What the index measures
The unit that ranks is a software product within a defined category. Brand pages roll up the published products connected to that brand; a brand is not inserted directly into a product leaderboard. Each result belongs to a promoted snapshot with a start date, end date, methodology version, score version, prompt-set version, roster version, and extraction version.
The index records visibility in answers produced for buyer-oriented category prompts. A high score means a product appeared often and prominently in the eligible runs for that category and platform. It does not establish product quality, customer satisfaction, market share, or suitability for a particular buyer.
Questions, rosters, platforms, and models
Each active category has a versioned prompt pack and a separately versioned, published product roster. Prompts focus on discovery and evaluation needs within that category. Runs are accepted only when their prompt and roster versions match the versions used by the snapshot, which prevents evidence from different cohorts being silently combined.
The public platform inventory currently supports Claude, Gemini, ChatGPT, Perplexity, Google AI Overview, and Google AI Mode when eligible evidence is available. Provider and model labels are retained with individual run records. Public rankings are presented at platform level and do not promise that a provider will keep one model version throughout a data window; model and retrieval changes are therefore a documented source of variation.
Cadence, window, geography, and personalization
The scheduled publication path is weekly and uses a rolling 30-day data window by default. A snapshot is not public merely because its scheduled date arrives: it must complete validation and be promoted. Pages continue to show the latest promoted evidence if a newer candidate fails validation.
Category runs are system-generated observations, not logged-in consumer sessions and not personalized recommendations. The public index does not claim geographic representativeness. Provider localization, model changes, retrieval changes, source availability, and run timing can all affect an answer, so results should be read in the context of the displayed window and platform.
Visibility-score formula
For an eligible category/platform cohort, mention rate is the number of eligible runs that mention a product divided by the total eligible runs. Position is weighted using the inverse square root of the product's observed answer position. That position score is normalized against the strongest position score in the same category/platform cohort.
Visibility score equals 100 multiplied by mention rate multiplied by 0.75 plus 0.25 times normalized position score. This gives frequency most of the weight while allowing earlier answer positions to distinguish products with similar mention rates. Products are ordered by visibility score; equal scores at six-decimal precision share the same rank.
- Mention rate = mentioned eligible runs / all eligible runs.
- Position score = average of 1 / square root of observed answer position.
- Visibility score = 100 × mention rate × (0.75 + 0.25 × normalized position score).
Eligibility and quality controls
Only succeeded, version-matched runs enter ranking calculations. Runs identified as Slate self-citations are excluded. Snapshot publication also requires an active category, a published roster meeting its minimum product count, a promoted or validated snapshot, finite bounded metrics, valid public entities, and sufficient eligible evidence.
Prompt answer pages apply an additional answer gate. A public answer must be complete, direct, externally cited under the existing policy, and free of unresolved entity markers. Ineligible answers are omitted rather than replaced with a generated placeholder. Public products, brands, categories, child pages, and prompt pages must also pass the existing public-page quality and route-readiness checks.
Samples, confidence, and variance
Pages expose the applicable window, snapshot, eligible-run counts, and source evidence where available. The current pipeline labels cohorts with at least 20 eligible runs as high confidence and smaller eligible cohorts as medium confidence. This is an operational evidence-volume label, not a statistical confidence interval and not a claim of causal certainty.
The current public methodology does not publish a variance estimate. Differences between platforms or periods can reflect model, retrieval, prompt, roster, source, or timing changes. Buyers should examine the underlying prompt answers and cited domains instead of treating small score differences as inherently meaningful.
Sources, attribution, and review
The evidence layer distinguishes observed AI answers, matched product mentions, external citation domains, vendor or product facts, and Slate-calculated ranking fields. A citation means the domain appeared with an eligible response; it does not mean Slate endorses every claim on that source.
Automated validation blocks nonfinite metrics, reserved slugs, unresolved workflow markers, stale route relationships, and insufficient evidence. Failed publication gates can create review tasks. Corrections are also reviewed against source evidence before a public record changes. Material updates receive a new publication timestamp through the existing snapshot or correction path.
Independence and limitations
Slate Index is not pay-to-play. Commercial relationships do not alter prompt inclusion, mention extraction, score calculation, rank, or publication eligibility. Editorial explanations must remain distinguishable from observed answers, vendor facts, and calculated metrics.
AI answers are probabilistic and can change. The index cannot prove product quality, universal buyer preference, or future assistant behavior. Missing products may be outside the published roster, absent from eligible answers, or withheld by a public-quality gate. Use the index as transparent directional evidence alongside direct product evaluation and independent research.