Brand Book and Aliases: Deciding What Counts as a Mention
Before any AI visibility number means anything, you have to define what a mention is. Here is how we set brand aliases and analysis rules for clients.
The least glamorous setup task in an AI visibility program is also the one that most affects the numbers: deciding what counts as a mention of the brand. We do it on day one, and we revisit it whenever the client launches something new.
Skipping it produces charts that look precise and are quietly wrong.
Why the definition matters
Imagine a company called Northwind Analytics, with a product called Compass. An engine writes "Compass by Northwind helps teams track pipeline." Is that a mention of the brand? Probably yes. What about "Compass" in an answer about hiking gear? No. What about "Northwind" alone, which is also the name of a sample database and several other companies?
Without rules, a tracker will either miss the first or count the second and third. Both errors distort share of voice, sentiment, and competitor comparison.
What we gather in the kickoff
We ask the client for:
- Every name the company is known by: legal name, trading name, abbreviations, former names.
- Product and feature names that buyers use.
- Common misspellings and alternative spellings.
- Names that collide with something else.
- Parent and subsidiary relationships and how they should be treated.
- Competitor names and their variants.
- Languages, since names can be transliterated or localized.
Most clients have never written this down. The exercise sometimes surfaces disagreements inside the client team about what the brand is called, which is useful to resolve early.
Setting it up in the platform
In Promptwatch we use two tools for this.
Brand aliases map alternative names to one entity. The brand, its abbreviations, and common spellings all count as the same thing. We apply the same to competitors, and use brand merging when the platform has discovered duplicates.
The Brand Book holds analysis instructions that decide what counts as a mention. This is where we write the nuance. For example: count the product name only when it appears with the company or in a software context; do not count a parent company mention as a brand mention unless the answer is about the subsidiary's offering; treat a mention in a list of "alternatives" as a mention, but read sentiment separately.
Writing these as clear instructions rather than keyword lists makes the analysis closer to how a human would judge it.
The edge cases we plan for
Ambiguous names. If the brand name is a common word, we specify context cues. Short and generic names get the strictest rules.
Product versus company. Decide whether a product mention counts toward the company. We usually count it, and report product-level separately.
Competitor mentions of us. When an answer names the client only to compare it unfavourably, it still counts as a mention. Sentiment handles the tone.
Mentions without a link. Counted as mentions. Links show up in citation data, not mention data.
Former names. After a rebrand, engines may use the old name for a long time. We alias it and watch the transition.
Local variations. For multi-country clients, brand names may differ by market. We add each variant and tag by country.
Test before trusting
After setup we review a sample of responses by hand. We read twenty or thirty, mark whether each was counted correctly, and adjust the instructions where the system got it wrong. This is a few hours of work and it protects every later report. We repeat the check after significant changes to the aliases or after adding new competitors.
Document the decisions
We record the definition in the client file. It includes the date, the list of aliases, the rules in the Brand Book, and who approved them. When someone later asks why a number changed in March, we can check whether the definition changed.
We avoid changing the definition mid-quarter without noting it in the report. If we have to, we annotate the chart.
Competitors need the same care
It is easy to be rigorous about the client and sloppy about competitors. If a competitor's name is spelled three ways and only one is tracked, their share of voice is understated and the client's relative position looks better than it is. We apply the same discipline, and Promptwatch's auto-discovered competitors list helps by suggesting names that appear in answers.
What clients get wrong
- Assuming the tool knows what the brand is.
- Treating the brand name as the only thing to track.
- Letting several people edit aliases without a log.
- Forgetting new products.
Revisiting the setup
We put a reminder in every quarterly review: new products, new names, new competitors, new markets. A small update here keeps the data meaningful as the business changes.
Our recommendation
Define the mention before you measure it. Write aliases and Brand Book instructions, test them on a hand-read sample, and keep a dated record. We do this in Promptwatch because the aliases, Brand Book, and competitor merging all feed the same dashboards. See the platform at promptwatch.com.