Methodology

How we measure the AI shelf.

Every chart on this site, every claim in our scanner, every entry in the Index — explained, sourced, and replicable.

1. What we measure

EvidenceSignal measures how AI engines — ChatGPT, Perplexity, Claude, and Gemini — describe and recommend supplement brands in response to real shopper queries. For each query, we capture: which brands each engine cites, the exact language used, the position in the answer, and the substantive claims attributed to each product.

We then check every claim the engines make against the largest indexed corpus of supplement regulatory enforcement in the category, classify it under the DSHEA structure-function vs disease taxonomy, and surface any precedent that suggests it may be challenged.

2. The regulatory corpus

SourceVolumeRefresh
FDA Warning Letters6,717 total · 222 confirmed supplementWeekly
NAD / BBB National Programs2,202 cases · 290 supplementWeekly
FTC enforcement actions271 cases · 115 supplement-flaggedWeekly
NIH DSLD label database214,780 labels (pipeline built)Monthly
TINA.org consumer callouts2,808 articles + 6,744 class actions (pipeline built)Weekly

3. Classification — the DSHEA taxonomy

Classification is performed by a prompt-based LLM classifier (v1) calibrated against a 26-case manual gold set. We run it against two engines (Gemini and GPT). On the gold set, agreement is 96.2%. Disagreement cases are flagged as gray-zone candidates for human review.

Every cited_claim entry in our database is annotated with the classification model, version, and disagreement flag. Replication tooling is available to anyone who wants to audit our scoring.

4. The Index

The Supplement AI Shelf Index aggregates citation share across all four engines for the two hundred most-prompted supplement shopper queries. For each query we record which brands each engine cites, position in the answer, and substantive context. We compute a per-query citation-share-weighted score (position decay: 1.0 / 0.7 / 0.5 / 0.35 / 0.25 / 0.18), aggregate equal-weight across engines, normalize to a category baseline (100.0 on April 1, 2026). Rebased nightly at 11:59 PM ET.

What the Index does not measure

5. Cross-validation

6. Limitations — what we cannot do

7. Replication

We publish per-source extract code, classification prompts (v1), and our gold set on request. Contact us for the bundle.

Three reproducibility commitments: every chart shows source + date range; every methodology change bumps the version number; disagreement cases are reported alongside confident calls.

8. Version

Methodology v0.1 · published May 24, 2026.