Forty Questions

How to check whether Claude recommends your software

A step-by-step way to find out whether Claude names your company when a buyer asks what to buy, how to read what it looked at before answering, and why one check is not enough.

Written 5 October 2026 from 450 Claude answers collected 21 September – 3 October · Satayu Isariyaphorn

To check whether Claude recommends your software, ask it the question a buyer would type, with your country in it, in a new conversation with web search turned on, and repeat the same wording on three separate days. Count the answers that name you and note which pages Claude cites, because with search on it builds the answer from those pages. In our data Claude named the measured company in 180 of 450 answers, and the set of companies it named was the same on two separate days only 45% of the time, so a single check tells you about the day, not about Claude.

Step by step

  1. Write the question the way a buyer would. Not your product name: the problem and the place. “What is the best rostering software for a small café in Australia?” rather than “Is RosterElf good?”. Write five of these and keep the wording.
  2. Start a new conversation, with web search on. In Claude, open a fresh chat outside any project, and make sure web search is enabled in the settings. Claude answers buyer questions by searching; without search it is answering from training data that may be a year old.
  3. Ask one question per conversation. No follow-ups. A follow-up (“what about us?”) is a different measurement: you have told it who you are. Close the chat and open a new one for the next question.
  4. Record the answer and the sources. Note every company named, and open the source list. Those pages are what Claude read before answering. If you are not named, the pages that are cited are the ones a buyer’s answer is built from, and the list of places you are missing from.
  5. Repeat on two more days, same wording. Claude’s answers vary from day to day more than ChatGPT’s or Perplexity’s (see why three runs). Three runs on three days gives you “named in n of 15” for five questions, which you can compare next month.
  6. Count under a rule you wrote first. Decide before you look what counts as being named: your company name or an agreed alias, anywhere in the answer. Do not change the rule on a bad month.

What Claude does when it answers a buyer question

Across 450 buyer-question answers from Claude (model claude-sonnet-5, via the API with the web search tool, up to eight searches per answer), Claude searched in all but a handful and cited on average about 21 pages per answer. The companies it names are overwhelmingly the ones on those pages: review sites, comparison articles, directories and the companies’ own pages. So the practical question behind “does Claude recommend us?” is “are we on the pages Claude reads for this question?”, and the source list in step 4 answers it directly.

Doing it by API instead of by hand

If you want the check to run itself, the Claude API with the web search tool does the same thing as a fresh chat with search on. Each call is stateless, so there is no history to worry about; set the search tool’s maximum uses so that cost is predictable. At our measured rate a Claude answer with search costs about $0.14 in API charges (how we measured that), so five questions, three days, is about $2 a month. The scripts we use are on GitHub.

If you are not named

The fix is rarely on your own site alone. Claude is citing the pages it found for that question; the work is to be on those pages (a directory listing, a comparison article that includes you, an answer to the question on a forum it reads) and to have one page on your site that answers the buyer’s question plainly. The $390 audit does this for forty questions across four assistants, with every cited page listed; or do it yourself with the steps above. If a competitor is named where you are not, the four fixes, in order.

The same check on the other assistants

The steps are the same for ChatGPT (search on, new chat, logged out if you can), Perplexity and Google’s AI Mode. The full rulebook, including how we handle names that are ordinary words and answers that are ranked lists, is on the method page.