How to measure whether AI assistants recommend your brand

When someone asks an AI assistant "which accounting software suits a five-person agency?" or "where can I get a reliable boiler service in Leeds?", the answer names a handful of brands. Whether yours is among them is now a question worth tracking. There is no Search Console for this, though, and the answers vary from one run to the next. Measuring it properly means treating it like a survey, not a single lookup.

Rule 1: never put your brand in the question

Ask "what do you know about Northfield Dental?" and the assistant will talk about Northfield Dental, because you asked it to. That tells you whether the model has facts about you, which is useful, but it is not visibility. Visibility is whether the assistant brings you up on its own when a customer asks a general question:

• Brand-free: "What should I look for in a family dentist in Northfield, and who is well regarded?"
• Brand-free: "Recommend a few project management tools for a small design studio."
• Not brand-free: "Is Northfield Dental good?" — useful as a separate reputation check, but keep it out of the visibility numbers.

Watch for indirect leaks too: a product name, your domain, a founder's name or a slogan in the question all count as mentioning the brand.

Rule 2: use several questions, written the way customers ask

One question is an anecdote. Five to ten questions that cover the main ways people look for what you sell give a picture you can compare month to month. Good sources for those questions:

• Search queries from your Search Console performance report, rewritten as full sentences.
• Questions customers ask on calls, in emails and in reviews.
• Comparison intent: "X vs Y", "best X for Y", "alternatives to Z" (a competitor's name is fine; yours is not).
• Location or segment where it matters: city, country, company size, budget.

Mix broad and narrow questions. Small brands rarely appear for "best CRM", but can appear for "CRM for veterinary clinics with SMS reminders" if that is genuinely what they do.

Rule 3: one model is one sample

Different assistants are trained on different data, use different search providers, and answer differently. The same assistant can also give a different list on a second run, or after a model update. So:

• State which model the result came from, and do not generalise from one assistant to "AI".
• Expect noise. A brand that appears in three of ten answers this month and two of ten next month has not necessarily lost ground.
• Look at trends over several runs rather than reacting to one.

What to record for each answer

Field	Why
Mentioned: yes / no	The headline number: mentions out of questions asked
The sentence that mentions you	Shows how you are described — accurately, vaguely, or wrongly
Position in the list	First of five is different from fifth of five
Who else is named	Tells you which competitors the model associates with the topic
Date and model	Needed to compare runs fairly

Match name variants when checking for mentions: the brand with and without spaces, common abbreviations, the domain name, and accented spellings. A mention of "northfielddental.example" counts.

Doing it by hand versus with a tool

By hand: open the assistant in a fresh session with no memory or custom instructions, paste each question, copy the answers into a spreadsheet, and fill in the fields above. Logged-in sessions, previous chats and location can all colour the answer, so keep conditions the same each time.

The AI visibility tool in DeskSEO runs the same procedure: you enter the brand, its variants and three to ten questions; it refuses questions that contain the brand, asks each one like an ordinary user, and records whether you were mentioned, the sentence, and your position in any list. It uses Anthropic's Claude model, costs $0.10 per question from a prepaid balance, and keeps every run so you can compare over time.

Reading the results

• Never mentioned, competitors always mentioned: the model does not associate you with the topic. This is usually a content and reputation gap, not a technical one.
• Mentioned, but described wrongly: the facts about you on the web are inconsistent or outdated. Find where the wrong version comes from — old directory listings, an old About page, a press article.
• Mentioned for narrow questions only: normal for small brands. Strengthen the niche first.
• Mentioned last in lists: treat position loosely; ordering is not stable between runs.

What tends to improve it

No vendor publishes how assistants choose which brands to name, so nobody can promise a lever. The things below are what is reasonable given how these systems are known to work: they learn from public text, and assistants with search fetch and quote pages.

1. Let AI crawlers read the site. If search-type crawlers are blocked, the site cannot be fetched or cited. Check with the free AI bot checker and read AI crawlers in robots.txt.
2. Put the content in the HTML. Many AI crawlers likely do not run JavaScript; see what crawlers see on JavaScript sites.
3. Answer the exact questions. Pages that plainly state who the product is for, what it costs, where it operates and how it compares are easier to quote than vague marketing copy.
4. Be mentioned elsewhere. Reviews, industry directories, comparison articles and press coverage on sites you do not control are part of the public record models learn from.
5. Keep facts consistent. Same name, address, prices and description across your site, profiles and listings.

Re-run the same question set every month or after a significant change, and keep the questions fixed so the numbers stay comparable.