How AI visibility tools count brand mentions (and why yours may read 0%)
What counts as a mention is a decision every AI visibility tool makes. A comparison of the approaches, the 0% name-matching bug, and Promptwatch's Analysis instructions.
Quick verdict
Before you compare two tools on visibility, ask how each decides that a brand was mentioned. The answer changes the number more than the engine does. A tool that matches on exact strings will miss "Acme" when your project is called "Acme GmbH." A tool that guesses will count a passing "also consider" as a recommendation. Promptwatch lets you write the rules yourself, in plain language, and fixed a related name-matching bug on 28 October 2026 (changelog).
The problem: same answer, different score
Imagine ChatGPT responds to a prompt with a paragraph that names your product once, in the middle, with a caveat. Is that a mention? A visibility metric must say yes or no, and every choice has consequences:
- Does the bare product name count, or only the company name?
- Does a misspelling count?
- Does a mention inside a negative sentence count?
- Does a competitor's product that shares a word with yours count against you?
- Does the model's passing reference to a parent company count?
Different tools answer these differently, often without telling you. That is why two dashboards can show different visibility for the same brand on the same day, and why a number that was stable can jump when a vendor changes its matching.
How tools approach it
There are roughly three styles.
Fixed matching. The tool compares answer text with the brand name and sometimes a list of aliases you supply. It is predictable and often wrong at the edges: legal suffixes, abbreviations, product lines.
Model-judged matching. A language model reads each response and decides whether your brand appears, with or without rules from you. More flexible, but only as consistent as the instructions it follows.
Rules you author. You write down what counts, and the tool applies your rules. This puts the decision where it belongs, with the team that owns the metric.
Our catalog does not record which style most vendors use, so we cannot place them on this scale. Where it is unknown, the practical test is the one at the end of this article.
What Promptwatch offers
Promptwatch's Brand Book holds brand aliases and an Analysis instructions card. According to the changelog, the card lets you write plain-language rules, up to 1,000 characters, for what should count as a mention of your brand and your competitors. Promptwatch applies them when it analyses new responses, so visibility, sentiment, share of voice and competitor rankings follow your rules. Responses that were already analysed keep their results.
That last sentence matters. If you rewrite your rules, history does not change. Plan to note the date of the edit on any trend chart, or you will compare two different definitions as if they were one.
On 28 October 2026 the changelog also records a fix: visibility no longer reads 0% when your project name carries a legal suffix like GmbH or Inc and AI answers use the bare brand name. If you saw zero on a project like that, it was a matching problem, not an absence from AI answers.
Why a score can read 0% or look too high
Work through these in order.
- Project name versus public name. If you registered "Acme Software Inc" and the world says "Acme," older systems may match nothing. The suffix fix addresses this case in Promptwatch. Check your project name against how people actually write the brand.
- Missing aliases. Product names, abbreviations and common misspellings need to be on the list.
- Overbroad matching. A common word as a brand name will match unrelated text. Write a rule that requires context, such as the product category.
- A competitor set that has drifted. If rivals are added or merged, share of voice moves without any change in your visibility.
- A prompt panel that never names anyone. Zero is correct when the engine answers with no brands.
Our guide on Brand Book and Scheduled Chats for weekly reporting shows how to write definitions down so they stay stable from one month to the next.
Writing rules that work
Short and concrete beats clever. A starting set might read:
- Count the company name and the product name as mentions of us.
- Count common abbreviations, but not the unrelated word that shares our spelling.
- Count a recommendation inside a comparison list.
- Do not count a mention that only appears in a cited page title.
- For competitors, apply the same rules to their product names.
These are examples to adapt, not a template that fits every brand. You have 1,000 characters, so spend them on the edge cases that confuse you, not on restating the obvious.
Comparison
| Question | Fixed matching | Model-judged | Your own rules (Promptwatch) |
|---|---|---|---|
| Who decides what counts | The vendor | The vendor's prompt | You |
| Handles a legal suffix | Depends on the tool | Depends on the tool | Fixed 28 Oct 2026 for bare-name answers |
| Changes apply to old data | Depends | Depends | New responses only |
| Competitors follow the same rules | Depends | Depends | Yes, per the changelog |
| Source | Not recorded in our catalog | Not recorded in our catalog | Changelog |
A test you can run on any tool
Pick ten prompts where you know your brand is named, plus ten where you know it is not. Run them in the tool, then read every response by hand and mark each as a mention or not. Compare your marks with the tool's. The miss rate on the first ten shows how often it undercounts. The hit rate on the second ten shows how often it overcounts. Do this before you trust a trend chart, and again if you change the brand name.
For a broader look at checking what AI says, our guide to checking whether your site is mentioned by AI search engines covers the manual side.
Who should pick which
- A tool with fixed matching if your brand name is unusual, unique and rarely misspelled.
- A tool with model-judged matching if you want less setup and accept that the logic is the vendor's.
- Promptwatch if your brand name is common, shared with a product, or carries a legal suffix, or if you need competitors judged by the same written rules. Analysis instructions turn a hidden choice into a documented one.
Bottom line
A visibility number is a definition with a decimal point. Know the definition, write it down, and change it on purpose. Of the tools we compared, Promptwatch is the one that puts that definition in your hands, and the one we would start with if your score looks wrong.