Your brand is five different companies to an LLM (and how to check in 10 minutes)

Most AI visibility advice is about being mentioned. That is the second problem. The first one is being mentioned as the wrong thing.

Here is the mechanic that people skip: a language model does not look up your brand in a table. At answer time it reconstructs a description of you from whatever fragments appear across its training data and its retrieval results. If those fragments disagree — different category, different audience, different founder, different spelling — the model does not resolve the conflict. It averages it, or it picks the most repeated fragment and discards yours.

So the question is not "am I visible?" It is "am I the same thing everywhere?"

Three tests you can run in ten minutes

Open ChatGPT, Claude, or Perplexity in a fresh session (no memory, no custom instructions) and run these. Write down the exact wording, not your impression.

Test 1 — Category stability. Ask What is <your brand>? five times in five separate sessions. Do not ask follow-ups.

You are looking for the noun. If it comes back as "a tool", "a platform", "a service", "a company", and "a website" across five runs, you have no stable category. Category is the single strongest signal a model uses to decide when to surface you. An unstable category means you are eligible for every query and selected for none.

Test 2 — Boundary test. Ask Is <your brand> good for <adjacent job>? — the job one step outside your actual scope.

A model with a clean model of you will say no, and explain what you actually do. A model with a muddy model of you will say yes and invent a reason. That invention is the tell: it means the model has no boundary for you, only a vibe.

Test 3 — Attribution test. Ask Who makes <your brand>? and What else does <founder or company> make?

Wrong answers here are the most expensive kind, because they attach your brand to someone else's reputation. This is also the fastest way to discover that your brand name collides with an unrelated product — common for short, generic names.

Why the fragments disagree

Almost always one of four causes.

1. Your pages describe you at different altitudes. The homepage sells a vision ("the future of X"), the product page lists features, the About page tells a founding story, and the pricing page names a category. Four writers, four categories. A model reading all four has to pick one, and it picks the one that appears most often in third-party sources, not the one you consider canonical.

2. Directory listings froze years ago. Every directory you submitted to captured a tagline on a specific day. Some of those directories are still what the model retrieves. If your tagline changed and the listings did not, you are permanently described by your earliest, least accurate sentence.

3. Structured data is missing or contradicts the page. Organization and SoftwareApplication schema with a description that differs from the visible copy is worse than no schema at all — you have now given the model two official-sounding descriptions that disagree.

4. No sameAs graph. If your site does not explicitly link your entity to its LinkedIn, Crunchbase, GitHub, and app store pages, the model has to infer that these belong to the same company. It often gets this wrong for brands with common words in their name.

The fix, in the right order

Order matters, because each step depends on the previous one.

Step 1 — Write one sentence and freeze it. Format: <Brand> is a <category> that <does the specific job> for <specific audience>. No adjectives, no superlatives, no "AI-powered". This sentence is now the canonical description. Every page, every listing, every profile uses a close variant of it — not identical (that reads as spam), but the same category noun and the same audience.

Step 2 — Make the category noun appear in the places models actually read. Title tag, H1, first sentence of the meta description, Organization.description, and the first line of your About page. Five slots, one noun.

Step 3 — Add the sameAs graph. A single Organization block listing every profile URL that represents the same entity. This is cheap and it is the highest-leverage structured data change most small brands have not made.

Step 4 — Re-run the three tests monthly, same questions, same model settings. What you are tracking is not the answer text, it is the category noun. When it stops moving across five runs, you have an entity. Until then you have a rumour.

What this looks like when it is broken

A pattern I see often: a tool whose homepage says "AI visibility platform", whose pricing page says "SEO audit plans", and whose directory listings say "ChatGPT brand monitoring". Three categories. Ask four models what it is and you get three different answers, all defensible, none stable. The fix is not more content — it is deleting two of the three categories.

That is the counterintuitive part of this whole discipline. AI visibility work is mostly subtraction: fewer claims, fewer categories, fewer synonyms for what you are. The brands that get cited consistently are boringly consistent.

I build a small tracker for the measurement side of this (bidvix), but the three tests above are free, they take ten minutes, and they will tell you more than a dashboard will. Run them before you buy anything.

Also in this series: The 20-minute AI visibility audit · The six kinds of pages LLMs actually cite · Your robots.txt might be hiding you from AI answers