A citation in a Claude reply may be useful as an observation, but it does not prove that a page is consistently preferred, that the commercial message is correctly understood, or that an editorial change produced the effect. For B2B SaaS, a serious test must separate three things: the occurrence of a citation, the cited source, and the accuracy of the product claim. Only after these three layers are measured separately is attribution worth discussing.
The experiment below is built for teams that have documentation, product pages, comparisons, and technical materials distributed in multiple places. It does not assume access to an official "citation share" metric and does not invent one. It works with saved responses, observed URLs and a log of content changes.
1. Start with the real questions in the buying journey
The set of prompts should reflect how a B2B buyer investigates a solution. Problem discovery questions are different from questions about integration, security, migration, or comparisons with a competitor. If you mix them up, a citation change may come from intent distribution, not site intervention.
Build prompt families. For example: "what solutions solve problem X", "how does product Y integrate with Z", "what options are there for security requirement A", and "what are the differences between Y and an alternative product". Keep exact wording, language and date. If you paraphrase, treat them as new observations, not duplicates of the same question.
2. Choose a control that can actually function as a witness
The control should not be a random page. Ideally, you use a document family with close intent, but to which you do not apply the tested editorial change. If the intervention aims to clarify the integration documentation, keep a comparable set of unchanged technical pages.
In B2B SaaS, the product changes often. Note releases, API updates, pricing changes, and navigation changes. A new feature released in the middle of the test can create new sources or change what existing pages say. Without this log, you will easily attribute the effect to the wrong editorial tactic.
3. Save the full response, not just the quoted domain
For each run it keeps the answer, order of citations, URLs, and the sentence or passage that the citation appears to support. This detail is important: a page may be cited, but the answer may misrepresent the product. Conversely, a correct statement may come from a third-party source, not from the company's documentation.
In the journal, separate at least four states: own domain does not appear, own domain appears without supporting the main claim, own domain appears and correctly supports the claim, and a third-party source describes the product. Do not convert these states into a single score until you have a clear methodological reason.
4. Change a single page family
If you change the product page, documentation, schema, internal linking, pricing page and comparison page at the same time, the experiment becomes impossible to interpret. Choose a narrow intervention: for example, clarifying the structure of the integration documentation or adding a section explaining the limitations of a feature.
The intervention must be noted before publication: what text changes, what URLs enter the treatment, what remains control and why. If a content error or technical setback occurs, define the rollback before release. A good experiment is reversible.
5. Don't demand immediate results
Observed citations may vary between runs, and the status of indexing or available sources may change. An appearance on the same day as the amendment does not demonstrate the effect of the amendment. Use an observation window sufficient for multiple runs and maintain the same collection discipline.
Don't compensate for uncertainty by volume of prompts out of control. More useful is a stable set, executed at several times, with a good log. If you change both the set of questions and the documentation, you can no longer tell what has actually changed.
6. Separately assess the correctness of the product description
For a B2B buyer, a quote is only valuable if the resulting information is useful and accurate. It verifies claims about integration, security, availability, plan limits, and compatibility at the primary source. If the answer attributes a nonexistent function to the product, treat the case as a factual issue even if the company's domain is cited.
This separation prevents a common error: considering any citation as a positive result. A GEO experiment should not optimize the appearance of a source at the cost of a less accurate description.
7. Control releases and external distribution
In B2B SaaS, information also lives in marketplaces, partner pages, public repos, changelogs, and review sites. If a partner simultaneously publishes a new tutorial, it can become a relevant source independently of the change being tested on the site.
Include in the change log important external events: releases, webinars, campaigns, releases, integration into a marketplace and documentation updates at partners. You don't need a perfect list, but you do need enough information to identify obvious confounders.
8. Define replication criterion before verdict
A first result may be interesting, but it is not enough for generalization. Repeat the set of prompts in a subsequent window and, if possible, on a similar document family. The replication criterion must be written before: for example, the same direction of change for the citation of the documentation and the absence of factuality degradation.
If the result doesn't replicate, don't hide it. Compare the conditions of the two rounds. Maybe the product release changed the context, maybe the third-party source became more relevant, maybe the intervention wasn't isolated enough. These explanations are more helpful than a forced verdict.
Claim ledger
- FACT/EVIDENCE: Anthropic's documentation describes web search and citation mechanisms for using the tool, but does not provide a guarantee that a particular domain will be cited for a particular prompt.
- FACT/EVIDENCE: Google Search documentation describes general conditions for accessing and understanding web content, without guaranteeing citations in third-party AI systems.
- PRACTITIONER GUIDANCE: control, changelog and replication are methodological elements for attribution, not official requirements of Anthropic.
- INFERENCE: clearer documentation may increase the chance that the product is described correctly, but the effect on citations needs to be measured.
- NOT PROVEN: that a Claude citation directly produces pipeline, revenue or brand preference.
Stop criteria and limits
Stop interpretation if the prompt set has changed substantially, if the product has had a major release, if the control documentation has changed, or if responses are no longer collected under comparable conditions. Do not publish an "uplift" without a denominator and without a description of the observation window. Additionally, citations are platform behavior and may vary over time.
