation slot, record what replaced it. A newer competitor page, an updated marketplace page, a publication article, a generic explainer, or no citation at all imply different diagnostic paths. Displacement analysis helps separate competitive movement from engine volatility.

Step 4: Cross-engine divergence check #

Compare the same query across engines. If the brand remains cited in one engine but disappears in another, the decline may be engine-specific rather than brand-wide. If the brand declines across multiple engines under comparable runs, the evidence for a broader visibility problem is stronger. Zhang, He & Yao's selection-versus-absorption distinction is useful here because citation count and answer influence can move differently (Zhang, He & Yao, 2026).

Step 5: Source freshness and availability audit #

For every previously cited page, check whether the page is still available, crawlable, canonicalized correctly, internally linked, semantically clear, and substantively current for the query. Do not apply universal 90-day or 180-day thresholds. Instead, compare the page's update pattern, dated claims, and evidence density against the sources now being selected.

What accelerates observed decline #

Not all citation loss has the same risk profile. The following factors can make observed decline more likely, but each still needs measurement in the query-engine panel.

1. Single-source dependency. A brand whose visibility relies on one URL has fewer alternate citation paths if that URL stops being selected.
2. Thin entity signals. Incomplete structured data, inconsistent naming, stale profile pages, or ambiguous product language can make a brand harder for engines to resolve.
3. Weak supporting evidence. Claims supported only by the brand's own page may be less resilient than claims also described by independent sources, but the cited research above does not prove a universal corroboration ranking factor.
4. Dynamic query categories. Queries with dates, vendor comparisons, regulations, pricing, or rapidly changing product categories are more exposed to source churn than evergreen explanatory queries.
5. Technical retrieval friction. Rendering failures, blocked crawlers, canonical mistakes, redirects, or pages with low extractability can remove otherwise useful content from consideration.

These are risk factors, not proof of causation. A decay report should name the measured decline first and then assign confidence to each suspected explanation.

How to respond to citation decay #

The countermeasure is not automatic content volume. It is source architecture: keeping the brand's evidence retrievable, current where the query requires currency, and clear enough for engines to cite.

Maintain measurement before prescribing action. Continue the frozen query-engine panel while investigating. If the measurement setup changes midstream, start a new baseline rather than mixing runs.

Repair availability and extractability. Confirm that previously cited URLs render server-side where possible, expose stable canonical URLs, include clear headings, preserve source facts, and avoid burying the answer behind scripts or vague copy.

Clarify the entity. Align organization names, product names, founders, categories, structured data, and internal links so the engine can resolve the brand consistently. Entity clarity can reduce ambiguity, but it does not guarantee citation durability.

Refresh only when the query needs freshness. For time-sensitive pages, update the evidence, dates, comparisons, and cited support. For evergreen pages, unnecessary date churn can be less useful than improving specificity and source clarity. Aggarwal et al. show that content optimization can change visibility in tested settings, but the live decay response still needs measurement (Aggarwal et al., 2024).

Strengthen external evidence without assuming permanence. Earned media, partner pages, public profiles, documentation, and independent references can create more possible citation paths. They should be treated as additional evidence nodes to measure, not as guaranteed protection against decay.

Separate recovery from attribution. If citation presence returns after a refresh, do not assume the refresh caused the recovery unless the design supports that claim. Engine changes, competitor changes, and random answer variation may also explain the movement.

Proposed protocol labels #

Some useful operating ideas are best treated as proposed protocols until measured against live longitudinal panels:

• Comparative decay tables by engine should be labeled as hypotheses unless backed by the same query set over time.
• Fixed decay windows, such as four to six weeks or two to three months, should be replaced with category-specific baselines.
• Three-period classifiers can be used as internal triage, but they should not be presented as validated decay stages without evidence.
• Freshness thresholds should be derived from the query category and competing sources, not universal 90-day or 180-day rules.
• A "net visibility" equation that subtracts decay from velocity may be a planning metaphor, but it is not a validated measurement formula.

Frequently asked questions #

What is AI citation decay? #

AI citation decay is an observed longitudinal decline in how often AI engines cite a brand across a frozen, comparable query-engine panel. In the Machine Relations Index, about one in four domains cited for a question is still cited for it by the same engine two to three months later.

How is citation decay different from SEO ranking loss? #

SEO rankings and AI citations are different measurement surfaces. A page can keep an organic ranking while appearing less often in AI answers, but that divergence must be measured; it should not be assumed from ranking data alone.

Can citation decay be reversed? #

Sometimes citation presence returns after availability repairs, content updates, entity clarification, or stronger external evidence. Those actions are plausible responses, not guaranteed reversal mechanisms. Continue the baseline panel to verify whether citation presence actually recovers.

How quickly does citation decay happen? #

In the Machine Relations Index, a domain cited for a question is cited again by the same engine for that question 49.6% of the time within 7 days, 32.2% at 29 to 56 days and 26.2% at 57 to 91 days. The speed differs by engine: ChatGPT and Perplexity lose about 27 to 29 points over two months, Claude about 10. No fixed four-to-six-week or two-to-three-month rule applies; report the observed curve for the engines and questions you track.

How do you measure AI citation decay? #

Run a fixed set of questions in each engine on a schedule. For every pair of answers to the same question in the same engine, compute the share of sources cited in the earlier answer that reappear in the later one, grouped by the time between answers. The one-week value is the noise floor; decay is how far retention falls below it as the interval grows.

What is the relationship between citation decay and citation velocity? #

Citation velocity measures new citation gains over a defined query-engine panel. Citation decay measures losses over that same kind of panel. Both belong in Layer 5 measurement, but they should be reported as separate observed metrics rather than collapsed into a deterministic equation.

Where does citation decay fit in Machine Relations? #

Citation decay sits in Machine Relations Layer 5: Measurement. It helps determine whether a brand's source architecture is sustaining, losing, or redistributing AI visibility over time. The metric observes movement; diagnosis of the upstream cause requires a separate evidence review.