Short answer: an operational system for Claude web citations does not start with "how we get citations", but with maintaining a consistent public version of the truth about location and services. Anthropic documents web search and citations, but does not publish a formula by which a local business can control source selection. In practice, the team must manage location identity, first-party sources, important external profiles, the set of monitored questions, and the changelog.

Precondition 1: registry per location

For each site or operational area, store internally the public name, address, phone, schedule, active services, coverage area, canonical URL, and last revision date.

The registry is not a public page. It is the source of truth for QA and the teams that update the site, local profiles and support materials.

If two locations have different rules, keep the difference explicit. Don't normalize the information just to make the template simpler.

Precondition 2: ownership

Every critical attribute needs an owner. The program may belong to operations. Prices may belong to the commercial team. The service page can have an editorial owner.

Without ownership, the audit only notes the drift. Can't close it.

Stage 1: first-party audit

Check homepage, location pages, service pages, footer, contact, policy and structured data. Look for schedule, address, phone, service or name conflicts.

These conflicts take precedence over any citation analysis because they are completely under the company's control.

Stage 2: canonical ownership

For each material question, determine the page that holds the answer. The location page can hold the schedule and address. The service page may contain the scope and terms of the service. Don't put the same information in five places without a sync process.

Internal linking must lead the user to the correct owner, not to a network of almost identical pages.

Stage 3: critical external profiles

Map only the platforms that matter for the buyer journey or appear repeatedly in monitored outputs. For each note what data it contains, whether it can be updated, and who has legitimate access.

Don't try to clean the entire web. A short list of relevant sources is more workable than an endless inventory.

Stage 4: set of questions

Build queries by category:

  • identification and location;
  • schedule;
  • services;
  • coverage area;
  • price or calculation method;
  • eligibility;
  • comparison;
  • contact and scheduling.

Version the set. Don't change prompts after every unfavorable result.

Stage 5: observation

For each run note the date, question, answer material, sources shown, brand mention, owned citation, third-party citation and factual accuracy.

Keep raw observations. Screenshot can be useful, but text and URLs are easier to compare.

Stage 6: classification of findings

Use operational states:

  • ``first-party conflict'';
  • ``external stale'';
  • ``source mismatch'';
  • ``incorrect output'';
  • ``incomplete output'';
  • ``no issue'';
  • ``not enough evidence''.

Do not turn `owned citation absent' into an automatic finding.

Step 7: The fix

The healthy order is:

  1. correct first-party;
  2. synchronize structured data with the page;
  3. update legitimate external profiles;
  4. clarify internal ownership;
  5. wait for a reasonable recrawl/refresh;
  6. rerun the same query set.

Don't rewrite content just because an engine chose another correct source.

Step 8: measuring

Watch separately:

  • factual accuracy;
  • first-party conflict rate;
  • external critical-profile consistency;
  • owned citation rate;
  • third-party citation rate;
  • referral;
  • time-to-resolution.

Do not combine these metrics into a single "GEO score". Each answers a different question.

Stage 9: cadence

Check weekly for deliberate program and service changes. Monthly, run the query set and main external profiles. Quarterly, review the registry and sources still worth maintaining.

The cadence depends on the volatility of the business. A seasonal service may require more frequent review.

Acceptance criteria

The system is operational when:

  1. each location has a registry;
  2. critical attributes have an owner;
  3. first-party conflicts are zero or have an active task;
  4. important external profiles are mapped;
  5. the query set is versioned;
  6. raw observations are kept;
  7. the denominators of the metrics are documented;
  8. findings have severity and closure criterion;
  9. the fix does not introduce duplicate pages;
  10. conclusions do not promise citations.

Rollback and limitations

If a new local page does not bring information gain, consolidate it. If an external profile cannot be controlled, mark `external unresolved'. If an attribute changes too often to copy manually, move it to a more suitable operational source.

Do not alter the first-party truth to imitate a stale external source.

A complete example

A company has two offices. Office A works on Saturdays, office B does not. The global footer says "M-F" but the A page says "M-S". A third director repeats "M-F". Claude quotes the director for a question about A's headquarters.

The main finding is first-party conflict, not "Claude prefers directories". Fix the footer or make it contextual, check the LocalBusiness markup, update the directory if you have control, and then rerun the observation. If Claude continues to cite the directory, but the answer becomes correct, the factuality has improved even if the source is not one's own domain.

Claim ledger

  • FACT/EVIDENCE: Anthropic documents web search and citations.
  • FACT/EVIDENCE: Google documents LocalBusiness structured data and recommends that the markup reflect the page.
  • PRACTITIONER GUIDANCE: registry, ownership and conflict states make the process workable.
  • NOT PROVEN: a public mechanism by which a location can guarantee a Claude citation.

Conclusion

Claude web citations for local services must be managed as an information and operation issue. When locations have consistent data, well-defined owners, and sources, the team can measure what's changing without turning every citation variation into a GEO crisis.

Sources reviewed

Related insights