Methodology

AI Live Index polls every tracked channel once per minute and stores one observation per channel per minute: stream state (live / paused / offline / unknown), viewer count when the platform exposes one, scene and model labels. Tracking began 2026-09-03. Everything published on this site is derived from those raw observations — nothing is interpolated, estimated from press numbers, or backfilled.

“Observed Current Viewers” is the sum of measured concurrent viewers across channels that are currently live and whose platform exposes a public viewer count. It is a lower bound on real audience, not a market total.

Some platforms (e.g. Twitch) expose stream status but no viewer count. Those channels are live-audience unknowns: they count toward live-channel counts and supply metrics, and they appear in the denominator of audience telemetry coverage — but they are never treated as zero viewers.

Active Observation Coverage — are the channels we should be collecting actually being collected? Valid observed minute-slots ÷ expected minute-slots over the minute-polled population (active, verified and temporarily-missing channels). Channels delisted at their source are quarantined to 6-hour rechecks and are excluded from this denominator by design.

Audience Telemetry Coverage — of channels that were live, how many can actually be measured? Example: 9 / 11 live channels with audience telemetry. This measures source transparency, not collector health.

Each complete UTC day yields one canonical row. A channel's daily viewer-minutes is the sum of its observed CCV over live minutes with a numeric viewer count. From viewer-minutes we derive:

• Daily Mean Observed CCV = total viewer-minutes ÷ live viewer sample minutes (market-wide).
• Per-channel live average CCV = channel viewer-minutes ÷ its sample minutes.
• Top-1 / Top-3 share and HHI — concentration over viewer-minutes shares among channels with measurable audience. HHI is the sum of squared percentage shares (0–10 000).

Current-snapshot concentration (a point in time) and daily concentration (viewer-minutes) are different metrics and are always labeled separately. Concentration is descriptive — we publish no normative “healthy market” claims.

Raw observations found to be wrong are excluded from every derived metric but preserved in the database for audit. Example: the 2026-09-07/08 fal.live identity contamination, where one channel's telemetry was misattributed to others — 3,413 minutes were invalidated (see the fal.live data-integrity report). Derived values computed before invalidation were recomputed from raw telemetry.

Lifecycle is a separate dimension from stream status — it asks whether the upstream source entity is still publicly observable: verified → active → temporarily_missing → delisted (and reserved retired). Only deterministic source evidence advances lifecycle; a fetch timeout never marks a channel delisted. Delisted channels keep their valid history and are rechecked every 6 hours; they are excluded from current leaderboards and from coverage denominators.

All public core rankings — Most Watched, Peak Viewership, Most Reliable, Fastest Growing — cover the core generative-video categories: continuous generative video, interactive generative TV and agent-directed generative TV (plus unclassified channels, which are never silently dropped for missing metadata). AI Avatar Live channels (e.g. Neuro-sama) remain verified, tracked with full minute-level telemetry, accessible on their channel pages, and included in all-inclusive market analytics (all_ai_native_live scope, labeled as such on the Industry page) — but they never receive a rank in the default core leaderboards.

The Most Reliable ranking uses reliability_v2:

• 75% Technical Availability — live minutes ÷ (live + offline minutes). Paused minutes are deliberately excluded from the denominator: a paused stream is not broken infrastructure.
• 25% Outage Stability — 100 minus 20 per confirmed outage per tracked day, floored at 0.
• Observation coverage is a gate, not a score component — ranking requires ≥80% coverage (and the 7-day leaderboard additionally requires ≥90% coverage plus 72h of tracking history). Low coverage makes a score untrustworthy, not a channel unreliable.

A companion liveAvailabilityPct (live ÷ (live + paused + offline)) is published for transparency but does not enter the score.

Growth rankings compare the current 7 days against the previous 7 days per channel. A channel is eligible only when ALL of the following hold:

• tracked across the full previous 7-day window (newcomers are not “fastest growing” — they are pending evaluation);
• ≥80% observation coverage in both windows;
• measurable live audience telemetry in both windows;
• previous-window average ≥ 5 viewers (tiny baselines produce noise percentages, not signal);
• currently observable (not delisted at source).

Rankings cover the core generative-video scope; AI Avatar channels are tracked but do not receive core ranks. If no channel qualifies, the page says so instead of loosening the gates.

When we discover and add a pre-existing channel, measured audience rises because coverage expanded — not because the market grew. Daily series expose new-channel and delisted counts per day, and week-over-week claims are made on comparable cohorts only. The first fully comparable both-sides-clean week-over-week window completes 2026-09-23.