How Claude Decides Which Businesses to Recommend
Progressive search, citation boundaries, deployment-level domain controls, and the index Anthropic does not name.
Start with the two Claudes: training and live search
Claude can answer from what its model learned, but that knowledge has a cutoff date. Anthropic publishes cutoff dates by model and warns that later events may be missing or wrong. When a question concerns a current business, a changing service, a local provider, or an explicit request to look something up, Claude can use web search instead. Treat these as two different visibility conditions. A business may be absent from the model's learned associations yet perfectly discoverable in live results, or present in training but described there in an outdated way. For a practical test, ask the same question with web search off and on, record the date, model, wording and cited pages, and do not treat either answer as a permanent position.
Make the business answer the exact local question
Anthropic documents that Claude generates a targeted search query and may refine it during a request, and that it can run several progressive searches where earlier results inform later ones. That makes the user's wording important. Build pages around the real combinations customers ask about: service, location, audience, problem, timing and meaningful constraints. State plainly who you serve, where you operate, what you do, and what is outside your scope. Use the same business name and location consistently so a retrieved page is unambiguous. Do not stuff a page with every city and every service. A page that directly answers "who provides this service in this place, and why?" gives Claude usable text when its query matches that intent.
Put proof on pages Claude can actually retrieve
A low-profile business should assume that its name alone is weak evidence. Publish first-party pages that substantiate the claims a recommendation would require: exact services, service area, qualifications, relevant experience, contact details, availability and clearly qualified outcomes. Use readable prose and stable URLs. Put important facts in text rather than only in images, scripts or downloadable files that are harder to process. Separate verified facts from testimonials, and date the material. The goal is not to manufacture authority but to make legitimate claims easy for a search system and a human reader to inspect. Anthropic publishes no required evidence count, so never promise that a specific number of pages will trigger a recommendation.
Build corroboration without manufacturing it
Anthropic says response quality depends on the underlying sources, and its help guidance tells users to inspect the original pages because synthesis may omit context. An independent legal-search test observed Claude using a mixture of directories, recognitions, first-party firm pages and deep topical articles. For a small business that suggests a practical evidence trail: maintain an accurate official site, claim profiles in relevant professional or local directories, and seek genuine third-party coverage or awards that can actually be verified. Keep names, addresses, phone numbers, credentials and dates consistent. Do not create fake listings, duplicate thin pages, invented awards, or paid placements presented as independent reviews. Corroboration reduces ambiguity, but Anthropic has not said that citation count, domain authority or review volume carries a fixed weight.
Write pages that support a precise citation
Anthropic's citation documentation explains that citable content is organized into source text, and that clear permanent URLs, descriptive titles and logical text blocks improve citation boundaries. Apply the same discipline to public pages. Give each important topic its own durable URL and a descriptive title. Put one claim, or a tightly related set of facts, in a short section, followed by any qualification or source. Make the page easy to scan without hiding the substance behind tabs or vague marketing language. Check that the page loads for ordinary visitors and that the cited claim is genuinely present on the linked page. A citation is useful evidence that retrieval happened, but it does not mean Anthropic has endorsed the business.
Keep current facts current
Claude's documented trigger for web search includes information that changes, such as current organizations, products, prices and statistics. Review the facts a customer might use to choose you: opening hours, service areas, staff, licences, pricing ranges, availability, offers and contact routes. Put an updated date only where it reflects a real review, and remove expired claims rather than leaving contradictory versions online. If a service has changed, update the page and the relevant profiles so retrieval cannot combine old and new facts. Ask Claude current, date-stamped questions and inspect the citations returned. Recency is a documented reason Claude may search. It is not documented as a universal ranking boost, and it is certainly not a guarantee of selection.
Test recommendations as evidence, not as a score
Run repeatable prompts that name the service, location, customer need and date. Test with web search enabled, because a no-search answer measures model familiarity rather than live discoverability. Save the exact response, the citations, the query wording where it is visible, and the pages Claude used. Repeat on different days and with neutral alternatives, then check whether the business was named, merely cited, or omitted entirely. A single result is not a ranking position. Anthropic publishes no stable score, weight or minimum threshold, and search availability plus domain controls can vary by account or deployment. Use the tests to find factual gaps and retrieval problems, not to claim that a placement has been achieved or will persist.
What is documented, and what we are inferring
Documented by Anthropic: model-specific knowledge cutoffs; optional or automatic live web search for current or explicitly requested information; targeted and sometimes repeated searches; cited web-sourced responses; user advice to inspect original sources; deployment-level domain controls; and citation structures for retrieved content. Independently observed in one legal-search test: query refinement and use of relevant directories, recognitions, first-party pages and detailed articles. Inferred, not proven: that clear local service pages and corroborating public evidence improve the chance of being considered, because they make a business more relevant and more verifiable in retrieved material. Unknown: the text-search index, the crawler rules, ranking weights, any reputation calculation, and any low-profile-business threshold. Keep every recommendation inside that boundary.
Did this guide answer your question?
See how visible your business is to AI
Get your free 0–100 AEO score across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews.