RadRank Methodology

RadRank measures how often AI assistants name your brand when people ask them about your category. That is the whole product. This page explains exactly how the number is produced, what it is, and — just as importantly — what it is not. We publish this in full because transparency is the only honest form of neutrality. If you can read how the score is made, you can check our work.

What AI Share-of-Voice (SoV) actually means

AI Share-of-Voice is the percentage of category prompts whose answers name you, relative to all brands named, measured per engine, then averaged across the engines that actually answered.

Concretely, for a single engine:

  1. We send a set of realistic, neutral category prompts (see "The prompt set"

below).

  1. For each answer, we extract every brand name the engine actually wrote down.
  2. Your SoV for that engine is: **the share of brand-mentions across all those

answers that are you.** If ten answers collectively name brands forty times and you are named six of those times, your SoV for that engine is 15%.

  1. We repeat this per engine and average across the engines that returned an

answer.

That is the entire calculation. There is no weighting by who pays us, no "adjusted" score, no proprietary multiplier. SoV is a count of real mentions in real answers, divided by the total real mentions, expressed as a percent.

The prompt set

RadRank ships a small, disclosed default prompt set for each category. The defaults are:

  • Short — eight words or fewer.
  • Realistic — phrased the way a person actually asks an assistant, not

keyword-stuffed.

  • Neutral — they ask the question; they do not lead toward an answer.
  • Disclosed — the exact prompts we run are shown to you. Nothing is hidden.

You can — and should — add your own real prompts. The closer the prompt set is to how your actual customers ask, the more useful the number. We do not claim the default set is exhaustive; we claim it is honest, repeatable, and the same for everyone in your category.

The engines, and honest engine stamping

RadRank queries multiple AI engines. Not every engine answers every time — keys expire, rate limits hit, an engine declines a prompt. When that happens, we tell you.

  • The result stamps engineMode and lists exactly which engines actually

returned an answer.

  • SoV is averaged only across engines that answered. An engine that did not

answer is reported as absent — it is never silently counted as a zero, and it never inflates or deflates your average.

  • We will never write "across 4 engines" if only 2 answered. The stamped count

always equals the number of real, non-null engine results.

If two engines answered, the page says two engines answered. This is a hard rule, not a preference.

Citation extraction: only names that are really there

When we read an answer to find the brands it names, we surface only names that appear verbatim in the answer text.

  • Every brand we attribute to an engine's answer is a literal substring of what

that engine wrote (matched after normalizing case and whitespace).

  • We never infer a competitor "that's probably also relevant." We never let a

language model complete or guess a likely-sounding brand. We never add a name the engine did not say.

  • If an answer names no brands, the result is an empty list — not a plausible

filler name.

This mirrors the load-bearing honesty rule in our sibling system's competitor-compare.ts (lines 28-32): a competitor signal is never fabricated; when something is not measured, it stays unmeasured. We would rather show an honest gap than invent a number — or a name — that makes anyone look more or less beatable than they are.

Show the receipt

Every score links to its evidence. For each engine and each prompt you can see:

  • the exact prompt sent,
  • the raw answer text the engine returned,
  • the brand names we extracted from it, highlighted in that text.

Nothing about the score is asserted without the underlying answer you can read yourself. If a mention is in your count, you can point to the sentence that earned it.

Sandbox / sample mode

If no engine keys are configured, or a daily spend cap is reached, RadRank does not invent a result to fill the gap. Instead it returns a clearly labeled sample (mode: 'sandbox'). Sample cards are visibly marked "sample" and fabricate no metric — there is no real-looking SoV or rank presented as a measurement. A sample shows you the shape of the report; it never pretends to be your report.

The honest fault line: synthetic vs. real

This is the most important thing on this page, so we state it plainly.

RadRank measures answers to a disclosed, synthetic prompt set. We do NOT claim access to a private stream of real user queries. We do not have a backdoor into ChatGPT's, Perplexity's, or anyone else's live traffic, and we will never imply that we do.

What RadRank gives you is a repeatable, transparent, apples-to-apples benchmark: the same prompts, the same extraction rule, the same engines, run on a cadence, for every brand in a category. That repeatability is the value. A number you can reproduce and audit is worth more than a number we ask you to trust.

When you add your own real customer prompts, the benchmark gets closer to your reality — but it is still a measurement of those prompts, and we will always describe it as exactly that.

In one sentence

RadRank counts, per engine and only across the engines that answered, the real brand mentions in real answers to disclosed prompts — shows you the receipts — and never invents a name, a number, or a query stream it does not have.