Never name your own brand in a tracking question

If your test question names your company, the answer you get back is not a measurement. Ask an assistant "is Acme a good accounting tool?" and it will tell you about Acme, because you told it what to talk about. Ask "what is the best accounting tool for a small business in India?" and whether Acme appears is a fact about the world rather than a fact about your question.

Why the difference is so large

An assistant answers the question in front of it. Putting a name in the prompt does two things at once: it tells the model the subject, and it signals that you already have an opinion about it. Both push the answer towards the name you supplied.

We have seen this produce the strongest possible false positive, and it is worth giving in full because it is the reason this rule exists here at all. On 27 July 2026 two runs of the same 78 questions happened on the same day, on ChatGPT, for our sister company clawlaw.in. The first used a wrapper that named the brand, and it came back ranking that brand first on almost every question. The second named nothing at all. It put the company second by breadth, and absent altogether from the litigation due diligence questions it most wanted to win.

The first run was discarded. The only thing it had measured was our own prompt. Had we kept it, we would have reported category leadership on the day we were invisible in the part of the category that mattered commercially, and we would have believed it.

That gap is worth sitting with. It was not a small bias to correct for. One run said the product led its category and the other said it was missing from the part of the category that mattered most commercially. Both were produced on the same day by the same assistant.

What a blind question looks like

It is the question a buyer types when they have not decided yet. It carries the context a buyer would include, and it never carries the answer.

• Not "why is Acme better than Beta?" but "what are the alternatives to Beta?"
• Not "does Acme cover all High Courts?" but "how do I search cases across all Indian courts?"
• Not "is Acme good value?" but "how much should I expect to pay for this kind of tool?"
• Not "why do agencies choose Acme?" but "how do I choose between agencies for this work?"
• Not "is Acme suitable for a 50 person company?" but "what should a 50 person company use for this?"

Notice what survives in the blind version: the constraint. Size, country, budget, use case and industry all belong in the question, because a buyer includes them. It is only the answer that must be left out.

The five other ways a question leaks the answer

Removing the brand name is necessary and not sufficient. These are the leaks we see most often, and each one produces a flattering number that means nothing.

• A description only you match. "Which tool covers all tribunals and has a legal chatbot?" names you without naming you.
• Your own internal vocabulary. If you invented the category phrase, asking about it retrieves your pages by construction.
• A leading premise. "Given that most tools miss tribunal data, which one does not?" has answered itself.
• Your competitor named and you not. This looks blind and is not: it produces answers about that competitor and the alternatives to it, which is a different question from what a buyer asks.
• Context earlier in the same conversation. If you mentioned yourself three messages ago, the question is no longer blind. Use a fresh conversation for every question.

Why a leaked question is so hard to catch

A leak is hard to see because the question looks neutral on the page. The test is not whether a brand name appears. It is whether any competitor's customer could plausibly have typed the same sentence.

What our own runs showed is how far a genuinely blind reading can sit from a flattering one. The same 78 questions, on the same day, on the same assistant, produced category leadership in one wrapper and absence from the commercially important zone in another. That gap was not a bias to correct for. It was two different pictures of the market.

There is a second reason to be careful, which came out of our own audit in September 2026. Asked afterwards how many of the questions it had actually run a live search for, the assistant withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had searched each one. That happened in three separate batches. If an assistant's account of its own behaviour needs checking, an assistant's answer to a question you wrote carelessly certainly does.

The one place naming yourself is correct

A small, clearly separated set of brand questions is worth asking, because it tells you what the assistants believe about you: whether your pricing is right, whether they have confused you with another company, whether they are repeating something out of date. Those answers are useful. They are just not a visibility measurement, and they must never be counted in the same number.

Keep them apart, label them, and never let them into the rate you report. A mention rate that includes questions naming the brand is a number that can only go up, which makes it a number that means nothing.

In practice the brand set is often the more urgent of the two. An assistant confidently telling buyers the wrong price is doing active harm right now, while being unnamed is a missed opportunity. Fix wrong facts before chasing visibility.

What this rule cannot protect you from

A blind question set removes the bias you control. It does not remove the noise. The same blind question asked twice can return different companies, so a single reading of a blind set is still only one reading.

It also does not tell you whether the question is one your buyers ask. A perfectly blind question about something nobody searches for is a clean measurement of something that does not matter.

Common questions

How do I know my question set is genuinely blind?

Read each question and ask whether a competitor's customer could have written it. If the answer is no, the question probably encodes something about you. Then check that no feature combination in the question describes only your product.

How many questions do I need?

More than you would like. Ten tells you whether you are roughly absent or roughly present. A hundred or more is where a rate starts to be worth reporting. Going from ten to fifty helps a great deal, and going from a hundred to a hundred and twenty barely moves anything.

Can I include my competitors' names?

In a separate set, yes, and it is useful for understanding how you are positioned against one of them. It is not a visibility measurement either, for the same reason, and it must be kept out of the rate.

What if the assistant asks me a clarifying question?

Answer with the constraint a buyer would give and nothing more. If you find yourself supplying detail that narrows it to you, the original question was too vague and should be rewritten rather than rescued.

Should I use the same questions every month?

Yes, for the tracked set, otherwise you are comparing two different measurements. Add new questions to a separate set and bring them into the tracked one deliberately, with a note of when they joined.

Does any of this apply to the free scan on this site?

Yes, and it is the reason the scan will not let a question name the brand. A tracked question that names you is refused rather than corrected, because a number that can only rise is worse than no number.

What to do first

Take the question set you or your agency are already using and check every line for the six leaks above. In most sets we have seen, at least one question encodes the answer. Remove or rewrite those, run the set in fresh conversations, and record the search count next to every reading.