Setting Up a Brand Book So Your AI Visibility Numbers Mean Something

Every AI visibility metric starts with a simple act: deciding whether a response mentions your brand. It sounds trivial. It is not. A brand called "Atlas" will be confused with the mapping product, the mythological figure, and a dozen unrelated companies. A brand with a long legal name will be shortened. A product line will be mentioned without its parent. If the tool counts the wrong things, every downstream number, from visibility to share of voice to sentiment, is wrong, and nobody notices because the dashboards still look tidy.

A brand book is the fix. It is a short document, kept in the tool, that tells the analysis how to decide.

What goes wrong without one

False positives. The tool counts a response that mentions a different entity with the same name. Your visibility looks higher than it is.

False negatives. The tool misses a response that uses a nickname or a product name. Your visibility looks lower than it is.

Wrong sentiment. The tool scores the tone toward an entity that is not you.

Split brands. Parent and product are counted separately, or a merger left two names in play.

None of these are dramatic on their own. Together they make trends unreliable, and a trend you cannot trust is worse than no trend, because it invites confident wrong decisions.

What Promptwatch offers

Promptwatch includes brand aliases and a Brand Book whose analysis instructions decide what counts as a mention. It also supports brand merging and auto-discovered competitors, so the competitor list can include brands you did not add. The same MCP and API surface lets you read and update the Brand Book programmatically, which helps for agencies with many clients.

What to put in your brand book

Think of it as instructions to a careful analyst who knows nothing about your company.

The canonical name. The full name as you want it written.

Aliases. Short forms, product names, common misspellings, and former names. Add the ones you see in real responses, not just the ones you would expect.

What is the brand. One or two sentences describing the company, so the analysis can tell your brand from a namesake.

What it is not. Explicit exclusions: "not the mapping software," "not the unrelated consultancy." Specific exclusions do the most work.

Edge cases. For example: does a mention of the founder alone count? Does a mention in a list of competitors count? Does a mention in a negative context count? Decide, and write the answer.

Products. Which product names roll up to the brand and which are tracked separately.

Step by step

1. Collect real examples. Pull 30 recent responses where your brand might appear, and read them.
2. Note the mistakes. For each, would a human say it mentions you? Compare with what the tool decided.
3. Write the first draft. Cover the failures you found.
4. Add aliases. Include every variant you saw.
5. Re-run on the same sample. Check whether the decisions now match your own.
6. Repeat until the match is good enough for your purposes.

You do not need perfection. You need agreement with a thoughtful human on most cases, and a record of what the rules are.

Competitors need rules too

The same problem applies to rivals. If you track them under inconsistent names, share of voice is skewed in your favor or against you. Add aliases for the main competitors, merge duplicates, and review the auto-discovered list monthly. New competitors appear, and sometimes the most frequent name in responses is one nobody on your team mentioned.

Document changes

When you edit the brand book, note the date and the reason. A change in rules can shift your numbers without any change in reality. Without a log, someone will later wonder why the line jumped on a certain Tuesday.

If you can, compare before and after on the same period. A/B period comparison in Promptwatch helps with that.

How often to review

Quarterly is a reasonable default, plus any time you launch a product, rebrand, or merge. Also review after you notice an odd response in the dashboard. One odd response may be noise, but three in a row usually means a rule is missing.

For agencies

If you run many clients, build a template with the questions above, and fill it in per client during onboarding. Agency plans in Promptwatch include unlimited projects, so each client gets its own setup. Keep the brand book in the onboarding file, so the account lead and the analyst read the same text.

Related measurement checks

Once the brand book is settled, check the other foundations: a stable prompt set, labeled prompt types, and a country and language that match your market. Together with a good brand book, they make trends trustworthy enough to base decisions on.

The recommendation

Spend an hour on the brand book before you spend a month on dashboards. Promptwatch gives you the aliases, the Brand Book instructions, and brand merging to make mention counting match your definition. Take 30 real responses today, check them against what the tool decided, and fix the first three mismatches.