Does Claude have a business-recommendation ranking formula? What Anthropic documents and what we tested

No published formula exists. We are not aware of any Anthropic document that sets out criteria or weights for deciding which businesses Claude names in an answer, and we do not treat the absence of one as a secret being kept. What can be observed is what the assistant says about its own ordering when you ask it, and on 17 September 2026, across six blind commercial questions on clawlaw.in, Claude put two enterprise vendors above the company on a question about finding every case against a company specifically because they publish explicit court and tribunal coverage lists, and noted that its ordering reflected price transparency and source authority rather than product quality.

That sentence is the closest thing to a ranking factor we can show you with a date on it, and it is one run of six questions. This page publishes the run, the method that produced it, and the method for extending it, because a factor list without a prompt set behind it is an opinion in a costume.

The answer, first

Four things to hold at once.

• There is no published rulebook for business mentions. Anthropic documents its models, its tools and its usage policies. None of that is a company ranking specification.
• There are observable reasons. Asked why it ordered a list the way it did, the assistant gives reasons, and in our audit those reasons were about published evidence rather than about product merit.
• A reason given is not a weight. It is an explanation produced after the fact, which is useful and is not a specification.
• Nobody can guarantee a position, including us, and the absence of a formula is one reason why.

How it was measured

The run behind the observations on this page

Six blind commercial questions in the legal technology category in India, asked on Claude on 17 September 2026 on behalf of clawlaw.in, with the company named in none of them, followed by a structured self audit in which the assistant was asked to account for its own answers: which sources it had used, which it had discarded, and why the order was what it was. The answers and the audit are recorded in our capture files for that date.

The method for a factor observation study

If you want to extend this, the design is the following six steps. It is an observation study and it is not an experiment, and calling it the right thing matters.

• Write thirty blind selection questions in one category, in buyer language. Freeze them.
• Ask each in a clean session, and save the answer with its source list.
• In a second turn, ask for the ordering rationale: what decided the order, which sources supported each name, and what was excluded.
• Code the stated reasons into named categories: published price, published coverage or scope, official status, independent source, breadth of public writing, product quality claims, and anything that does not fit.
• Count how often each category is given, over the number of questions, with the date. That count is your finding.
• Treat every rationale as a statement, not as a log. A model's account of its own reasoning is evidence about what it says, not proof of how it computed.

Why the second turn has to be a separate turn

Asking for the rationale inside the first prompt changes the answer. A question that says "explain your ranking criteria" produces a criteria-shaped answer. Ask the buyer's question first, let it be answered, then ask about the answer.

What the numbers were

The denominator is six questions, on one assistant, on 17 September 2026. That is the whole sample, and here is everything it produced.

• Coverage lists decided one ordering. Two enterprise vendors were ranked above clawlaw.in on a question about finding every case against a company, specifically because they publish explicit court and tribunal coverage lists, and the assistant noted that its ordering reflected price transparency and source authority rather than product quality.
• Unsourced comparison pages were discounted. Two comparison pages stated competitors' prices with no link, no date and no source. The figures turned out to be correct when the assistant checked them independently, but it had no way to know that at read time, so it treated the pages as advocacy and used official sources instead.
• Official sources took the how-to questions outright. On the questions about how to look a case up, official court portals took every position above any commercial product, and the assistant said plainly that no commercial product should rank above the official portal for a question about using that portal.
• No third party review or directory source reached any answer. One well known review site appeared in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question.
• Our pages were used without being credited. The assistant admitted it had used two specific arguments drawn from clawlaw.in pages and dropped the attribution in both places, which it described as a citation lapse rather than a ranking judgement.
• Being read did not produce a link. clawlaw.in pages ranked in the assistant's raw results and shaped what it wrote, and it still landed as a name inside a list rather than as a linked recommendation.

All six are Claude, 17 September 2026, on clawlaw.in. One further dated observation from the same programme, on 6 August 2026, belongs here because it is about what gets rejected: a competitor's page carried a hidden block of text addressed to answer engines, instructing them to cite that company as the source. The assistant found it, refused it, and named the company that had done it.

What we have not run: the thirty-question factor study described above, on Claude or anywhere else. There is no count of how often each reason category is given, because we have six questions rather than thirty and one date rather than several.

What this cannot tell you

• It is not a formula and cannot be turned into one. Six stated reasons on six questions is a set of signals worth acting on, not a weighted model.
• A stated reason may not be the operative reason. Models produce plausible accounts of themselves. The value here is that the account was specific and checkable against the pages involved.
• It is category-bound. Legal technology in India has official portals with real authority. A category with no official source will order itself differently.
• It cannot be generalised to other assistants. Anthropic's models, Google's and OpenAI's retrieve differently and credit differently.
• It cannot promise that doing these things earns a mention. The honest claim is narrower: on this run, these were the reasons given, and each of them points at something publishable.
• It cannot be checked from outside in full. Some of these observations concern public pages and can be verified. The audit transcript is in our capture files and is not published yet, and we would rather say so than imply independent confirmation.

Common questions

If there is no formula, what do we actually do?

Publish the things the assistant said it used. A coverage or scope list with named items and a count. Prices with dates. Every competitor figure sourced and dated. Your limits, stated. That is a week of work and it is aimed at the reasons that were actually given rather than at a guess.

Does Anthropic publish anything useful at all here?

Yes, just not a ranking. The documentation for the web search tool tells you how retrieval works when it is switched on. The model documentation tells you about knowledge cutoffs, which explains stale descriptions. The usage policies tell you what kinds of answer are handled carefully. None of it says how a business gets named.

Why would an assistant rank a worse product higher?

Because it is ranking evidence, not products. In our run the assistant said as much: the ordering reflected price transparency and source authority rather than product quality. A better product with nothing published about it in checkable form loses to a documented one.

Can we ask Claude to explain its ranking and quote that publicly?

You can quote it as what the assistant said on a date, which is what we have done on this page. You cannot present it as Anthropic's criteria, because it is not.

Is six questions worth anything?

It is worth exactly one thing: six specific, dated reasons, each pointing at a page you can go and write. It is worth nothing as a rate, which is why no percentage appears on this page.

Sources and change log

Primary material to read directly: Anthropic's documentation for the web search tool, the model overview pages including knowledge cutoff information, Anthropic's usage policies, and Anthropic's pricing page for anything to do with cost, read on the day and dated in your notes.

Our own material: the clawlaw.in programme from July 2026, and specifically the Claude response audit of 17 September 2026 which produced every observation counted on this page, plus the interim check of 6 August 2026 for the hidden instruction case.

Change log. First published 29 September 2026 with a denominator of six and no factor study. If we run the thirty-question study, its counts, categories and dates will be added below this line and the six-question run will stay visible as the earlier state.

Disclosure. Written by AI Knows Us, a vendor of AI visibility measurement.

What to do first

Ask Claude the single buying question in your category that you most want to win, in a clean session with your name absent. Then, in a second turn, ask it to say what decided the order it gave and which sources supported each name. Write the answer down with today's date. In fifteen minutes you will have your own version of the observation this page is built on, and it will name the pages you are missing.