Brand Book: Define What Counts as a Mention in AI Answers

Two analysts look at the same AI answer. One counts your brand as mentioned. The other does not, because the model wrote "the Amsterdam startup" instead of your name. Who is right? Neither, until you write down a definition. Most visibility disputes inside companies are definition disputes in disguise.

Promptwatch has a feature for this: a Brand Book whose analysis instructions decide what counts as a mention. Setting it up takes an hour and prevents months of arguing.

What goes wrong without one

• Name variants. Your product is "Acme Analytics" in docs, "Acme" in conversation and "AcmeHQ" on social. A tracker that matches only the legal name undercounts.
• Similar names. A common word, or another company with a similar name, produces false positives.
• Product versus company. Does a mention of your product count as a mention of the brand?
• Implied references. "The leading vendor from Berlin" may or may not count.
• Competitor confusion. If a rival's brand includes your name, you may inherit their mentions.

Each of these shifts your share of voice by a few points, quietly, and compounds across reports.

The parts of a Brand Book

Canonical name and aliases

List the official name, then every variant you want counted: abbreviations, product names, common misspellings, translated names for other markets. Promptwatch supports brand aliases, so these roll up to one brand in the reports.

Exclusions

Name what should not count: homonyms, a similarly named competitor, a product you have retired. Exclusions matter as much as inclusions.

Analysis instructions

Write plain-language rules that decide ambiguous cases. For instance: "Count a mention if the answer recommends, compares or describes the brand. Do not count mentions that appear only in a quoted user question." Short, clear rules work better than long legal ones.

Competitor set

List competitors with their aliases too. Promptwatch also auto-discovers competitors and supports brand merging, so you can clean up duplicates when two names refer to the same company.

How to write the instructions

Start with five real answers from your own panel. For each, decide yes or no on "mentioned," and note why. The reasons become your rules. Then test: could a colleague apply the rules and reach the same decisions? If not, tighten the wording.

Examples of rules that hold up:

1. A mention requires the brand or a listed alias to appear in the answer text.
2. A recommendation, a comparison and a neutral description all count equally as mentions. Track recommendation separately if you care.
3. Mentions inside cited quotes of third-party pages count if they appear in the answer text.
4. A reference to the product line counts if the product is on the alias list.

You can adjust. The point is that the rule exists and is the same every week.

Keep mentions and citations apart

A mention is the brand named in the answer. A citation is a URL used as a source. They overlap but are not the same thing, and a Brand Book governs mentions. Report them in separate columns. A brand can be mentioned without a citation and cited without a mention.

Changing the definition mid-stream

Sometimes you must change it. A rebrand, a product launch or a discovered false-positive source forces an update. When you do:

• Date the change in your log.
• Note the effect on historical numbers. Compare an A/B period before and after the change, which Promptwatch supports, and annotate the chart.
• Tell everyone who reads the dashboard. Surprise shifts erode trust faster than bad news.

Sentiment depends on it too

Promptwatch scores each response's sentiment from 1 to 100 and produces a weekly AI summary. If the definition of a mention is loose, sentiment gets calculated on answers that barely involve you. Tight instructions give you sentiment on the answers that matter.

Share of voice depends on it most

Share of voice is mentions of you divided by mentions of everyone in the set. Change the numerator rules or the competitor aliases and the percentage moves with no change in reality. If a competitor's board sees different numbers from a different tool, it is usually because the two tools have different definitions, not because one is wrong.

A one-page checklist

1. Canonical name and aliases, including local-language versions.
2. Exclusions and look-alike names.
3. Product versus company rule.
4. Treatment of implied references.
5. Competitor list with aliases.
6. Review date and owner.

Recommendation

Write the Brand Book before you build the dashboard, not after the first argument. Promptwatch puts aliases, analysis instructions and competitor merging in one place, and applies them to share of voice, sentiment and position in answer consistently. That consistency is worth more than any single feature, because a number your team trusts is a number your team uses. Set it up in your first week, review it each quarter, and keep a dated changelog.