Short answer: you can't correctly assign a business result to a Claude citation if you haven't defined beforehand what you're observing, what population you're targeting, and what other changes occurred during the same period. Anthropic documents that web search can use sources and citations. This allows you to see if a URL is shown as a source. But it doesn't give you a public attribution model for pipeline, revenue or brand lift. In B2B SaaS, good discipline separates four levels: visibility, citation, referral and commercial outcome.

Start with the observation unit

The smallest useful unit is not "the brand appears in Claude", but a documented run: a question, a moment, a way to use web search, an answer, and the sources shown. For each observation note the prompt, language, date, buyer journey category and visible URLs.

Without this drive, the dashboard becomes a collection of favorable captures. You cannot calculate a denominator or compare periods.

Separate the four layers

1. Visibility

The brand or product appears in the response. This is observable, but does not say where the information came from.

2. Citation

A URL is shown as the source. Separate first-party from third-party. A citation to your documentation and one to an independent publication are both relevant, but they answer different questions about the source of the information.

3. Referral

A session arrives on the site from that product and can be identified in analytics. The absence of the referral does not prove that the site was not used as a source. Its presence does not prove that a particular passage produced the click.

4. Outcome

Trial, demo, contact, SQL or revenue. These results have their own windows and their own confounders. Don't compress them into an "AI visibility score".

Baseline before optimization

Select a fixed set of questions: awareness, category, shortlist, comparison, integration, security, pricing logic and implementation. Don't just choose prompts where you know the brand appears.

Predetermine the number of observations and the frequency. If you do ten runs before the change and two after, the comparison is fragile. If you change prompts along with the site, you won't know what you tested.

The denominator matters

For owned citation rate', the denominator should be the eligible observations where the system displays sources and the question is related to the domain being measured. Forbrand mention rate', the denominator can be the complete set of followed prompts.

These two rates need not automatically share the same denominator. Otherwise, a change in web search availability can look like a change in brand visibility.

Conservative attribution model

Use labels, not manufactured certainties:

  • Observed after intervention: the result appeared after the change.
  • Compatible with contribution: the change could explain part of the result, but there are confounders.
  • Stronger evidence: more observations and a comparison group support the same direction.
  • Not attributable: the data does not allow the connection.

In B2B marketing, this discipline is more useful than a percentage of revenue arbitrarily assigned to a new channel.

Frequently occurring confounders

A product release can change search volume and media coverage. A PR campaign can introduce new external sources. A major review may change which sources are available. The team can update docs, pricing and homepage in the same week.

There is one more important confounder: the AI ​​system itself can change. Therefore, it preserves the date and version or mode when the product displays them.

How to use referral data without over-interpreting it

If analytics identifies traffic from an AI product, treat it as observed referral traffic. It links the session to the landing page and the outcome, but does not assume that all interactions with the brand in that product produce detectable clicks.

For B2B SaaS, a visit to the docs may have value without immediate conversion. A pricing visit can be part of a buying committee that converts weeks later.

What report is worth sending to leadership

A mature report has five columns: metric, denominator, baseline, current observation, interpretation limit. Then add concrete examples of sources and claims, not just percentages.

If the `owned citation rate' increases, it shows which URLs appear. If the referral traffic increases, it shows the landing pages and the quality of the sessions. If the pipeline is growing, it tells what other campaigns ran during the same period.

Acceptance criteria for an experiment

An assignment test is good enough for decision when:

  1. the set of prompts is fixed before;
  2. the denominators are defined;
  3. the baseline has enough observations to see the variation;
  4. the intervention is documented;
  5. other material changes are logged;
  6. the displayed sources are captured;
  7. the conclusion is proportional to the design.

What I wouldn't report

I wouldn't report "Claude brought us 18% of the pipeline" without a verifiable trail. I wouldn't turn an increase in mentions into proof of authority. I would not combine mentions, citations and referrals into a single score unless the formula has a real decisional justification.

Claim ledger

  • FACT/EVIDENCE: Anthropic documents web search and the use of citations in web-based answers.
  • PRACTITIONER GUIDANCE: separating visibility, citation, referral and outcome reduces misattribution.
  • INFERENCE: repeating observations can provide stronger evidence than a single screenshot.
  • NOT PROVEN: a universal revenue attribution model to a Claude citation.

Conclusion

Attribution for AI search needs to be more modest than the marketing that sells it. Measure what you can observe, define the denominator, and say clearly where the inference begins. In B2B SaaS, this discipline allows you to invest in resources and content without confusing visibility with revenue.

Sources reviewed

Related insights