How to tell whether AI is actually recommending your brand

The short version

To find out whether AI recommends your brand, ask several assistants the same questions a potential customer might ask. Record whether your brand appears, what the answer says and which sources it cites. Repeat the more subjective questions in fresh conversations. A single appearance tells you what happened once, not how reliably you appear.

This method comes from how I check brands in audits, including cases where assistants disagree about which business a name refers to. The point is to keep the questions fixed while you compare the answers. If you keep changing the question until your brand appears, you’ve measured your ability to prompt for it.

Why can one AI visibility score mislead?

A combined score can be useful for spotting a broad change, but it can hide the problem you need to fix. One assistant may describe your business correctly while another confuses it with something else. An average won’t tell you which answer a customer sees.

The assistants also draw on different sources. In an April 2026 study, BrightEdge found that pairs of AI engines shared between 16% and 59% of their top 100 cited sources. That measures source overlap, not how often they recommend the same brands. I explain the distinction further in the 14% problem.

There’s another reason to look beyond a score: being named isn’t always good news. A client in the financial sector has a name similar to a popular software product. Different assistants have sometimes identified the wrong business. A dashboard that counts the name as a mention could miss that mistake entirely.

Which prompts should you run?

Start by checking whether an assistant knows which business you mean. Ask a simple question such as “Tell me about [brand]” or “Who founded [brand]?” Use the same wording across assistants. If the basic answer is confused, a detailed question about services or suitability is harder to interpret.

Then build a small, fixed set of questions around decisions your customers make:

1. Category: “What are some options for [service] in [area]?”
2. Comparison: “How does [brand] compare with [relevant alternative] for [need]?”
3. Suitability: “Which businesses would be good for [specific customer need], and why?”
4. Brand knowledge: “What does [brand] offer, and who is it for?”

The brand questions check whether the assistant understands you. The open questions check whether it brings you up without being told to. Keep those results separate: an assistant can answer a direct question about your business while never suggesting it to someone looking for help.

Write the open questions as a person would ask them. Someone looking for a family activity in Brighton might ask for something creative to do together, without mentioning bakeries. If a bakery offers in-person courses, that broader question may reveal whether the assistant understands the activity well enough to suggest it. A question that narrows from “bakeries in Sussex” all the way to the bakery’s exact offer is much less revealing.

How do you run the prompts across four assistants?

Run the fixed questions in ChatGPT, Claude, Gemini and Perplexity. I use these four because they cover much of the observable chatbot referral activity, as I explain in usage is not citation. They’re a practical starting set, not a claim that every customer uses them equally.

Start each prompt in a fresh conversation so earlier questions don’t supply hints. Keep the location and customer need identical across assistants. Save the date, exact prompt and answer. For each answer, record:

• whether your brand is absent or named
• where it appears in a list, if there is one
• what the assistant claims about the business
• which sources it cites, if any

For open questions, run the same prompt again in fresh conversations and count how many answers name your brand. Ask each question at least three times per assistant. Report the count alongside the number of runs, per assistant and per question. Don’t turn a handful of answers into a claim about every future result.

How much repetition matters depends on the question. A clear founding fact supported by several reliable sources may produce consistent answers. A request for the “best” local options is more subjective. The assistant may draw on different sources or offer a different shortlist each time. That’s why an occasional recommendation and a repeated one need different descriptions.

How should you read the differences?

Compare answers by question and assistant before you combine anything. First ask whether the assistant identified the right business. Then ask whether it named the brand for an open customer question. Finally, check whether the claims and cited sources support the recommendation.

Three outcomes need different responses:

Don’t treat every cited source as equally useful. A source can appear often without explaining why an assistant trusts it or whether it is relevant to your customers. Looking at the source itself helps you decide whether the answer rests on a clear business description, an outdated listing or a mix of both.

What should you do with the result?

Start with repeated factual errors. If an assistant keeps getting a basic fact wrong, check your website and the other pages that describe your business. Try to find the correct fact yourself. If it takes more than five seconds, making that information clearer is a sensible first investigation.

For an absent brand, look at the question before changing your website. Does it describe a need you genuinely meet? If so, check whether your pages and relevant independent sources explain that connection in terms a customer would recognise. My GEO service covers the wider work of making those facts clear and useful across sources.

Keep the prompt set and your recorded answers, and check again after a change. Re-run the whole set weekly or monthly so you measure a trend rather than a single result. The free audit offers a quick first look across assistants. For a closer investigation, I’d want to know which sources keep appearing over time and what makes them useful for those particular questions. A mention count alone can’t answer that.