Short answer: a useful experiment is not testing whether you can "force" Claude to cite its own domain, but whether a controlled intervention on the consistency of local information reduces errors and stably changes the observed source mix over time. Anthropic documents web search and citations, but does not publish a formula that provides direct control over source selection. Therefore, the experiment must be built around hypotheses that you can falsify.

Hypothesis

A defensible wording is:

For locations with consistent first-party information, aligning priority external profiles and clarifying page ownership will reduce the rate of factually wrong answers in the monitored query set and may change the source mix, compared to similar locations without the same intervention.

The main result is factual accuracy. Own domain citation is a secondary outcome, not the sole purpose.

Population

Choose locations comparable in service type, maturity and operational complexity. Do not compare a new headquarters with one that has years of reviews, consolidated external profiles and high local brand awareness.

If you have only one location, you can use a before/after design, but you have to be more careful with the assignment.

Baselines

Before the intervention save:

  • the canonical name of the location;
  • address or service area;
  • the schedule;
  • active services;
  • the canonical URL;
  • Local Business markup;
  • critical external profiles;
  • first-party conflict count;
  • factual accuracy in a fixed query set;
  • owned citation rate;
  • third-party citation rate;
  • the date of each observation.

Without a baseline, a favorable screenshot is not an experiment.

The intervention group

Changes must remain limited to the tested layer:

  1. fix first-party conflicts;
  2. clarify an owner URL for each material information;
  3. align structured data with the page;
  4. you update the external profiles that you legitimately control;
  5. keep the same offer and the same commercial copy, as much as possible.

Do not change pricing, design, service pages and campaigns at the same time if you want to isolate the effect of the intervention.

The comparison group

It uses one or more nearby locations as a structure, but without the full intervention in the first window. If a material error occurs, fix it even if it affects control. The integrity of the information takes precedence over the purity of the experiment.

If the comparison group receives a redesign, a campaign or a wave of reviews in the meantime, mark the event as a possible confounder.

The query set

Construct questions in four categories:

  • identity: "What is the address?", "What schedule is it?";
  • offer: "Offer service X?";
  • eligibility: "Does it cover area Y?";
  • comparison: "What is the difference between location A and B?".

Keep the formulation stable within a phase. If you change the prompts, a new version of the experiment starts.

Metric 1: factual accuracy

For each verifiable claim, use the same rubric: correct',incomplete', wrong',unverifiable'.

Report the numerator and denominator. For example, 34 correct claims out of 40 eligible claims says more than "very good accuracy".

Metric 2: source mix

Separate owned source, local directory, review platform, publication and other sources. Don't assume owned source is always superior. For a customer experience, an independent source may be suitable.

Metric 3: naming stability

Track whether the system consistently uses the correct location name and does not fall back to historical aliases. This metric is especially useful after a rebrand or relocation.

Metric 4: correction lag

It measures the time between the first-party correction and the moment when external profiles or monitored outputs no longer repeat the error. This lag is operational and may vary between sources.

Observation window

Define the forward range. There is no universal number of days. For schedule changes, the window may be short. For external profiles and recrawls, longer observations may be required.

Use the same frequency for the compared groups.

Confounders

Keep a journal for:

  • seasonal program;
  • changes in offer;
  • new reviews;
  • campaigns;
  • local PR;
  • relocations;
  • website changes;
  • changes to the AI ​​platform;
  • independent directory updates.

If a confounder is large enough that the groups are no longer comparable, stop the interpretation and start a new phase.

Stop criteria

Stop the experiment when:

  • the query set has been materially modified;
  • a location changes its basic offer;
  • the comparison group receives a major intervention;
  • first-party data can no longer be kept comparable;
  • the results have become stable over several rounds and new observations no longer change the conclusion.

Don't keep going just to find a positive result.

A realistic positive result

Let's say that before the intervention you have 72% factual accuracy, and after the intervention you have 91%, while the comparison group remains roughly stable. At the same time, the owned citation rate increases modestly, but some answers continue to cite third-party sources.

The mature conclusion is that the intervention was associated with an improvement in factual accuracy in the observed population. You don't conclude that "entity optimization forces Claude to cite the website".

A useful null result

You can get first-party conflict rates close to zero with no detectable change in citations. This is still a good result: the public system is more coherent and the team reduces the risk of conflicting information.

Acceptance criteria

The experiment can be reported if the hypothesis, population, control, intervention, observation window, denominators, confounders and raw observations are documented and reproducible.

Claim ledger

  • FACT/EVIDENCE: Anthropic documents web search and source citations.
  • FACT/EVIDENCE: Google documents LocalBusiness structured data for eligible local entities.
  • PRACTITIONER GUIDANCE: factual accuracy and source mix must be measured separately.
  • INFERENCE: reducing conflicts can contribute to more stable outputs.
  • NOT PROVEN: that a local intervention produces directly or guaranteed owned citations in Claude.

Conclusion

Editorial A/B for Claude web citations only makes sense if the experiment can also produce an unfavorable result. If any variation is interpreted as "GEO success", you no longer have experiment, you have storytelling. It defines the hypothesis, protects control, and measures factuality before vanity.

Sources reviewed

Related insights