How to check if ChatGPT (and other AI assistants) recommend your brand

The most common mistake is typing your own name into an assistant and seeing a flattering description. That only shows the assistant has heard of you. It doesn't show whether it would recommend you to someone who has never heard of you, which is the situation of every new customer.

What matters is whether you appear when the buyer doesn't know your name yet. That means asking the category question, for example 'best invoicing tool for freelancers', not 'what is [your brand]'.

Do this once to get a baseline. It is manual, so keep it small.

• Write 5 questions a real buyer would ask. Mix the broad ('best X for Y'), the specific ('X with feature Z') and the comparison ('X alternatives to Competitor').
• Never include your own brand name in these questions. If you do, you've answered your own test.
• Ask each question to at least three assistants, for example ChatGPT, Claude and Perplexity, with web search turned on where it is offered.
• Run each question twice, in a fresh chat each time. Answers change from run to run.
• Record, for every answer: were you named, in what position, who was named instead, and which sources were cited.

Count the answers that name you out of the total. That is your mention rate. A brand that appears in 2 of 30 answers has a different problem from one that appears in 20 of 30 but always in fourth place.

The sources matter more than the answers. If the same review sites, comparison articles or forum threads keep getting cited when you're absent, those pages are where the assistants learn who to recommend. Getting accurately listed there is usually the fastest way in.

A single try can be wrong in either direction.

• Personalised results: assistants that remember you may favour tools you've mentioned before. Use a fresh, logged-out session where possible.
• Run-to-run variation: the same question can name different brands on different runs. One answer is an anecdote, not a measurement.
• Memory versus search: some answers come from the model's training data, others from live web search. They move on different timescales.
• Small samples: five questions is a starting point. Treat anything you conclude from it as a hypothesis.

Once you've done the manual baseline, the useful questions are how it changes and what to do next. That is repetitive work, which is what tools are for.

Citeably's free check runs a set of buyer-style questions across assistants for you, shows the real answers with your name highlighted, who was recommended instead and which sources were cited, and ends with a fix list. Paid plans repeat it weekly so you can see whether what you changed worked.