Methodology

Principles

• Nothing is invented. When a value is not publicly available the site says Not publicly disclosed or Unknown.
• Every meaningful value is stored as a fact with value, source, source type, URL, published / retrieved / last-verified timestamps, confidence and verification status.
• Estimates, community data, calculations and editorial seeds are labelled as such and never presented as verified.
• History is appended, never overwritten: metrics are time series, prices are dated observations, events are an append-only log.
• There is no single "best AI" score. Separate measurements are shown with their inputs.

Data integrity & bias controls

We treat reliability as something to check, not assert. Every ingest applies these controls, and anything that fails goes to the public review queue rather than onto the site as fact:

• Prices are checked against the official page. A plan is marked “verified” only when its exact price appears next to the plan's name on the vendor's own pricing page. Pages that redirect to another site, block automated access in robots.txt, or render prices with JavaScript are labelled “not checkable” - never assumed correct. Mismatches are reviewed against the page text and corrected from the official figure.
• No single-language bias. Public-interest signals sum Wikipedia views across English and 9 other major languages, so products popular outside the English-speaking world are not systematically under-counted.
• No keyword inflation. Hacker News discussion is measured with specific phrases for ambiguous names (“Cursor IDE”, not “cursor”), each story counted once.
• No double counting. When a product's Wikipedia title redirects to another tracked entity's article, the views are not attributed twice; suspicious step changes (e.g. article moves) are flagged on the product page.
• No single-signal winners. The coverage penalty stops an entity with one strong signal from outranking broadly measured ones.
• No paid placement. Rankings and inclusion cannot be bought. Companies can send corrections, which are published only when a primary source confirms them.
• Conflicts stay visible. When sources disagree, both claims are kept and labelled.

Known residual biases we disclose rather than hide: Hacker News skews toward developer tools; Wikipedia attention favours consumer brands; benchmark coverage favours models that Epoch AI has evaluated; and seed data for products without machine-readable pricing pages remains unverified until checked.

Where data comes from

Source adapters run in the ingestion pipeline (npm run ingest). Each adapter is isolated: failures are retried with exponential backoff, rate-limited per host, protected by a circuit breaker and reported without stopping the run. Official sources rank above independent press, which ranks above community data. Full list with health on the sources page.

Rankings

Rankings measure attention and adoption signals, not quality. For each signal we take the trailing window sum (daily sources) or the latest snapshot (GitHub, Hugging Face), log-scale it relative to the day's maximum (0–100), and take the weighted mean of the signals that are available. The result is multiplied by 0.4 + 0.6 × coverage, where coverage is the share of total weight with data - so an entity with only one strong signal cannot outrank broadly measured ones.

Because inputs are stored as daily series, every day's ranking is recomputed from history, making movement arrows (▲ 4, ▼ 2) reproducible. Snapshot-based rankings (developer, open source) only have history from the first day a snapshot was recorded - older values are never back-filled.

Enterprise presence is not ranked: no measurable, public, terms-compliant source is configured. Website traffic and app downloads are not used for the same reason; the Similarweb and app-store adapters are present but disabled.

Trend & momentum

Trend (7d) compares the last 7 days with the average week of the previous 28 days. Momentum (30d) compares the last 30 days with the 30 before. Both are weighted across Wikipedia views (40%), HN stories (30%) and SDK downloads (30%), each capped at −90%/+500% and requiring a minimum baseline (700 weekly views, 2 weekly stories, 5,000 weekly downloads) so tiny numbers do not produce huge percentages. "Why is this trending?" lists the per-signal changes as likely contributing signals; no cause is inferred.

Discovery score

A calculated metric for emerging entities, separate from popularity. Products launched within 18 months: 35% recency, 25% trend, 20% momentum, 20% event activity (60 days). Models released within 90 days: 50% recency, 30% capability (ECI), 20% event activity.

Verification & confidence

• Verified

  Read directly from a primary or structured source at the retrieval time shown.

• Reported

  Stated by a named source (company, publication, benchmark body) but not independently confirmed.

• Estimated

  A derived estimate - always labelled; never shown as fact.

• Community-sourced

  From a community-maintained reference (Wikidata) - generally reliable, open to error.

• Calculated

  Computed by our documented methodology from the underlying sources.

• Unverified seed

  Editorial seed awaiting automated verification (e.g. consumer prices).

• Unknown

  No value could be established.

Confidence (high / medium / low) reflects the source tier and extraction method. Low-confidence items (e.g. unclassified official posts, unmatched benchmark organisations) are routed to the review queue rather than published silently.

Freshness

• Verified <24h - Read from its source in the last 24 hours.
• Verified 1–7d - Verified between 1 and 7 days ago.
• Verified 7–30d - Verified between 7 and 30 days ago.
• Stale >30d - Not verified for more than 30 days - treat with caution.
• Not verified - No verification date recorded (editorial seed or unknown).

If a source fails, the last known value stays visible with its original verification date - it ages into "stale" instead of being silently presented as current.

Conflicting sources

When sources disagree (e.g. an editorial license vs the GitHub repository license, or country per seed vs Wikidata), both claims are kept, the page states "Sources disagree", and a review item is created. Where fields conflict, the source priority is Official > Filing > Benchmark body > Trusted news > Reference > Community > Editorial - applied only where that source type actually supports the field.

Price normalisation

API prices are normalised to USD per 1M tokens. "Blended" prices assume a 3:1 input:output token ratio (stated wherever used). Subscription plans are converted to a monthly equivalent (annual ÷ 12); one-time, usage-based and custom prices are never converted. Image, video and audio generation prices are not normalised until a structured source exists. The cost calculator is informational and excludes reasoning tokens, tool fees and discounts.

Event engine

Events come from official feeds, press feeds, model listings (new models, retirement dates), Epoch benchmark runs and change detection between ingest snapshots (prices, context windows). Articles are classified by rule into normalised types (MODEL_LAUNCH, PRICE_INCREASE, ACQUISITION, …), linked to entities by name and alias (short or ambiguous names require exact case), and clustered: articles within 72 hours with similar titles or the same primary entity and type become one event. The strongest source - official first, then earliest - is shown as primary. Article text is not copied.

Limitations

• Wikipedia views and HN stories measure attention, not usage; products without their own article lose that signal. Views of redirect titles are summed with the article so page moves do not create false jumps.
• HN phrase matching can include unrelated stories for generic names (e.g. "Cursor", "Gemini").
• Package downloads include CI and mirrors (PyPI figures exclude known mirrors).
• Model release dates fall back to the OpenRouter listing date when Epoch has no record; this is labelled.
• Consumer prices and capability lists are editorial seeds until automated verification of official pricing pages is added.
• Regional availability and privacy/compliance details are only shown where captured from vendor pages, and are labelled as vendor claims.