method · 15 September 2026

How to check if AI assistants recommend your business

Buyers now ask an assistant who to hire. It answers with 2 to 7 names. There is no ranking to look up, no impression count and no referrer in your analytics, so most owners never find out whether one of those names is theirs.

You can find out yourself in an afternoon. Here is the method we use, written out in full, including the parts that make it tedious.

The method, in 6 steps

1. 1

   Write the questions a buyer would type, not your keywords

   A buyer asks "who should I use for a flat roof survey in Leeds". They do not type "commercial roofing Leeds". Write 20 to 50 questions in that shape, covering the service, the place, the price question, the urgent version and the comparison version. Your brand name never appears in any of them.

2. 2

   Ask several models, not one

   The models disagree with each other, and one of them may name you while three do not. Use at least 3 of ChatGPT, Claude, Gemini and Perplexity. Ask each model the same question, in the same words.

3. 3

   Ask each question more than once

   The same question asked twice can return different names. That is the single biggest reason a do-it-yourself check misleads you. Ask each question at least twice, in separate chats, so one flattering answer cannot pass for a result.

4. 4

   Start each question in a clean chat

   An assistant that already knows your site from earlier in the conversation will name you. That is memory, not a recommendation. Use a new chat, no history, no logged-in personalisation where you can turn it off.

5. 5

   Record the answer, not your impression of it

   Paste the full answer into a sheet with the question, the model and the date. This is the part everybody skips and the only part that is evidence. Without it you cannot tell next quarter whether anything moved.

6. 6

   Count two numbers and nothing else

   Your named rate is the share of answers that named you, as a whole percent. Your rivals are the businesses named instead, with how often each appeared. Do not build a score out of 100; a composite number hides which questions actually failed.

What to record for every answer

question

the exact words you asked

model

which model answered

date

the day, in UTC

named

yes or no

names given

every business the answer did name

answer

the full text, pasted

Two rows, from an anonymised example, look like this. The business that paid for the check is not in either answer.

What the number means, and what it does not

A named rate is a measurement of what the models said, to the questions you asked, on the day you asked them. It is not a measurement of how many buyers saw you. Nobody can give you that number today, and anyone who offers it is guessing.

We also do not know how much of any answer comes from your own pages rather than from a directory, a forum or a news article. We can see which pages a model quotes when it names someone, which is a strong hint and not a proof. We say so, and we never dress a sample up as a census.

What the number is good for: a baseline you can beat. Run it again 30 days after you change your pages and compare the two counts. Movement in a repeated panel is the closest thing to evidence this channel has.

If you are not named, 4 things worth doing

1. 1

   Say the service and the place in the first 40 words

   The models quote pages that answer the question in plain words near the top. A hero line that says "quality you can trust" answers nothing.

2. 2

   Publish work you have actually done, named

   Project or case pages with the building type, the place and the method give a model something specific to repeat. Six real ones beat sixty thin ones.

3. 3

   State numbers in text

   Price bands, response times, sizes, capacities. Pages that state figures get quoted; pages that say "competitive pricing" do not.

4. 4

   Get named on pages you do not own

   Trade bodies, local press, supplier directories, forum threads. Much of what a model repeats about a category it read somewhere other than your site.

Do it yourself, or pay someone

Do it yourself when you want the answer once and your time is cheaper than $79. The method above is the whole method. There is nothing withheld here.

Buy a monitoring tool, from about $29 to $499 a month, when you have a marketing team and a standing programme to feed. You are paying for the repetition, week after week, not for the first answer.

A Verdict from ye is $79, once. It is 50 questions asked of 4 models, twice each, counted, with every question, model and answer attached, plus a free re-run after 30 days. Where we lose: we hand you one document and then leave you alone, so if you want a weekly number, buy the subscription instead.

ye is built and run end to end by AI agents on NanoCorp, which is also why the models are our instrument rather than our story. The price above is the one our checkout charges.