Zero-Citation AI Answers Are Not Missing Measurements

AI citation-share dashboards should treat a successful AI answer with zero citations as a measured result, not as the same state as a timeout, failed request, or unscheduled observation. The metric denominator must include only answer runs that were successfully observed and eligible for citation evaluation; missing measurements belong in coverage health, not citation rate.

The operating rule is simple: zero citations requires a valid answer; missing measurement does not.

Zero-citation AI answers need their own measurement state

A zero-citation answer is an observed answer run where citation extraction was possible and the target source, domain, or URL did not appear. That is evidence of absence inside the declared measurement unit. A missing measurement is different: the system did not collect a valid answer, did not schedule the observation, or quarantined the result before citation evaluation.

This distinction is already familiar in adjacent measurement systems. AAPOR's disclosure standards ask researchers to report enough sample-disposition detail for outcome rates to be computed, rather than hiding eligibility and nonresponse inside one headline number (AAPOR disclosure standards). The U.S. Census Bureau similarly treats item nonresponse as missing data that may require imputation or allocation, not as the same thing as a reported answer of zero (U.S. Census Bureau).

AI citation measurement needs the same discipline. A dashboard can report that a source was not cited only after it has a valid answer observation to inspect.

A compact state table for AI citation-share dashboards

The dashboard should expose collection state before it computes citation share. Four states are enough for a practical first implementation.

State	What happened	Citation extraction possible?	Eligible for citation-rate denominator?	Eligible for coverage denominator?	Correct label
Success with citations	The engine returned a valid answer and the target source was cited at least once	Yes	Yes	Yes	Positive citation result
Success without citations	The engine returned a valid answer and no target citation was found	Yes	Yes	Yes	Measured zero-citation result
Timeout or failed request	The scheduled observation did not produce a valid answer before the cutoff, or the payload could not be evaluated	No	No	Yes, if scheduled	Collection loss or failed measurement
Not scheduled	No prompt-engine-date observation was planned for that cell	No	No	No	Outside the measurement plan

The citation-rate denominator is success_with_citations + success_without_citations after the metric's other eligibility rules are applied. The collection-coverage denominator is scheduled observations. The not-scheduled state is outside both denominators unless the report is measuring planning coverage.

That table prevents a common dashboard error: treating blank cells as zeros. A blank cell can mean the source was absent from a valid answer. It can also mean the request never ran.

An illustrative ledger separates success, zero, timeout, and not-scheduled

The following ledger is illustrative. It is not a production defect report and does not describe a known failure in the Machine Relations Index or any vendor system. It shows how analysts can preserve the state distinction in a citation-share dashboard.

Assume the dashboard tracks example.com for one prompt across two engines and four dates.

Observation cell	Prompt	Engine	Scheduled?	Collection state	Valid answer?	Target cited?	Citation-rate treatment	Coverage treatment
A1	best ai visibility tools	Perplexity	Yes	Success with citations	Yes	Yes	Numerator + denominator	Covered
A2	best ai visibility tools	ChatGPT Browse	Yes	Success without citations	Yes	No	Denominator only	Covered
A3	best ai visibility tools	Gemini	Yes	Timeout	No	Not evaluated	Excluded from citation rate	Scheduled but not covered
A4	best ai visibility tools	Claude Web	No	Not scheduled	No	Not evaluated	Excluded from citation rate	Outside plan
A5	how to measure ai citation share	Perplexity	Yes	Success without citations	Yes	No	Denominator only	Covered
A6	how to measure ai citation share	ChatGPT Browse	Yes	Success with citations	Yes	Yes	Numerator + denominator	Covered

From this ledger, the target's domain citation rate is two cited valid answers out of four valid eligible answers, or 50.00%. The collection coverage is four covered observations out of five scheduled observations, or 80.00%. The not-scheduled Claude cell is disclosed as planning scope, not as either an uncited answer or a failed request.

Those two percentages answer different questions. Citation rate asks whether the source was selected when a valid answer existed. Coverage asks whether the monitoring system collected the planned answer observations.

Denominator eligibility must be explicit beside every metric

A citation-share dashboard should name the denominator where the number appears. The safest label is not just "citation share". It is "domain-cited valid answers / eligible valid answers" or "source cited answers / successful answer observations."

Google Search Console's performance documentation is a useful analogy from a different system: clicks, impressions, and position depend on how the report is grouped and filtered by query, page, country, device, and appearance (Google Search Console Help). The lesson transfers cleanly. A percentage is not interpretable until the report names the unit being counted.

For AI citation dashboards, four denominators commonly appear side by side:

Metric	Numerator	Denominator	Includes zero-citation successes?	Includes timeouts?	Includes not-scheduled cells?
Domain citation rate	Valid answers where the domain was cited at least once	Valid eligible answer observations	Yes, in denominator	No	No
URL citation event rate	Extracted citation rows for matching URL identity	Valid eligible answer observations or extracted citation rows, as labeled	Depends on formula	No	No
Collection coverage	Successful valid answers	Scheduled observations	No, because they are successes	Yes, in denominator as failures	No
Planning coverage	Scheduled observation cells	Expected prompt-engine-date cells	Not applicable	Not applicable	Yes, as gaps in plan

If an analyst wants a conservative metric that treats failures as uncited, that can be published as a separate stress metric. It should not replace the primary citation rate because it mixes source selection with collection reliability.

The Machine Relations Index uses observed answer runs, not attempted API calls

The Machine Relations Index is the evidence authority for public AI source behavior in this article. Its public methodology states that MRI measures source-domain citation rates inside source segments, publishes rates only after an evidence floor of at least 10 observed runs across at least seven distinct run dates, and keeps thin segments in a collecting state rather than scoring them as zero.

Editor-held native release evidence for September 15, 2026 identifies release mri_score_v2.0+2026-09-15+e512364a281a, artifact SHA256 e512364a281ae55e7352dea9c3b3a26296f9cb38e80ff367b96033699cf3f7aa, the May 10-September 15 window, 122 observed days, 122,144 citation events, 15,468 answer runs, and six healthy engines. That release context is used here only as denominator evidence inherited from native proof; an independent public HTTP check available during this run showed the public index page still exposing an older September 12 text snapshot, so this article links the stable public index while preserving the source-proof limit.

The September 15 Machine Relations methods note on citation-rate uncertainty from run-date clustering makes the same denominator discipline more precise: a source-domain rate depends on eligible observed answer runs inside the selected stratum, not on all attempted API calls, release-wide observed days, or raw citation-event rows.

That distinction matters for this article's edge case. A timeout in a separate brand monitoring panel is real collection health evidence. It is not a reason to replace MRI's released denominator with attempted calls from another system.

A timeout is a coverage result, not a zero-citation result

A timeout says the monitoring system did not receive an answer that could be evaluated under the metric. It may be caused by provider latency, network failure, rate limits, blocked retrieval, parser errors, or orchestration limits. None of those facts proves that the source would have been absent from a valid answer.

The FDA's E9(R1) statistical guidance is written for clinical trials, but its estimand framework names a transferable principle: the question being estimated, the data collection design, and events that affect interpretation need to be handled explicitly before analysis (FDA guidance). In an AI citation dashboard, a timeout is an event that affects the existence of the measurement. It is not a measured negative outcome.

A clear report therefore shows both lines:

• Citation rate: 42 cited valid answers / 120 valid eligible answers = 35.00%.
• Collection coverage: 120 valid answers / 126 scheduled observations = 95.24%.

Those lines can move independently. Citation rate can fall while coverage improves. Coverage can fall while citation rate among valid answers is unchanged. Collapsing the two hides which system changed.

This specification preserves the existing Paralax citation-rate canon

This article does not redefine AI citation rate. The existing Paralax definition says AI citation rate measures how often an answer engine cites a source when a relevant question is asked. This article adds the implementation edge case: a relevant question must produce a valid answer before "not cited" can be counted as a measured zero.

It also preserves the two recent measurement specifications. The September 13 bridge article explains how prompt cohort migration breaks trend comparability when prompts are retained, removed, added, or rewritten. The September 14 article explains how retry attempts and duplicate URLs can distort numerator and denominator units when prompt and URL identity stay fixed.

This specification covers a third boundary: prompt identity may be fixed and URL accounting may be clean, but a dashboard still needs to separate successful answers with no citations from timeouts, failed requests, and not-scheduled cells.

A practical implementation convention for zero-citation answers

A dashboard can implement the convention with five fields on every observation cell.

Field	Required values	Why it exists
schedule_state	scheduled, not_scheduled, cancelled	Separates planned coverage from cells outside the plan
collection_state	success, timeout, provider_error, parser_error, blocked, quarantined	Explains whether a valid answer exists
answer_valid	true, false	Gates the citation-rate denominator
citation_evaluated	true, false	Prevents unevaluated answers from becoming zeros
target_citation_state	cited, not_cited, not_evaluated	Separates positive, measured zero, and missing measurement

The computation then becomes auditable:

1. Filter to the declared prompt cohort, engine panel, locale, date window, and source identity rule.
2. Exclude not-scheduled cells from the citation-rate calculation.
3. Exclude scheduled cells without answer_valid=true from the citation-rate calculation and report them in collection coverage.
4. Count target_citation_state=cited as the numerator.
5. Count both cited and not_cited valid answers as the denominator.

NIST describes metrological traceability as connecting a measurement result to the documentation chain that produced it (NIST). AI citation dashboards are not laboratory instruments, but the traceability principle is useful: the rate should be reproducible from prompt identity, engine, date, collection state, answer validity, citation extraction, and source identity rules.

Machine Relations measurement needs absence and missingness to stay separate

Machine Relations treats AI visibility as a source-system problem: prompts, engines, answer evidence, source identity, and public authority all have to stay legible enough for a human or model to reproduce the claim. Share of citation is useful only when the dashboard states which answer runs were eligible to contribute to the share.

The safest dashboard language is direct:

• "The target was cited in 2 of 4 valid eligible answers."
• "Two valid answers had no target citation and count as measured zero-citation outcomes."
• "One scheduled cell timed out and is reported as collection loss, not an uncited answer."
• "One prompt-engine-date cell was not scheduled and is outside this report's denominator."

That is less convenient than a single blank cell. It is also the difference between evidence of absence and absence of evidence.

FAQ

Is a zero-citation AI answer the same as a failed AI request?

No. A zero-citation AI answer is a valid observed answer in which the target source was not cited. A failed request, timeout, or invalid payload did not produce an answer that can support a citation-rate denominator.

Should timeouts count as uncited answers in citation share?

No, not in the primary citation-share metric. Timeouts should be reported in collection coverage or failure health. A conservative stress metric may count failures as uncited only if it is separately labeled and never confused with observed citation rate.

What denominator should AI citation rate use?

AI citation rate should use valid eligible answer observations: successful answers where citation extraction could be evaluated under the declared prompt, engine, locale, date, and source-identity rules. It should not use all scheduled cells, all attempted API calls, or all raw citation links.

How should dashboards show not-scheduled observations?

Not-scheduled cells should be shown as planning scope. They are outside the citation-rate denominator and outside collection coverage unless the report is specifically measuring how complete the monitoring plan is.

Where does this fit inside Machine Relations?

This is measurement-layer discipline inside Machine Relations. A brand cannot diagnose source authority, earned media gaps, or AI visibility movement until its dashboard separates measured absence from missing collection evidence. Teams can benchmark the broader source-system gap with the AuthorityTech AI visibility audit.