How to measure your brand's AI visibility: a free manual method

A common first check is to ask ChatGPT about yourself. That tells you what one answer said once, to one person, on one day. A measurement needs a fixed set of questions, repeated answers, a record and a denominator. This post is a procedure you can run with a browser and a spreadsheet, and the result is a table you can repeat next month and compare honestly.

The method follows the one the Dopbot Index uses on its category pages, scaled down to what one person can do by hand. The index asks each question through developer APIs. You will use the consumer apps, which is a real difference that API answers vs app answers covers. Perplexity's own FAQ notes that its API and app outputs may differ (as of October 2026).

What do you decide before asking anything?

Five decisions, written down first, because changing them later makes your results incomparable.

1. Your brand's names. The brand, common misspellings, the product lines, and the name an assistant might use for the parent company. Decide in advance which of these count as "your brand".
2. The categories. One category per run. A store selling water bottles and dog harnesses runs the procedure twice. Mixing categories hides which one is visible.
3. The assistants. At least two, and each is its own series. In October 2026 OpenAI and Perplexity shared 47.0% of recommended brands per category on average (excluding the alternative-to-a-brand question, 160 categories, 1,271 answers). That overlap is the brands both engines recommended, divided by the brands either recommended, averaged over categories with equal weight; OpenAI vs Perplexity has the detail. Being recommended by one assistant does not mean being recommended by the other: per category, on average, 3.9 brands were recommended by both engines and 5.3 by only one (excluding the alternative-to-a-brand question, 160 categories).
4. The country and language. Pick the market you sell in and use its wording and currency.
5. What counts as a recommendation. The index counts a brand only when the answer presents it as a pick. A brand named in a warning or a comparison is a mention. Log the two separately.

Which questions do you ask?

Use one question for each of five types. The index writes five prompts per category, one for each type below, so your results line up with the public pages.

Table: the five question types of the Dopbot Index, water bottles category, as asked in the October 2026 edition and as changed from the November 2026 update. The category page shows the latest update, so it may show only the November wording. Other categories have their own wording on their pages.

For your own set, use the November versions of the last two rows. The October fifth question named a brand, which is why the index replaced it with a buying-advice question that names none (see the methodology).

Three rules for writing them. Never put your own brand's name in a tracking question, because a question that names a brand is a test of that name, not of visibility. Write each as one question. Then freeze the wording: if you change a word, you have started a new series.

How many times do you ask?

Four times per question per assistant. That is 5 questions x 4 repeats = 20 answers per assistant, or 40 answers in all. The total, 40 answers, is what the index collects for each of its 40 updated US categories from November 2026 (5 prompts, 2 engines, 4 samples each; see the methodology). The arithmetic of why it is a floor and not a comfortable size is in how many answers you need before trusting a change. Allow about an hour in total if each answer takes a minute or two to ask and paste.

Why repeat at all? Because the same question can return different brands each time. A single answer is one draw. Four repeats of five questions give you a rate for those five questions on that assistant.

How do you ask in a clean session?

A clean session is one where earlier turns, saved preferences and your account history cannot shape the answer.

• One question per new conversation. Never ask the second question in the chat of the first. Earlier turns become context.
• ChatGPT (as of October 2026): use Temporary Chat and choose the unpersonalized option before you send the first message. OpenAI's help page says that option does not use memory, custom instructions or plugins.
• Perplexity (as of October 2026): use incognito mode. Perplexity's help center says memory and previous searches are always off in incognito.
• Same place. Run everything from the same city, with the same browser settings, and note the location in your log. Location can change search results (OpenAI's API documentation lets developers set an approximate user location for a web search), and we have not verified how the apps use it, so hold it constant.
• Same day. Collect all 40 answers within a day or two, so the underlying model and web results are as constant as you can make them.

Your answers will still differ from those of a logged-in customer with a history. That is a limit of the method, and the next section's log is where you write down what you controlled.

What do you record?

Open a spreadsheet with one row per answer and these columns:

Record answers that recommend no brand as rows too. In October 2026, 3.4% of answers (excluding the alternative-to-a-brand question, 1,271 answers) recommended no brand at all: 4.8% of OpenAI's 640 answers and 1.9% of Perplexity's 631. Leaving those rows out inflates your share.

Log source pages if the assistant shows them, but read them as what the assistant displayed, not as proof of what shaped the recommendation. In the index's data, for example, Perplexity's source lists are the pages its search retrieved, and the API does not say which of them the answer relies on.

How do you turn the log into a number?

For each assistant, count the rows where your brand is R, call it k, and the total rows, n. The share is k divided by n. Then add the 95% Wilson interval, which stays between 0% and 100% and behaves well with small counts. In a spreadsheet, with k in cell B2 and n in C2:

• Lower end: =(B2/C2+1.96^2/(2*C2)-1.96*SQRT(B2/C2*(1-B2/C2)/C2+1.96^2/(4*C2^2)))/(1+1.96^2/C2)
• Upper end: the same formula with + in front of 1.96*SQRT

Here is the layout on made-up counts for a hypothetical Brand A. These are not Dopbot data.

Table: a made-up report. The first two rows are usable. The third row shows why you should not slice a small sample by question type: 2 of 4 is compatible with anything from 15.0% to 85.0%.

The two assistants' intervals overlap, so from 20 answers each you cannot say that Brand A is truly more visible in one. Report the counts and the interval, never only the percentage.

The interval treats the 20 answers as independent draws, which they are not quite (four of them answer the same question). Read it as the least uncertainty you have, and as covering only these five questions.

What goes wrong most often?

• Asking once. One answer is an anecdote. Four per question is the floor.
• Naming your brand in the question. "Is Brand A good?" can return Brand A as the subject of the answer; we have not measured this pattern. A question that names a brand measures that name, not whether the brand is recommended unprompted.
• Mixing assistants into one number. They return different brands and different numbers of them. Report each.
• Reading budget and premium as one. In October 2026 the brands in a category's budget answers and premium answers overlapped by only 8.2% on average per category (brands shared by the budget and premium answers, divided by brands in either, over 160 categories, 155 of them with two budget and two premium answers; October 2026 edition). Your brand can be present in one and absent in the other, which is why the type column exists. See budget vs premium questions.
• Moving the goalposts. Rewording a question, adding a competitor to your brand names or switching mode between rounds makes the new numbers incomparable with the old.
• Counting mentions as recommendations. The share rises and the meaning changes.
• Forgetting the brand-name variants. A brand that comes back as "Acme" in some answers and "Acme Home" in others splits its count. The index keeps a hand-made list of merges for this and says so in the methodology.

How many brands should you expect in an answer?

Few. In October 2026, excluding the alternative-to-a-brand question, an answer recommended 2.3 brands on average and the median was 2 (1,271 answers, both engines). OpenAI averaged 2.2 across its 640 answers and Perplexity 2.5 across its 631. The index gives no benchmark for what share of answers one brand should reach, so compare your own rounds with each other, using the interval from the table above, and see what is AI shopping visibility for the definition and the October sample.

How often should you repeat it?

A monthly round is a workable rhythm, and it is the one the index uses from November 2026 for its 40 updated US categories. Keep the questions and the procedure identical, and compare the counts with the test described in the sample-size post, not by eye. If you would rather not run this by hand, the free snapshot shows how often assistants recommend your brand, who they recommend instead and which sources appeared in their answers.

Frequently asked questions

Can I measure with the paid apps and the free apps interchangeably?

No. We have not measured whether plans differ, so you cannot assume they match. Pick the plan you will keep using, write it in the Mode column and keep it constant, so that a change in your numbers is not a change in what you are asking.

Should I use my work account?

Not for measurement. A logged-in account with memory enabled carries your history into the answer. Use the temporary or incognito modes described above, or a logged-out browser window.

Is 20 answers enough?

For a rough picture, yes. For deciding whether something you changed worked, it is not: a rise from 6 of 20 to 10 of 20 is not distinguishable from noise. The sample-size post gives the exact thresholds.