Short answer: in professional services, measuring Claude citations must start from a stable baseline, a versioned query set and separate metrics for factual accuracy, source mix, owned citations, third-party citations and time-to-resolution. Anthropic documents web search and citations, but does not publish a brand authority formula or official "viewability" metric. Therefore, the attribution must remain conservative.

Correct baseline

Select questions on several stages of the buyer journey: awareness, problem diagnosis, shortlist, methodology, implementation and comparison. For each question, keep the language, persona, intent and target entity.

In parallel, it saves the first-party status: the page that owns the service, expert profiles, methodologies, case studies and material claims. Without this layer, you can't tell whether an external response reflects a source problem, an internal problem, or simply variation.

The population and the denominator

Do not use all answers in the same denominator. For citation rate, only include remarks where the interface exposes sources. For factual accuracy, use only verifiable claims. For owned citation rate, the denominator is the total of eligible observations with citations, not the total of all prompts.

This detail prevents spectacular percentages obtained from incompatible populations.

Metric 1: factual accuracy

Classify each material claim as correct',incomplete', wrong' orunverifiable'. It separates the identity of the company, the role of the expert, the service, the methodology and any commercial claim.

A company may have a low owned citation rate, but good factual accuracy through independent sources. This is not automatically defective.

Metric 2: owned citation rate

Measure how often company-controlled URLs appear in eligible observations. It also shows the absolute number of observations.

Don't turn this metric into a single KPI. A first-party citation does not prove that the page is the best source for any question.

Metric 3: third-party source mix

It groups together publications, professional directories, associations, review platforms, speaker profiles and other sources. Measure each category separately.

For professional services, the source mix can vary between awareness and shortlist. This is normal.

Metric 4: identity stability

Track whether the company, service and experts are identified consistently. An obsolete role or a subbrand confused with the parent organization is more severe than a lexical variation.

Metric 5: first-party conflict rate

Measure material claims that differ between service pages, bios, methodology pages and other internal surfaces. This is a directly controllable metric.

Metric 6: source freshness

For sources cited, note date of verification and, where available, date of update. Freshness is not synonymous with quality, but it helps to identify stale profiles or historical articles presented as current.

Metric 7: detectable referral

If analytics identifies referral traffic from an AI product, track landing page, session quality and lead path. But keep the limit: many influences do not produce measurable clicks.

Do not extrapolate revenue directly from citation count.

Metric 8: time-to-resolution

For each finding, it saves the moment of detection, the owner, the date of remediation and the date of reverification. This metric shows whether the program can close problems, not just notice them.

Observation window

Define the period before analysis. For operational monitoring, use the same frequency and query set across the entire window. If there is a rebrand, a launch or a migration, it marks a new version.

Don't extend the window just to get a positive trend.

False-attribution risks

Confounders frequently appear in professional services:

  • PR and new interviews;
  • changes in the role of experts;
  • acquisitions or rebrand;
  • new backlinks;
  • site rewrites;
  • new profiles in directories;
  • query set changes;
  • updates of the AI ​​platform.

Keep a change log.

How to report without vanity metrics

Good report: "In 60 eligible observations, factual accuracy remained stable and first-party conflicts decreased after cleanup; source mix became more varied. We do not attribute the change solely to the intervention."

Poor report: "Claude authority increased by 37%."

Acceptance criteria

The measurement is auditable when:

  1. the query set is versioned;
  2. populations are defined;
  3. the denominators are clear;
  4. raw observations are kept;
  5. identity claims have an owner;
  6. source types are separated;
  7. observation window is fixed;
  8. confounders are logged;
  9. external outcomes are distinct from quality metrics;
  10. the conclusions do not claim unproven causality.

How do you compare practices without falsifying the population

A legal, tax and technical practice can have very different volumes, source footprints and buyer journeys. Do not compare raw citation counts. Keep the denominator at the level of eligible observations and report the coverage of each practice separately. If a group has too few observations, use NOT_PROVEN instead of a firm conclusion.

How do you deal with role changes and rebranding

If a partner leaves, a practice is renamed, or the firm enters into an alliance, it closes the old version of the series and starts a new one. Historical mapping can remain for audit, but do not mix before and after into a single trend.

Claim ledger

  • FACT/EVIDENCE: Anthropic documents web search and source citations.
  • PRACTITIONER GUIDANCE: factual accuracy, source mix and conflict reduction must be measured separately.
  • INFERENCE: a versioned query set reduces false attribution and allows cleaner comparisons.
  • NOT PROVEN: the existence of a universal Claude authority score or a public brand visibility formula.

Conclusion

In professional services, citations become useful when they are treated as operational evidence, not trophies. It measures what you can reproduce, separates populations, and preserves the line between association and causation. Thus, the program remains useful even if the source mix changes from one week to another.

Sources reviewed

Related insights