Cold probes vs. live-search citations
"Does ChatGPT know about my business" and "does ChatGPT cite my business when it actually searches" are not the same question, and a tool that only asks one of them is only half-measuring AI visibility. WUWI AI asks both, on a schedule, and keeps them as two separate readings rather than blending them into one number.
The cold probe
A cold probe asks a model about the business with no web access — it measures what the model already knows from training alone. It never asks the model about the business by name and then treats a correct answer as a win, because that guarantees a match and measures nothing; instead it asks the kind of open question a customer would ask, and checks whether the business comes up unprompted. A model answering from memory is not treated as a business being passed over — memory alone isn't the same failure as a live search that finds nothing.
The live-search probe
A live-search probe asks a question the assistant answers by actually searching the web, and checks what it cites. This is the one that behaves like traditional rank tracking, except the result is a citation or the absence of one, not a position number — and it's run on request rather than a schedule, because each one costs roughly a penny in model usage.
A worked example
Take a fictional business — Harborview Roofing, Tampa, FL. A cold probe asking "who does roofing in Tampa" might come back with Harborview mentioned, because the model's training data included it somewhere. A live-search probe asking "who handles emergency roof leak repair in Tampa" might come back citing two competitors and not Harborview at all, because whatever the model's live search actually surfaced didn't include their site. Both readings are true at once, and they point at different fixes — the first says the model has heard of the business, the second says the business isn't winning the specific, high-intent search that matters.
How often each one runs
Cold probes run on a schedule, since they cost nothing to repeat and don't depend on what's currently in a model's live search index. Live-search probes are run on request rather than continuously, both because of the small per-probe model cost and because live search results genuinely change day to day — running one constantly wouldn't make the reading more true, just noisier.
Answer & AI authority — 14% of WVGI — starts at a cold-probe reading (up to 75% covered) and only reaches its fuller range once live-search probes are added on top. Neither one alone is "AI visibility"; together they're the closest honest answer to what an assistant actually tells a customer.