How the Core AI Index is measured.

AI category memory: brand × category × week × model lane. Each week we ask a panel of large language models which brands belong in each Knowledge Information Map (KIM) electronics category, collecting tens of thousands of responses. A brand earns a category association when multiple independent models name it, consistently, across multiple weeks, at sufficient saliency.

The result is a substrate: per (brand, category, week, lane) cell, a record of whether the brand was surfaced and at what strength. Every published rank is computed from that substrate and nothing else.

Two views over the same substrate, each answering a different question.

Brands in Core are also in Full (by construction). Brands in Full but not Core typically rely on niche-LLM consensus — the major buyer-facing systems don’t reinforce the same association at the Core threshold. That divergence between views is the substrate truth, not a defect.

AI mentions component manufacturers, distributors, aggregators, and design-tool vendors — each in their own role. A single rank template would score DigiKey as if it manufactured microcontrollers. It doesn’t, so we don’t.

• manufacturer vs manufacturers — ST, TI, Infineon, Analog Devices. Live.
• distributor vs distributors — DigiKey, Mouser, Arrow, Avnet. Live (DigiKey #1 in Microcontrollers among distributors).
• aggregator vs aggregators — Octopart, FindChips. Live (Octopart #1 in 15 of its 25 ranked categories).
• eda vs design tools on an EDA-native category set — Cadence, Synopsys, Altium. Live (Cadence #1 in IC design tools).

The substrate is shared across all role views — same foundation mentions, same identity normalization, same thresholds. Only the ranking population changes. (Role-classification mechanics in the appendix.)

The substrate occasionally surfaces brand-category associations that don’t match the brand’s real product line. We do not hide them. We label entity types, and we let the Core-vs-Full divergence show the mess honestly.

The cleanest illustration from a recent week: PolymerSearch in tantalum-capacitors. The wider Full Substrate panel surfaced PolymerSearch (a polymer-search platform, not a capacitor manufacturer) at high share across 8 LLMs. The Core AI Index — restricted to ChatGPT / Gemini / Claude / Perplexity — does not reinforce that association at its 2-of-4-vendors threshold. Core renders KEMET at #1; Full shows PolymerSearch as the substrate truth it is. If we were massaging rankings, you would never see this.

Associations that persist even in Core are labeled by entity type and kept in the evidence layer rather than silently removed.

This connects the substrate to the buyer. Per category, we embed AI’s canonical description and the top Google buyer-search terms for that category, and measure the match: what fraction of real buyer searches use words matching AI’s language (embedding similarity, 0.55 cosine per keyword).

• AI ≈ buyer search — ≥55% of buyer searches match. No badge renders on rows.
• partial — 30–55% match.
• AI ≠ buyer search — fewer than 30% match. This is the red badge on category rows: buyers may not find what AI says about that category. The five-vantage panel on the category page shows both vocabularies side by side.

This is a substrate property of the category, not a per-brand score. Mismatched categories tend to be broad lexical terms (“memory”) where both vocabularies stretch across topics; aligned categories are tight lexical leaves (nand-flash).

The substrate is refreshed weekly. The observed week is stamped on every public page; when the substrate for the requested week isn’t shipped yet, pages render the most recent available week with an as-of stamp — we do not zero out priors.

Every published number traces to a named artifact identified by SHA-256. Public pages name the substrate kind and week in their footer; paid surfaces render the exact artifact, week, SHA, and threshold at the point of use, so any number a customer quotes is replayable.