Methodology

Semantic stems

A semantic stem is one distinct buyer question. Different wording can express the same need. Atlas groups those paraphrases under one stem, preserves the exact prompts, and counts the buyer question once. Twenty rewrites remain one idea. The thesaurus does not receive voting rights.

Each provider answer is an observation attached to that stem. Reports distinguish the number of stems from the number of provider observations. The first describes buyer-question coverage. The second describes how broadly those questions were tested.

Ranking stems and brand diagnostics

A ranking stem asks for a recommendation within the declared lens. A brand diagnostic asks about one firm or one explanatory detail. Diagnostics can explain how a provider describes a brand, but they do not add to comparative share. Naming a firm in the question is not evidence that the firm won an open recommendation.

Every stem receives its role before the run. Atlas does not relabel an awkward answer after it arrives.

Missingness and incomplete evidence

Missingness

An expected usable observation is absent. This describes the evidence pipeline. It is not zero visibility.

Insufficient evidence

Too few eligible stems or observations survived the declared rules for a stable conclusion. Observed facts may remain visible, but the result is labelled directional or withheld.

Incomplete cohort

The intended firm set has not passed its identity, evidence, or eligibility requirements. The table is not a complete comparison of the declared market.

An observed zero requires an eligible stem and a usable response with no qualifying recommendation. It does not fill missing cells.

Paraphrases, refusals, and extreme responses

Paraphrases remain attached to their semantic stem. Refusals are preserved and classified rather than treated as negative recommendations. Extreme responses enter anomaly review. A reviewer records whether the answer remains eligible, is excluded, or is rerun under the existing contract.

The original evidence and the decision remain linked. Provider unpredictability is unavoidable. Quiet deletion is optional, so Atlas declines it.

What Atlas measures

Atlas measures AI recommendation behavior for a declared lens and date. It does not measure legal quality, establish causality, or promise remediation outcomes. A compatible before-and-after run can describe movement. It cannot prove why the movement occurred.

Versioning and reproducibility

Contracts, cohorts, prompt sets, and provider panels are versioned. Past snapshots are retained with their original inputs. Reproducibility means the run conditions and decision rules can be inspected and repeated. It does not mean a hosted model will emit identical prose forever.

Every methodology change receives a new version and a dated note. The note states what changed and whether longitudinal comparison remains compatible.