How to Check If AI Mentions Your Business (A Repeatable Method)

Ask ChatGPT, Gemini and Perplexity the same buying question twice — once naming your business, once without it. The branded ask tests whether the engine knows you and describes you accurately. The unbranded ask tests whether it recommends you unprompted. Run all 3 engines separately, in a logged-out session, and repeat the set weekly, because a single answer is a sample of 1.

Most owners find out they are invisible to AI by accident. Someone types their company name into ChatGPT, gets a confident and slightly wrong description back, and the obvious next question — what does it say when nobody mentions us at all? — never gets asked systematically.

That second question is the one that matters commercially. A buyer who already knows your name was going to find you anyway. The buyer worth measuring is the one who asks an assistant who to hire and gets 3 businesses back. Whether you are among those 3 is a measurable fact, and measuring it takes a method rather than a tool.

How do you check if AI mentions your business?

Build a list of 10 questions a buyer would actually type, then ask each one in ChatGPT, Gemini and Perplexity — first without your name, then with it. Record the exact wording of every answer, which businesses were named, and which sources were cited. That log, not any single reply, is the measurement.

The questions worth asking

The list is the hard part, and it is where most self-audits go wrong. People test the phrases they wish they ranked for rather than the ones buyers use. A workable set of 10 covers 4 shapes:

• Hiring questions — “who is the best roofing contractor in Tucson”, the closest thing to a purchase intent an assistant sees.
• Problem questions — “my roof is leaking after a hailstorm, who do I call”, how buyers phrase it before they know the category.
• Comparison questions — “X vs Y for commercial HVAC”, where a shortlist gets built.
• Qualifier questions — “which solar installers in Phoenix handle permitting”, where being named depends on published detail rather than reputation.

Run each in a logged-out or temporary chat. A signed-in session carries your history, and an assistant that has watched you research your own company for 6 months will name it back to you in a way no stranger ever sees. That single detail invalidates more DIY audits than anything else on this page.

What is the difference between a branded and an unbranded prompt?

A branded prompt names your business and tests accuracy: does the engine know what you sell, where you operate, and does it describe you correctly. An unbranded prompt describes the need instead and tests recommendation: are you named at all when a buyer has never heard of you. Only the second one predicts new demand.

The reason the two must never be averaged

Branded and unbranded prompts fail for unrelated reasons and get fixed by unrelated work. An engine that describes you accurately when asked directly has read your site properly — that is an answer engine optimization success. An engine that never names you unprompted has read your site and still declined to recommend you, which is a corroboration problem no amount of on-page work solves alone.

Blending them into one percentage produces a number that rises whenever you add branded prompts to the set. It is the most common way an AI-visibility report flatters the business paying for it.

Why do ChatGPT, Gemini and Perplexity disagree?

Because they read different sources. Perplexity leans on live retrieval and cites URLs, Gemini draws on Google’s index and Business Profile data, and ChatGPT blends trained knowledge with live browsing depending on the mode. The same question can return 3 different shortlists on the same afternoon, all correct for their own inputs.

The practical consequence for reporting

Keep 3 columns, never 1 average. A business can be named reliably by Perplexity — which favours pages it can retrieve and cite right now — while being absent from Gemini because its Google Business Profile category is wrong. Those are different invoices. Averaging them yields a middling score that describes neither problem.

Access is the precondition underneath all 3. If your robots.txt blocks the crawlers, the question of what an engine thinks of you does not arise. The user agents to check are documented by the vendors themselves — OpenAI publishes its 3 bots and Google documents Google-Extended — and a great many sites disallowed them in 2023 and 2024 to stay out of training data without realising the same rule also removed them from being recommended.

How often should you run the check?

Weekly for the unbranded set, monthly for the branded one. Model outputs vary run to run, so a single check is noise and a quarterly check is too slow to catch a drop. Repeating 10 prompts across 3 engines every 7 days gives 30 observations a week — enough to see a trend rather than a mood.

The log that makes the trend readable

Five columns are enough, and the fifth is the one people skip:

1. Prompt — verbatim, because a reworded prompt is a different experiment.
2. Engine and date — never merged across engines.
3. Named businesses, in order — position matters less than presence, but record both.
4. Your status — recommended, mentioned in passing, or absent. Three states, not a yes/no.
5. Cited sources — every URL the answer leaned on.

That fifth column is the most actionable thing the exercise produces. The domains an engine cites for your buying question are the pages it currently trusts more than yours, and they are frequently directories and review sites rather than competitors’ websites. That tells you where corroboration has to come from, which is a different task from publishing another page.

Can a free checker do this for you?

Partly. A checker automates the repetition and scores the output, which removes both the tedium and the temptation to read one good answer as proof. DaxReach’s own GAS Score methodology reports Search, Answer and Generative as 3 separate sub-scores from 0 to 100 instead of blending them.

The limits worth knowing before you trust a number

Any score, ours included, is a diagnostic rather than a goal. It compresses a messy sampling problem into one figure, and a figure invites optimisation of the figure. The full GAS methodology is published — including the tie rules and the exclusions — specifically so the number can be argued with rather than taken on faith.

Two honest limits apply to every tool in this category. Model output is non-deterministic, so a score is a snapshot of a distribution, not a fixed position. And no checker can tell you what a buyer did next; only your own booking data does that.

It is also worth knowing what does not move the needle here. Structured data is the most commonly sold AI-readiness deliverable and the evidence for it is weak — Google itself states structured data is not required for its AI features.

The options compared

Four ways to answer “does AI mention us”, with an honest case for each:

Most businesses should start in row 1 and stay there for 2 or 3 weeks. Running the prompts by hand is what teaches you which questions your buyers ask, and that list is the durable asset — every tool downstream is only as good as the prompts you feed it.

The measurement is not the work, though. Once you know which engines omit you and why, the fixes divide cleanly into access, structure and corroboration — which is how our AI Visibility service is organised, and how the terms in this piece are defined in the glossary. If you would rather see the approaches side by side first, the comparison page covers that ground.

Check it yourself before you buy anything. A method you ran by hand for a month is a far better brief than a report you cannot reproduce.