PRAGMA™ · AI Visibility Institute™

How the AAIR™ score is established

A score is only worth trusting if you can see how it’s built. So we publish it. Here is exactly how PRAGMA measures whether AI names your business — the engines, the tiers, and the formula.

1. What AAIR™ measures

AAIR — the AI Answer Inclusion Rate™ — is the percentage of the real, high-intent questions your customers ask AI where your business is named in the top answers. It runs 0–100, like a credit score: higher means AI recommends you more often when a buyer is deciding who to hire.

It is not a guess and not “SEO.” It is a direct measurement: we ask the questions your customers actually ask, to the AI engines they actually use, and count how often you show up.

2. The engines we ask — in two tiers

We ask the same fixed set of buyer questions across a frozen panel of AI models. We label each by its harness — the runtime we read it through — because the same model can answer differently depending on how it’s run. That transparency is part of the method.

The LLM tier is where your customers are today — the cloud assistants they open on their phone and laptop. The SLM / on-device tier is where they’re heading, as AI moves onto the device itself. Most tools measure neither honestly; we measure both, and we tell you which is which.

3. The formula — published, not hidden

Two numbers, and they do different jobs. AAIR™ is the ruler — one factor, nothing else: were you named in the top answers, yes or no, across every engine in the panel. No tone parsing, no human judgement, no weights to argue about. That austerity is deliberate: every subjective factor is a factor nobody can audit, and an unauditable number can’t be a standard.

AIVI is the diagnostic. Where AAIR™ tells you whether you’re named, our fuller visibility index tells you why — it weights five measured signals, and we publish the weights so you can recompute our composite yourself from the subscores on our own live dashboard:

AAIR™ is the score we’d let an auditor freeze. AIVI is the instrument panel we work from. We don’t blend them, and we never report one as the other.

What the denominator leaves out — and why that’s the whole point. If an engine rate-limits or errors, that cell is a no-read, not a “no.” We drop it from the denominator and publish the count beside your score, so you always see named / scored and no-read as separate numbers. Scoring an error as an absence invents a low baseline — and then a fake improvement when it clears. We don’t do that, including to ourselves: in July 2026 we found our own scanner dividing by the panel size instead of by the cells that answered, and we restated our own published baseline downward.

No secret sauce in the number itself. We publish the factor and the arithmetic — what stays proprietary is the exact question corpus, so the score can’t be gamed.

4. Why your score moves — the two clocks

AI answers run on two clocks, and honest measurement respects both:

So a real lift shows first on the fast clock and compounds on the slow clock. Anyone who promises an overnight jump across every model is either gaming the ruler or doesn’t understand the mechanism.

5. Open the outcomes, close the recipe

We publish the outcomes — your score, every engine’s result, the formula, the timestamp — so anyone can check our work. We keep the exact prompt corpus proprietary, so the score stays a fair yardstick instead of a checklist to cram for. The measurement ruler is frozen: it never changes between your “before” and “after,” so any movement in your number is real movement in your visibility — not us moving the goalposts.

6. Reproduce it yourself — run a live sample

Neutrality you can't check is just a claim. So here's the part no vendor-controlled score will give you: the exact way we build the questions, and a real sample you can run right now. We ask the leading AI answer engines the buyer's real "who's best" questions for a category × metro — the way a customer actually asks. The full corpus stays proprietary so the score can't be gamed, but the method isn't a secret, and neither is a representative slice of it.

Try it — ask any AI engine (ChatGPT, Gemini, Perplexity, Google's AI) these, for the commercial-signage × Jacksonville cell we publish in the registry:

Record who each engine names, and how often. Then compare it to the dated receipt in our registry. Do that, and you've done exactly what we do — that's the whole point. A score you can run yourself isn't a leash; it's a measurement.

7. Why you can trust the number

AAIR™ is a versioned, dated, published benchmark methodology. It is being placed under the independent governance of the AI Visibility Institute™, a separately governed body — so the score is built to belong to the market rather than to any one agency. That separation is structural, not cosmetic: under U.S. trademark law the body that certifies to a standard cannot also sell the service being certified. Every scan is timestamped and reproducible. Third-party validation of the methodology is on the roadmap, so an independent party can reproduce your score from this published method and confirm it matches.