AI visibility: how to tell whether AI recommends your business

There is a number most businesses have never measured: out of the questions their buyers ask an AI assistant, how often does the answer contain their name.

For a lot of companies it is zero. Not low. Zero. And because nobody checks, nobody knows.

Why it is worth measuring

Buyers have started asking assistants for shortlists. Not everybody, and not for everything, but enough that being absent from those shortlists is now a real, silent loss. You do not see it in analytics. There is no referrer for a recommendation that never mentioned you.

Rank tracking will not catch it either. You can rank third for your main keyword and be named in none of the answers about your category, because the assistant read a directory rather than the search results.

The method

You need four things: a list of questions, a set of models, a record, and a schedule.

1. Write the questions a buyer would ask

Twenty is a good number. Rules:

• No brand names, including yours. You are measuring recommendation, not recognition.
• Buying intent, not research. Best X for Y, who should I use for Z, recommend a company that does W.
• Include your locations and your niches. Best moving company in Brooklyn, not best moving company.
• Include the objection questions. Cheapest X, most reliable X, X for small business.

2. Ask across the four assistants

ChatGPT, Claude, Gemini and Perplexity, with web search enabled. They differ substantially: in our own category one competitor is named in 40% of Claude's answers and 0% of ChatGPT's. Averaging across models is the only way to get a stable picture.

3. Record two things

Who was named. Every business, not just yours. You need the competitive picture or the number is meaningless.

What was cited. The sources listed under or inside the answer. This becomes your work list, because those are the pages that decide the answer.

4. Repeat weekly

Same questions, same models, same day of the week. The absolute number matters less than the trend and the gap to your competitors.

Reading the result

Here is a real example from our own market, twenty questions about Australian SEO services, asked across four assistants:

Three things to read from a table like that.

A gap between named and cited is a message. A business cited often but named rarely is being used as a reference rather than recommended. Usually it means the content is good and the commercial proof, reviews and third party listings, is thin.

Zero is not a mystery, it is a diagnosis. A business at 0% is almost always missing from the sources the assistants read. Not penalised, not disliked. Absent.

Per model differences tell you where to work. If you are strong on Perplexity and absent on ChatGPT, that is usually a difference in which sources each one favours.

What to do when your number is zero

In order.

1. List every source cited across your twenty questions. Count them. The ones appearing three or more times are your targets.
2. Get into the top five. Directory submissions, review platform profiles, an email to whoever wrote the listicle. Most of this is free and takes an afternoon each.
3. Fix your own facts. Make sure the basics of what you do, where, for whom and at what price exist as plain text on your site.
4. Make your review count visible on your own site and accurate everywhere else.
5. Re measure in two weeks. Directories get recrawled quickly and this moves faster than classic SEO.

The trap to avoid

Do not optimise for one question. It is tempting to find the phrasing where you do get named and report that number. It is also useless. The whole point of a twenty question set is that it averages out the luck.

Equally, do not panic at a bad week. Answers move. Look at four weeks before concluding anything.

Doing it without a tool

Everything above works with a browser and a spreadsheet. Twenty questions across four assistants is roughly two hours if you are quick, once a week.

Figo does it on a schedule and keeps the history, because two hours a week is exactly the sort of task that gets done twice and then quietly dropped. But the method is the valuable part, and the method is free.