Methodology.

For every reading we probe each model the same way, then read four signals off how it responds. The probes themselves, and how we tune them, are part of our calibration and aren't published. What is always open: the four signals below, the weights we apply, the composite math, and, on every reading, the actual answers each model gave, so you can see the evidence your score is built from.

Each model's reading is the weighted sum of four field measurements. The weights sum to 100%.

Per-model scores roll up into the composite using these product weights. They express our current view of relative discovery importance; they are not an independently audited measure of consumer traffic.

The weights are a published product assumption, not market-share telemetry. We renormalize the composite over whichever model slots are live at scan time, so an unavailable provider never flatters or punishes your reading.

For each model:

model_score = presence × 0.30
            + prominence × 0.25
            + accuracy   × 0.25
            + sentiment  × 0.20

Then the composite reading:

composite = Σ (model_score_i × model_weight_i)
          / Σ model_weight_i

(renormalized over models live today)

• NoteAccuracy is scored by a calibrated LLM judge, not a fact-check. It reads how concretely and confidently a model describes you — which correlates with the model actually knowing you — and the judge is validated against a hand-labeled rubric set before any methodology change ships.
• NoteSentiment is scored by the same judge, reading how the whole answer frames you. It cleanly separates positive from neutral from critical framing; sarcasm can still slip past it.
• NoteIdentity is verified only when you provide a role/field: a direct answer that clearly describes a different person with the same name is excluded from your presence and flagged in the reading. A name-only scan can't be verified — it's marked unverified rather than silently trusted.
• NoteCanonical facts you save on a subject are checked against each model's answer, and clear contradictions are flagged in the reading's transcript. Findings are receipts, not inputs — they never change the score.
• NotePer-model scores are stable within a week but not bit-identical. Treat the index as a calibrated reading, not a deterministic score.
• NoteThe familiar labels name model families, not the consumer products themselves. Consumer apps may add web search, citations, private system prompts, routing, personalization, geography, or conversation history. This benchmark controls the prompts and API basket; it does not claim product parity.
• NoteA reading is one timestamped sample in English, using the model basket shown above. It is not a confidence interval, a geography-wide survey, or proof of what every user will see. Weekly comparisons are useful for direction; repeated samples are required before claiming a small movement is durable.
• NoteAll seven slots are attempted on a full reading. If one is temporarily unavailable, its weight is renormalized out of the composite. The transcript and scan time are the receipt for which answers actually contributed.