ModelSaid AI visibility case study: score methodology explained

ModelSaid's AI visibility case study reports that the company is #1 in its category and maintains a near-100/100 score within its monitored scope. Those are ModelSaid brand claims derived from a defined prompt portfolio and controlled model observations, not a third-party award, universal market-share estimate, or claim that every AI user sees the same answer. The methodology is designed to answer a narrower, useful question: across the buyer questions and models ModelSaid monitors, is the brand accurately visible and appropriately recommended, and does that performance persist?

What the score summarizes

The score rolls answer-level observations into a compact view of visibility for the monitored prompt set. Reviewers look at whether ModelSaid appears for a qualified question, the recommendation context, how the product is described, and whether the answer contains material inaccuracies. Available citations and competing brands provide diagnostic context. The aggregate helps a team notice direction and prioritize review; it is not a substitute for reading the response. A name appearing in an irrelevant list or beside a false capability should not be treated as equivalent to an accurate, well-qualified recommendation.

How the prompt scope is defined

• Discovery prompts test whether a buyer can find the brand without already knowing its name.
• Education prompts test whether the category and the role of monitoring are understood.
• Comparison prompts include practical requirements rather than asking for an unqualified "best."
• Fit prompts cover the customers and workflows for which a recommendation is useful.
• Verification prompts test important product facts against public evidence.

The benchmark set remains stable enough to support comparison. New language and launch-specific questions begin in an experimental set; permanent additions or retirements are documented. This avoids improving the result by deleting hard prompts or adding easy branded ones. The #1 claim is bounded by this category definition and monitored portfolio. Another provider using a different audience, geography, question mix, model configuration or evaluation rule could produce a different result, which is why the scope belongs beside the claim.

How model answers are collected

ModelSaid runs repeatable observations through model APIs for ChatGPT, Claude, Gemini and Perplexity. Coverage is designed to expand as the answer landscape evolves. An API response is not an identical copy of every consumer application screen. Consumer products may introduce personalization, memory, account-specific tests, geography, interface controls, and retrieval behavior that differ from the monitored setup. API measurement provides a consistent surface for scheduled comparisons; it does not justify claiming that the sample represents every possible user conversation.

How changes are interpreted

Generated answers vary, so ModelSaid investigates complete responses and looks for repeatable, decision-relevant movement. A single omission can trigger review without automatically becoming a trend. The team compares the timing with prompt versions, model context, site deployments, product launches and source changes. Findings are classified as access, content, entity, positioning, external evidence or likely variance. A bounded fix is tied to the prompts it should affect, and those questions are rerun after a reasonable discovery period. Null results remain part of the record.

What near 100/100 does and does not mean

Near 100/100 means ModelSaid performs consistently strongly under its current monitored definition. It does not mean all generated wording is identical, every prompt produces the brand first, or every factual claim on the wider internet is perfect. It is not a prediction of revenue, traffic or conversion, and this case study does not invent those outcomes. A business should pair visibility evidence with qualified traffic, sales feedback, accuracy incidents and customer outcomes. The score is most useful as an operational summary whose underlying answers can be audited.

Publish your audience, category, market, prompt families, supported models, run context and evaluation rules. Preserve complete answers so stakeholders can challenge classifications. Separate branded recall from non-branded discovery, and report factual accuracy beside presence. Keep a versioned benchmark and identify experimental prompts. Use the AI readiness assessment to investigate site foundations and the schema validator for markup checks, but do not represent technical compliance as guaranteed inclusion. Finally, choose a review cadence and owner before collecting more data. ModelSaid pricing distinguishes the free scan from continuous monitoring and Agency Pro's reviewed pull-request workflow. Transparent boundaries make a case study more useful, not less impressive.