Citation Decay

Definition and Measurement #

Machine Relations definition: Citation Decay names a downward change in an entity's observed citation rate between comparable measurement windows.

Treat Citation Decay as a descriptive measurement, not an explanation of why citations changed. A valid comparison keeps the query set, engine, interface, observation procedure, and denominator consistent. If the protocol changes, report a new measurement basis instead of presenting the result as decay.

A citation-rate observation needs four fields:

1. the queries in the measurement panel;
2. the engine and interface observed;
3. the number of eligible responses; and
4. the number of those responses that cited the entity or its defined source set.

The comparison should retain the answer text, cited URLs, and observation time so the result can be inspected later. Report each engine separately before calculating any combined view.

What Existing Evidence Establishes #

SalesPeak reports that roughly half of the content cited in the AI-search responses it analyzed was less than 13 weeks old (SalesPeak AEO Research).

Foglift reports that 61.7% of the top-cited domains in its comparison appeared in only one engine's rankings (Foglift, 2026).

Moz analyzed nearly 40,000 search queries and found that 88% of AI Mode citations are not in the organic SERP for the exact match query (Moz). That observation describes one engine's citation behavior for one sample; it does not define a universal relationship between organic rank and AI citation presence.

These observations justify measuring freshness, engine-specific citation presence, and the separation between organic rank and AI citation. They do not establish a universal decay window, prove that inactivity caused a decline, or show that every engine responds to source age in the same way.

What the Machine Relations Index Observes #

The Machine Relations Index reruns its library of buyer questions in each engine on a recurring schedule, so one question can be compared with itself over time. For its release of 31 October 2026 (observation window 10 May to 31 October 2026), it measured how fast the sources an engine cites for a question turn over: of the domains cited in an earlier answer, what share the same engine cited again for the same question in a later answer.

Time between answers	Domains cited again	Exact pages cited again	Answer pairs compared
1 to 7 days	51.6%	46.0%	11,989
8 to 28 days	43.3%	37.1%	22,465
29 to 56 days	34.9%	27.8%	10,042
57 to 91 days	26.6%	19.6%	10,022
92 days or more	21.7%	14.9%	3,623

This is source turnover at the engine level, not Citation Decay as defined on this page. It pools many domains and describes how quickly the set of cited sources changes; it is not any one entity's citation rate, and it is not a minimum or a threshold an entity must clear. It does show three things for measurement. About half of cited domains repeat within a week and about half do not, so a single missing citation inside a week is expected turnover and is not, by itself, Citation Decay. Exact pages turn over faster than domains, so a publication can stay in an answer while the specific article cited drops out. Engines differ: within 1 to 7 days the repeat rate runs from 34.1% (Google AI Overviews) to 60.7% (Perplexity), so read each engine against its own row. The engine-by-engine table, the exclusions (Google AI Mode and Perplexity comparisons that span its source-count changes are left out) and the method are in AI Citation Repeat Rates Across Five Engines.

Citation Decay Across Engines #

Citation Decay should be calculated independently for each observed engine and interface. A decline in one panel is evidence about that panel only. It is not evidence that the same decline occurred elsewhere.

Observation	What it records	What it does not establish
Citation rate in an earlier window	The entity's presence in that declared panel	A permanent baseline
Citation rate in a later window	The entity's presence in the comparable panel	An explanation for the difference
Difference between the windows	The direction and size of the observed change	An engine-wide ranking rule
Cross-engine comparison	Whether the observed direction differs by engine	Shared retrieval architecture

The engine label is part of the measurement. Do not merge observations from different products, modes, or retrieval states without disclosing the change in basis.

Citation Decay vs. Citation Velocity #

Citation Velocity records additions to observed citation presence, while Citation Decay records losses from that presence.

Read the two measures as separate descriptions of change. Neither measure proves why a source entered or left an answer, and neither establishes a business outcome.

Metric	Observable event
Citation Velocity	A citation appears where the comparable prior observation did not contain it
Citation Decay	A citation is absent where the comparable prior observation contained it
Share of Citation	The entity's citation presence within the declared response panel

Investigating a Decline #

Working hypothesis (not measured): A citation decline can coincide with source changes, query drift, competitor activity, retrieval changes, or changes to the measured entity. Those are investigation paths, not causes established by the decline itself.

For each lost citation, compare:

• the earlier and later answer text;
• the earlier and later cited URLs;
• the query wording and engine mode;
• whether the previously cited page remained available;
• whether a different source occupied the same answer role; and
• whether the entity name or source set changed.

This comparison separates an observed loss from an explanation. If the available evidence cannot identify an explanation, record it as unknown.

What Citation Decay Is Not #

Citation Decay is not proof of an algorithm penalty, competitor displacement, source-age effects, or the effect of earned media activity.

It is also not an SEO-ranking metric. Record search position and AI citation presence as separate observations. A relationship between them must be demonstrated with data rather than assumed from either metric.

How to Detect Citation Decay #

Use a stable query panel and retain the evidence from each collection window. For every comparison:

1. verify that the engine, interface, query set, and eligibility rules match;
2. calculate the citation rate from the recorded responses;
3. identify citations present in the earlier window and absent in the later one, reading a single absence against the engine's source turnover (see the Index observation above) before calling it a decline;
4. report the changed observations by engine; and
5. disclose any protocol or source-access differences.

Machine Relations treats a decline as detected only when comparable observations show a lower citation rate. The label says what changed; it does not assign responsibility for the change.

---

FAQ #

How fast does Citation Decay happen? This page specifies no general timeline for an entity's decay. Record the interval between comparable observations and report the observed change without converting it into a universal schedule. What the Machine Relations Index does measure is how fast cited sources turn over: in its 31 October 2026 release, 51.6% of domains cited in an answer were cited again by the same engine for the same question within 7 days, and 26.6% were cited again 57 to 91 days later. That is engine-level turnover across many domains, not the decay rate of any one entity.

Can Citation Decay be reversed? A later observation may contain the citation again. That observation does not, by itself, identify the intervention responsible for the return.

Does updating existing content prevent decay? This page does not claim that it does. Test the question by recording the update, preserving an unchanged comparison panel, and comparing later citation observations without treating correlation as causation.

How do you measure Citation Decay? Compare citation rates from matched observation windows using the same query panel, engine, interface, and eligibility rules. Retain the underlying answers and cited URLs, report results by engine, and label the explanation unknown unless separate evidence establishes it.