How ClientRadar measures AI visibility: methodology

What gets asked, and how

Every prompt is a question a real customer might ask — “best dentist in Austin”, “emergency plumber near me” — with the business's city as the location and never the business's own name in the question, so an engine names whoever it would name for any customer asking from that city. A free scan asks three questions once on two engines (ChatGPT and Google AI Mode); paid plans ask up to ten prompts weekly across up to five engines depending on plan, with ChatGPT sampled multiple times per prompt so the result is a rate rather than a single answer.

Each engine is called through its own official API, never a scrape of the consumer app: OpenAI's Responses API with web search for ChatGPT, Google's Gemini API with Search grounding for Gemini, Perplexity's Agent API for Perplexity, and DataForSEO's SERP API for a recorded Google AI Mode or AI Overview result at the given location. Every result on this site carries a disclosure naming exactly which of these produced it.

How an answer becomes “mentioned” or not

Claude Haiku 4.5 reads each answer against the business's name, known aliases, domain, phone and address, with tolerance for punctuation, “LLC”, “&” versus “and”, and possessives — a domain match in a citation counts as a mention even when the name itself is paraphrased. A deterministic check runs first for the exact, unambiguous cases; the model is called only when that check is ambiguous or to extract a factual claim (a phone number, an address) to compare against the profile. Rank, when the answer lists businesses in order, is the position in that order; an unordered mention gets no rank.

The visibility score

For each engine and prompt, a mention rate (mentions ÷ samples, 0 to 1) is combined with a rank weight — 1.0 for first place, 0.8 for second, 0.65 for third, 0.5 for fourth or fifth, 0.35 for sixth or later, and 0.5 for an unordered mention — and averaged across prompts. The visibility score (0–100) is 100 times the weighted sum of that figure across every active engine, weighted by engine (ChatGPT counts for more of the score than Google AI Overview, for instance, renormalized to whichever engines the plan actually runs). Share of voice is a separate figure: the target's mentions divided by the target's plus tracked competitors' mentions, per engine and overall.

A run-over-run change of ten points or more in either figure is flagged as notable; everything else is kept in the per-engine history so a real trend is visible even when week-to-week noise isn't.

What this number doesn't claim

It is not a claim about what any specific customer will see at any specific moment — these engines vary between calls, and the APIs called here closely approximate but are not identical to the consumer apps, which add personal history, device location and interface features the APIs don't expose. It's a measured, repeatable snapshot, stored with its date, the model that produced it and its citations, so it can be checked rather than taken on trust.