Build log · collimer
Can AI mention your brand and still describe someone else? Yes
Yes. An AI engine can name your brand and describe a different company in the same sentence, and a visibility score counts that as a win. In one case we measured, a brand had 57 classified mentions and only 70% of them were about the right business: 8 were confused with another entity and 6 could not be resolved. That is why we built a detector that reads each mention for who it actually describes, not just whether your name appears.
What are the two failures a mention count hides?
They look the same in a score and have different fixes.
- Entity independence. The engine names you but describes you as a feature of a parent company, or puts you in the wrong category. That is a positioning problem: the wider web tells your story in someone else’s terms.
- Identity accuracy. Some of your mentions belong to an unrelated business with the same name. That is a disambiguation problem, and a distinctive, findable description fixes it where a generic one cannot.
A pooled “60% visible” can mean either, or both, in any mix. It is the same trap as what a pooled score hides, applied to identity instead of query type. Both are different from getting no mentions at all, which is the zero that was not a verdict.
Who is most at risk of being described as someone else?
Any brand that shares its name with another business, a common word or a place. A roofing company, a law firm and a software tool can all be called “Summit”. Brands that are new, small or recently renamed are also exposed, because engines have less about them to go on and fill the gap with whatever is better known.
How does the detector decide?
A judge model reads each answer and classifies every mention as matching the brand, confused with another entity, or unknown. The detector flags when the confused-or-unknown share is above 15%, and only when there are at least 5 classified mentions. Below 5 it says nothing, because a thin signal is more misleading caveated than suppressed.
| Case | Classified mentions | Matched | Confused | Unknown | Flag? |
|---|---|---|---|---|---|
| The measured brand | 57 | 70% | 8 | 6 | Yes (30% > 15%) |
| Fewer than 5 mentions | under 5 | any | any | any | No, silent |
When it flags, the report recommends a fix: a short, distinctive description of who you are, what you do and where, repeated on your site and the profiles engines read, so the engine has something to tell you apart by.
What did we learn about using AI to judge AI answers?
Check what the judge actually reads. Ours was flagging too many non-SaaS brands as confused, and the cause was not the brands. Long answers were being cut down before the judge saw them, so on a multi-brand comparison the sentence that described the target brand often never reached it. The judge was guessing from text about other companies.
It also read extra detail as a mismatch. When an engine’s answer was more specific than the description we gave the judge, the judge treated the difference as a contradiction. We fixed both: the judge now reads the whole answer and treats a more specific answer as corroboration. Two different companies with one name are still flagged.
The 70% case was measured before those fixes, so some of its 8 confused mentions may not be real. We have not re-run it.
What we’re still figuring out
- How many brands flagged under the old judge stop being flagged with the fixed one.
- Whether the 15% threshold and 5-mention floor are right. We have one published case behind them.
- Whether a second measured case looks anything like the first.
For agents: try this yourself
- Read for identity, not mentions. Ask five AI engines “what is [brand]?” and classify each answer as the right company, a parent company’s feature, a same-name business, or unclear.
- Check your own truncation. If you pass long text to a judge model, print the length you send and compare it with the budget you think you set.
How this was made
Drafted by the Chronicler from Collimer commits and a published guide, then edited and published by Chase.
See how the Chronicler works →Try this with your own agent
2 prompts you can hand to your own agent (or run by hand) to work with what this post documents. Edit the bracketed parts for your context.
Read for identity, not mentions
Ask five AI assistants "What is [brand]?" and paste each answer here: [paste]. For each answer, classify it as: the right company, described as a parent company's feature, a different company with the same name, or unclear. Quote the sentence that decided each classification. Then give the share that is not "the right company".
Check a judge model’s input length
Here is the code that prepares text for our judge model: [paste]. State the intended budget, then compute what length the code actually sends for an input of [N] characters. Flag any slicing, division or unit mismatch (bytes vs characters) that sends less than intended. More in Collimer Build
-
Which sources do AI engines cite most? Your greatest resources are within reach: your LinkedIn and your GitHub
-
How does Microsoft Copilot decide what to cite? It asks Bing a few words at a time
-
Should you measure AI visibility per product line? Yes, because one average can hide the product AI has never heard of