Skip to content
Public · Methodology

How to read AI-visibility claims

Six questions any vendor in this category — including us — should answer before you trust their numbers.

AI-visibility tooling is young and its numbers are easy to inflate. This checklist is written to be fair to every vendor: each question has a real methodological answer, and a good tool — ours or anyone else’s — should be able to answer all six in public documentation. We answer each one below with a link to the mechanism, so you can hold us to the same standard.

1. Are the prompts real user demand or synthetic phrases — and does the tool disclose which?

Why it matters: A visibility number over prompts nobody asks measures nothing. Synthetic prompts can be useful for coverage, but only when they are labeled as synthetic.

Our answer: Every prompt keeps its provenance (GSC query, autocomplete, PAA, community, CRM persona, support snippet, manual, fanout) and the label travels into reporting. See how →

2. Is visibility a single number or a range? What sample size backs it?

Why it matters: Generative engines are non-deterministic. A single percentage without a run count and interval is a weather report from one thermometer.

Our answer: Wilson 95% intervals over the effective run count; below the minimum sample the label is "inconclusive", never a confident percentage. See how →

3. Are branded prompts mixed into the headline visibility number?

Why it matters: Asking an engine "what is <brand>?" and counting the echo inflates visibility. Branded and unbranded demand are different questions.

Our answer: Branded prompts are excluded from headline visibility and reported in a separate, labeled lane. See how →

4. When counters mix units — prompts, responses, runs, "checks" — what exactly was counted?

Why it matters: A big number built from prompts × models × days × regions can describe a small product. Units should be defined before they are multiplied.

Our answer: The plan contract itemizes active prompts, prompt checks per month, crawl pages, and credits as separate meters with published overage prices. See how →

5. What happens when an AI engine’s output format changes — are miscounted runs excluded or silently kept?

Why it matters: Engines change response shapes without notice. Without drift detection, a parser break silently becomes a "visibility drop".

Our answer: A schema-drift canary flags affected runs and excludes them from metrics, with the degraded reason stored on the run. See how →

6. Do the customer counts and usage stats on the homepage match the ones in the pricing FAQ and the case studies?

Why it matters: Inconsistent counts across a vendor’s own pages are the cheapest possible credibility check — run it before trusting their measurements of anything else.

Our answer: We publish no customer counts or usage statistics until real tenants exist, and our copy tests fail the build on fabricated counts. See how →

The standard we hold ourselves to

Every number in our marketing must trace to code, a public document, or a dated verification — and our build fails on fabricated counts, dead brand names, and outcome guarantees. If you find a claim of ours that fails this checklist, email security@nimble-seo.com and we will fix it or retract it.