A founder runs their domain through a trial AI visibility product and gets back prompts about remote telecom sites, power distribution at cell towers, and a leaderboard where the fifth slot is a company that is not theirs — only spelled almost the same. The percentages look precise. The story is wrong.
This is not rare. It is what happens when measurement skips two steps that humans still have to own: defining which questions matter, and confirming which legal entity the system is actually talking about.
Entity collision: similar names are not the same company
Language models resolve names probabilistically. Two companies that share a stem — Ansentra and Asentria, Lumen Pay and Lumen Technologies, dozens of fintech names built from the same Latin roots — get conflated in answers and, downstream, in dashboards that parse those answers without a canonical entity ID.
Asentria is a real vendor in remote site management for telecom infrastructure. Ansentra is an AI Search Growth company for B2B buyers. They operate in unrelated categories. A mention rate that ranks “Asentria” when you typed “Ansentra” is not a weak result for your go-to-market; it is a misattribution. Acting on it — rewriting positioning, chasing telecom keywords, briefing the board on an eighteen percent share — wastes time and erodes trust in measurement.
What to check before you trust a number
Read the competitor list aloud. If a name is one edit distance away from yours but sells something you have never sold, stop. The tool is measuring a homonym, not your entity.
Wrong query sets: the category was invented for you
Many products ship with “suggested prompts” or auto-generated query libraries so the trial feels instant. The generator often infers industry from thin signals: domain age, homepage keywords, structured data from a template, or nothing at all beyond the brand string. The output can be a coherent vertical that has nothing to do with how your buyers search.
Telecom site monitoring prompts for a marketing services company are a symptom, not an edge case. The tool is answering a different market’s questions and comparing you to Vertiv or DPS Telecom because those names dominate retrieval for that topic. You were never in the running for those answers, and optimizing for them does not move the queries your ICP actually uses.
Signs the query set is not yours
- Prompts use vocabulary your sales team has never heard from a prospect.
- Competitors are incumbents from another industry, not the three names you lose deals to.
- The product will not show raw responses, denominators, or which model produced each row until you pay.
- Mention rates are round percentages with no query count and no date range.
What a defensible baseline actually requires
Whether you use software or a spreadsheet, the sequence is the same. First, write a canonical fact set: legal name, one-line category, primary geography, and what you are not. Second, build or approve a query set from how buyers ask — discovery, comparison, problem, and brand — in their words, not your internal category label. Third, pick competitors explicitly; never accept a leaderboard the tool chose without review. Fourth, run repetitions per surface, store verbatim responses, and label every aggregate rate with numerator and denominator.
Only then does “we appear in eighteen percent of prompts” mean something. Until then it is a screensaver.
What we do on our own site (and what we refuse)
We publish insights like this one as normal HTML pages with real headings — not hidden markdown blocks in the footer, not duplicate blogs for crawlers only. Hidden or keyword-stuffed footers do not disambiguate entities; they add noise and contradict the retrieval hygiene we audit for clients.
We maintain a curated llms.txt index and Organization schema with accurate descriptions. We leave sameAs empty until a profile URL is real and owned. We would rather show zero mentions on a correctly defined query set than a flattering score on the wrong one.
If you are evaluating a platform
- Force custom prompts before you draw conclusions from a trial.
- Export or screenshot raw answers for at least ten queries you wrote yourself.
- Search your brand plus each suggested competitor; confirm industry alignment.
- Treat homonym hits as a finding for entity clarity work, not as visibility performance.
- Ask whether the vendor separates mention of your string from mention of your company as a resolved entity — most do not.