Unavailable Isn't the Same as Not Mentioned: A Better Way to Read AI Visibility Tests

Two very different failures that look identical

Imagine you ask a friend to check whether your restaurant is mentioned on a food blog. They come back and say "no." You'd probably ask a follow-up: did you actually read the blog, or did your phone die before you opened it? The answer changes what you do next. If they read it and your name wasn't there, you have real information. If their phone died, you have nothing — you just haven't checked yet.

AI visibility testing has the same problem, except the failure mode is quieter. A request to an AI system can fail for reasons that have nothing to do with your brand: the provider is rate-limited, a search-grounding layer didn't activate, the response timed out. If a tool doesn't distinguish that from a clean, completed answer that simply doesn't mention you, it's reporting a coin flip as a fact.

Why this needs three states, not two

A binary mentioned-or-not model breaks down the moment a request fails partway through. Growthract's AI visibility observations are recorded in one of three states instead: grounded, ungrounded, or unavailable.

A grounded observation means the request ran with a live, search-backed retrieval layer and returned an answer based on information the system actually looked up. An ungrounded observation means a usable answer came back, but without confirmation that a retrieval or search layer was involved — still a real response, just a weaker evidentiary basis for claims about what the system "knows" versus what it retrieved. Unavailable means the observation didn't complete at all. No answer, no evidence, nothing to interpret.

Only the first two states tell you anything about whether your brand showed up. The third tells you the test didn't run — full stop.

Why this distinction changes what you do next

If you're a marketer looking at a dashboard that says "not mentioned in ChatGPT," the natural response is to start rewriting content, adding schema, chasing citations — treating it as a content or entity problem. But if that result was actually an unavailable observation mislabeled as a clean miss, you'd be solving a problem that hasn't been diagnosed yet. You might fix nothing that was actually broken, or worse, convince a team that a real gap doesn't exist because the one test they saw happened to come back "fine."

This is also why a single test run, on its own, should never be treated as a verdict. Assistants can behave differently across sessions, and grounding availability can vary by provider and by moment. A pattern that holds across repeated, correctly-labeled observations is worth acting on. One ambiguous result is not.

What to ask any AI visibility tool you're evaluating

Before you trust a "not mentioned" result from any product — Growthract included — it's worth asking a direct question: when a request to an AI provider fails, what does your report show? If the honest answer is "it looks the same as a real miss," treat every negative finding from that tool with real skepticism.

This same principle — that a failure to observe something is not the same as observing its absence — shows up elsewhere in how we think about evidence. It applies just as much to a crawler that fails to load a page as it does to an AI assistant that fails to answer a prompt.

It's also part of why we separate real observed answers — whether retrieval-backed or captured from an actual assistant session — from simulated ones, which is a distinction worth understanding on its own. That's covered in Simulated AI Answers vs. Real Ones.

A couple of follow-up questions