Skip to content
AVE StudioFor marketing agenciesENHU

Cookbook

Does ChatGPT know your brand? A free twelve-prompt self-test

One answer to one prompt is not a measurement — the variance is the result, and it is the part people throw away.

Updated ·

The cheapest way to measure AI visibility is to ask yourself. It does not replace regular statistical measurement, but in an afternoon it tells you roughly where you stand and, more usefully, which layer your problem lives in. The key is prompt ORDER: ask unaided first, because that shows whether you come to mind at all, and only then prompt with your name — once you have supplied it, you are no longer measuring what you set out to measure. The second key is repetition. These systems are not deterministic; the same prompt can return a different answer run to run, so one answer is one sample rather than a result. Anyone who asks once, sees themselves, and relaxes has looked at a lucky sample. This recipe covers what to ask, how many times, and how to record the outcome so it is still comparable in two months. You do not have to assemble the twelve prompts by hand: there is a form at the end that generates them from your brand name and category, along with the scoring sheet.

The procedure

  1. Write down what you actually want to learn

    Do not start with "what does it know about me". Start with which buyer question you want to appear in. Write three or four sentences a real buyer would genuinely type — category, market, maybe a budget, but WITHOUT your name. That becomes the spine of the set. If you cannot phrase them, that is itself a finding: your positioning is not in a form a search engine or a model could grip.

  2. Ask unaided first

    The first six prompts are the unaided block: the category, the problem, the alternatives. This measures the thing that actually matters — whether you come to mind when nobody has prompted for you. Note who appears instead; that list is useful on its own, because it shows who the system treats as the category's defaults.

  3. Then add help, gradually

    The second six are the aided block: first narrowing by category plus market, then naming your brand outright. The goal is to separate "does not know you" from "knows you but files you elsewhere". If it describes you accurately once named but you never surface unaided, the problem is classification or evidence rather than knowledge. If it describes someone else even when named, that is entity ambiguity.

  4. Repeat, and vary the order

    Ask each prompt at least three times in separate conversations — not consecutively in one, because an earlier prompt influences the next. Vary the order between rounds too. This is the most commonly skipped step, and skipping it produces confident false statements in both directions: "we're in there" and "they've never heard of us" can each be a single sample.

  5. Score as a proportion, not a yes/no

    For each prompt, record how many runs out of how many you appeared in, and how: named, described without naming, or cited as a source. Those are not the same thing. "Named 3/3" and "named 1/3" are commercially very different while both look like "we're in there". Record the date and which surface you used — surfaces differ, and averaging across them produces a meaningless number.

  6. Repeat later with the same set

    A measurement is worth something when there is something to compare it to. Keep the prompt set unchanged and re-run it in a month or two, the same way, the same number of times. Change the prompts in between and comparability is gone — and the most common self-deception is asking "better" prompts next round and crediting the improvement to your own work.

Frequently asked questions

How many times should I ask each prompt?
At this free tier, at least three, per prompt, in separate conversations. That is enough to separate the obvious cases — the difference between 0/3 and 3/3 is real. It is not enough to claim anything about a smaller change between two dates; that requires computing an interval.
Why start without my brand name?
Because once you supply it you are no longer measuring what you intended. With your name in the prompt the model will talk about you — that measures whether it KNOWS you, not whether you COME TO MIND. Commercially the second is the interesting one, and only the unaided block measures it.
Which surface should I use?
At least one that shows sources and one that does not. The difference between them is the diagnosis: present on the sourced surface but absent on the other means you are findable but not held in memory — and those two states call for different fixes.
What should I do with the competitors that appear?
Write them down, because it is some of the cheapest market intelligence available. If the same three names return every round, the system treats them as the category's defaults. Look at which of their pages the answer cites — that shows what content this surface finds quotable.
Does this replace paid measurement?
No, and it is not meant to. A self-test is a snapshot you run, with your prompts, at a low repetition count. Regular measurement differs in that the prompt set derives from real demand, the repetition count supports a confidence interval, and change separates statistically from noise. What the self-test is genuinely good for is deciding whether going further is worth it at all.

Sources

  1. OpenAI — Bots (separating search from training user-agents). https://developers.openai.com/api/docs/bots
  2. Perplexity — Crawlers (the retrieval side of a sourced surface). https://docs.perplexity.ai/docs/resources/perplexity-crawlers