How Does Claude Decide What to Recommend, and Why Does It Name So Few Businesses?

Ask Claude to recommend a coach and you often get a thoughtful answer with nobody in it. A list of things to look for, a warning about credentials, a suggestion to check reviews, and then it stops. No names.

That reads like caution, and it is. But it is not vagueness for its own sake. There is a mechanism underneath it, and once you can see the mechanism you can tell the difference between an assistant that will not name anyone and an assistant that will not name you.

This post covers when Claude looks something up and when it answers from what it already knows, why that produces so few business names, and what changes the answer.

How does Claude decide what to recommend?

It makes two decisions in order. First, whether the question needs information it does not already have. Second, what it can say without overstating what it knows.

Anthropic documents the first decision plainly. Claude searches when a request depends on information that is current, changing, or outside its training data, and it answers directly when the request draws on stable knowledge. Its web search documentation lists both sides: recent events, current prices, and information about specific organizations that might have changed on one side, established facts on the other.

That is the whole hinge. "What should I look for in an executive coach" is a question about a profession, and a profession is stable knowledge. "Who coaches first-time engineering managers in Denver" is a question about specific organizations, and it is far more likely to send Claude looking.

When it searches, and when it answers from memory

When Claude does search, it shows its work. The same documentation says citations are always enabled for web search, and each cited source carries a title, a link, and the passage that was used.

That matters more than it sounds. An answer built from a search is an answer you can audit: you can read the page it read. An answer given from stored knowledge has no sources attached, because there is nothing to attach. The two look similar on screen and mean completely different things about your visibility.

So the first question to ask about any answer is not whether you were named. It is whether anything was cited at all. If nothing was, the assistant never went looking, and no amount of publishing would have changed that particular answer.

Why it names so few businesses

Because a broadly worded question does not have a business as its answer, and a cautious model will not pretend otherwise. Look at what an answer to a broad question is actually made of.

WHAT ONE ANSWER IS MADE OFILLUSTRATIVE

Asked broadly "Can you recommend an executive coach?"

Sets up the question

Lists what to look for

Caveats and how to verify

Names

Asked narrowly "Who coaches first-time engineering managers in Denver?"

Sets up

What to look for

Caveats

Names, with what each one is known for

Illustrative proportions, not measured. The point is the shape: the part of the answer that can carry a business name is small when the question is broad, and it is the largest part when the question is narrow enough to have a real answer.

AnswerHalo answerhalo.com

CLICK TO ENLARGE

This is the same finding the research keeps arriving at from the other side. The Princeton study on generative engine optimization found sources gained visibility when they added citations, quotations and statistics, which is to say when they gave a model something specific enough to repeat. A page of adjectives gives a cautious reader nothing it is willing to say out loud.

What makes it willing to name you

Facts that survive being repeated by someone who did not write them. That is the whole test, and it is stricter than it sounds.

"Award-winning leadership coach" cannot be repeated, because whoever repeats it has to vouch for it. "Coaches first-time engineering managers, works with Denver software teams, twelve-week engagement, $6,000" can be repeated exactly as written, and the person repeating it is only reporting what your page says.

The table below is the same idea applied across the shapes a buyer question can take. It is worth reading as a map of which questions you can realistically win.

Google's guidance on creating helpful content describes the same page from a different angle: first-hand experience, a clear answer to the question asked, and enough detail that a reader leaves satisfied. Written that way, a page is quotable by a cautious assistant and useful to a person, which is not a coincidence.

What to do about it this month

Ask Claude the four questions in the table above, in your buyer's words, and note two things for each: whether anything was cited, and whether any business was named. Those two columns tell you which problem you have.

Nothing cited and nobody named means the question never sent anyone looking, and it is not a question worth chasing. Sources cited and competitors named means the looking happened and your pages were not good enough to use. That second case is the one worth work, and the same reading applies to the other assistants: our walkthrough for the one most of your buyers use is in how ChatGPT decides which businesses to recommend, and the equivalent for Google's assistant is in how Gemini decides which businesses to recommend.

Then fix the narrow question first. It is the one with names in its answer, it is the one your buyer types once they are serious, and it is the one your site can actually influence. The mechanics of what gets pulled into an answer at all are covered in what retrieval means in AI answers.

If you would rather see the reading done for you first, the free check asks ChatGPT 25 real buyer questions from your niche at $0 and emails a plain-English readout: your score, who got named in your place, and the one fix to start with. It runs as soon as you ask for it, and the report lands within the hour between 7am and 5pm Central, or by 10am the next morning outside those hours.