How the Score works
The Score is a weekly, data-computed measure of cultural and commercial momentum. It is not a poll, a panel, or an editor’s list. Nobody decides it — a formula does, computed from public data, and it is the same formula for every person on the chart.
The five principles
Observable data only. The Score counts what verifiably happened on public platforms. No votes, no vibes, no editorial thumb.
Flows, not stocks. It measures what moved this week — not follower counts accumulated over five years. You can't sit on a number.
Measured in the present, validated against the future. The Score reflects this week's momentum, and the formula is judged by whether it predicts what happens next — not by whether it flatters anyone now.
The number never bends. No Score is ever adjusted by hand, for anyone, for any reason.
Rules public, thresholds private. Every rule on this page is public. The exact numeric tripwires inside the anti-manipulation rules are not — publishing them would be a manual for gaming them.
The five pillars
| Pillar | Weight | What feeds it |
|---|---|---|
| Consumption | 35% | How much of this person people actually watched this week: Twitch and Kick viewer-hours, YouTube long-form and Shorts views, TikTok views, Instagram Reel views, podcast plays — pulled from platform APIs and platform data exports. Raw attention, weighted by how much time each format really holds. |
| Engagement | 20% | How hard their audience engages per view: interactions, comments, and shares relative to plays, on TikTok and Instagram. Measured as rates, not totals, so a smaller creator with a rabid audience registers. |
| Momentum | 20% | The direction of travel: week-over-week growth of their consumption and demand, plus follower net-adds. A flow measure by construction — being big is not the same as moving. |
| Commercial | 15% | Disclosed commercial activity: ad-marked and sponsored posts, and tiered commercial events. Only disclosed work counts — it's read from public disclosure marks, not guesses. |
| Demand & Mentions | 10% | How much the world asked about them this week: press mentions (weighted by outlet), Google and YouTube search interest, Wikipedia pageviews, clip views about them made by other channels, mentions and appearances. Sourced from press feeds, search-interest data, and the Wikimedia API. |
Each pillar is compared against the person’s vertical, then combined as a weighted geometric mean — strength in one pillar cannot fully paper over absence in another.
What the number means. The Score is displayed as an index where 100 = a typical week for the vertical — every pillar exactly at its reference. 120 means a week twenty percent bigger than typical on the combined measure; 80, twenty percent smaller. The underlying computation and stored values are unchanged; the display divides the stored score by ten and rounds. Ranks always sort on the full-precision stored value.
The clock
A chart week runs Monday through Sunday, Eastern Time. The Score settles every Monday at 9AM ET. Each new chart is provisional for 48 hours while late-arriving platform data lands, then final. If a Score is ever revised, the revision is logged publicly — see the changelog below.
Anti-manipulation
Stated at the rule level; the numeric thresholds are deliberately private.
Winsorization. Extreme spikes are capped against a person's own trailing history, so one anomalous week moves the Score, not breaks it. Individual mention posts are view-capped the same way.
Velocity quarantine. Implausibly sudden surges are held out of the Score until they validate as real.
Per-account caps. No single account can dominate an input — a mentioning account contributes only its top posts each week; the rest are dropped.
Self-exclusion. Your own accounts can't feed your demand signals. Posts authored by any of your own handles never count as mentions of you. Demand means other people talking about you.
Promotion exclusion. Paid promotion doesn't count as organic demand. Disclosed commercial work is scored in the Commercial pillar — never as momentum.
Mentions are measured two ways at once. The mentions signal blends how much the mentions were watched with how many distinct accounts (above a minimum-views floor) did the mentioning — volume alone or a single loud account can't carry it. When a hashtag returns as many posts as the scrape requested, the counts are treated as a floor and flagged.
The number never bends. No Score is ever edited by hand — not for advertisers, not for friends of the house, not for anyone. If an input is wrong, we fix the input, re-run the formula, and log the revision in public. There are no exceptions.
Changelog
Every methodology amendment, newest first. Rendered from the public log — if the formula changes, it changes here.
- 2026-08-13: v0.2.8 mentions ENABLED in production via the two-stage publishing architecture — stage 1 publishes the chart on essential inputs (nothing can block it), stage 2 always runs the hashtag mention scrapes afterward at RESULTS_PER_TAG=50 with hard budget caps and enriches the same week's chart when it lands, or fails loudly without touching the published chart. First-enable receipts (who mentions lifted, and by how much) ship with every stage-2 commit as reports/<week>-mentions-impact.md.
- 2026-08-12: second publication-reliability incident — no chart shipped for weeks 2026-31 or 2026-32. Root causes: (1) the Apify account began returning HTTP 403 on every actor trigger on 2026-07-31 (account/billing side) and was never restored, so both Monday runs died at the scrape step, exactly as the publication guard intends; (2) the Monday external dispatcher was never created, so both runs fired hours late via GitHub's throttled cron; (3) no alerting existed, so two red runs went unseen for twelve days. Weeks 31–32 backfilled from data genuinely measured or archived for those weeks — Twitch/Kick viewer-hours from live snapshots (week 32 missing ~7 hours on 2026-08-06: lost hours are lost, not interpolated), Wikipedia pageviews and Google/YouTube Trends from their historical APIs. TikTok, Instagram, YouTube, press, mentions, and follower inputs for those weeks were not measured in-window and are absent, floored per v0.2.1a — both charts are provenance-labeled partial. Failure alerting (auto-issue), a client-side stale-site banner, and an audit system-health header added so silence is now impossible at every layer.
- 2026-07-22: first roster identity-integrity incident — six dead/squatted TikTok handles (Kai Cenat, Jynxzi, DDG, Rayasianboy, Danielle Pheloung, Jessica Wang) detected via ingest failure and corrected per founder verification; correction notes recorded on each roster entry, and the three distinctive new handles (pontiacmadeddg, rayasianboy, jynxziontwitch) added as query aliases.
- 2026-07-23: second identity-integrity audit — Instagram handles for 11 roster entries founder-verified via ingest audit; one correction applied (Brigette Pheloung acquiredstyle -> acquired.style, previous handle squatted/wrong), the other ten already matched the verified values in the roster.
- 2026-07-27: publication-reliability incident — the first scheduled Monday chart run never fired (GitHub cron throttling; no run in the Actions log at all, so no Apify scrape and no chart). Fixes: external cron dispatch is now the primary Monday trigger (in-repo cron demoted to backup); chart week resolved as the COMPLETED week (date -d yesterday — the old $(date +%G-%V) would have scored the hours-old new week on Mondays); a publication guard fails the workflow without computing or committing a chart unless same-day Apify consumption data landed (data/freshness.json stamps); the weekly audit report opens with a per-source freshness table. Week 2026-30 TikTok/IG rows re-ingested Monday 2026-07-27 from a fresh full-week scrape, replacing Thursday's mid-week test rows.
- v0.2.8 (2026-07-27): mentions and follower deltas go live. (1) mentions_by_others is now computed, not manual: TikTok + Instagram hashtag scrapes over each person's APPROVED mention_tags (excluded_from_scoring tags are never queried), ~200 results per tag, plus Instagram tagged-posts as an additional source when the actor is available (probed at run time; skipped loudly otherwise). Rules, public at rule level: chart-window filter; self-exclusion (posts by any of the person's own handles never count); per-post view winsorization; per-account weekly cap (top posts by winsorized views); score = 0.6 x log-scaled capped mention-views + 0.4 x log-scaled unique mentioning accounts above a minimum-views floor; results-cap floor flags when a tag saturates the scrape. Thresholds private. Input reclassified kind=ratio (recomputed weekly, duplicates are errors). (2) follower_netadds is now emitted weekly from consecutive fan snapshots (TikTok + Instagram from week 2026-31; YouTube subscriber snapshots begin 2026-31 via the existing channels call, joining deltas from 2026-32), summed across platforms, negative deltas pass through — falls count. Per-platform breakdown in data/follower_deltas/<week>.json feeds the site's "Followers this week" box, which appears only where a person-week has data.
- v0.2.7 (2026-07-23): Instagram consumption + engagement live via ingest/apify_instagram.py (manual Apify exports, like TikTok). ig_reel_views joins consumption (existing ig_reel_view coefficient); NEW ig-scoped engagement inputs ig_interactions_per_view_trimmed (weight 0.25) and ig_comments_per_view (0.10), existing three engagement inputs rebalanced x0.65 (interactions 0.39, shares_comments 0.195, chat 0.065). Reels are OWNER-ONLY in v1: records whose ownerUsername is not a roster handle are dropped. OPEN METHODOLOGY QUESTION — collab coauthors: 10.8% of scraped reels (95/881) were foreign-owned but carried a roster person as coauthorProducer; v1 rule is owner-only (no credit), revisit with a split-credit or full-credit rule once collab volume is better understood. ENGINE duplicate-row policy: every config input now declares kind=additive|ratio (documented in duplicate_row_policy); load_week SUMS duplicate additive rows (e.g. TikTok + IG disclosures feeding disclosed_partnerships_tiered) and raises a loud error on duplicate ratio rows.
- v0.2.6 (2026-07-21, founder-approved batch): (1) Verticals renamed: "Streamers" -> "Gen Z Streamers", "NYC Creators" -> "NYC Influencers". Verticals are now separate charts: rank_overall and all overall ordering removed from chart output (ratio math unchanged); Person of the Week and Riser of the Week are per-vertical (two crowns weekly, in chart "crowns"; riser is null until a prior week exists). (2) autocomplete_rank pillar weight set to 0 — still collected weekly for history. TODO: re-enable at 0.04 as a within-person delta vs a trailing 4-week baseline once history exists; cross-sectional autocomplete rank rewards name rarity, not demand. Its 0.04 redistributed proportionally: trends 0.1693, wikipedia 0.1017, mentions 0.2032, appearances 0.1355, sound_clip_reuse 0.0677, youtube_clip_views 0.1195, press_mentions_tiered 0.1406, youtube_search_index 0.0625. (3) query_terms upgraded to {primary, aliases} (founder-reviewed alias table); news and autocomplete query primary + aliases with cross-query dedupe; trends uses aliases only to assist topic-entity resolution (raw fallback stays the primary; entity-resolved people unchanged). DDG's primary is "DDG streamer" per founder hotfix; his news/trends/autocomplete rows re-ingested for 2026-30. (4) Weekly query audit: ingests write receipts to data/audit/<week>.json and ingest/audit.py renders reports/query-audit-<week>.md (terms, wikipedia article, trends entity/raw, top-3 matched headlines with sources, autocomplete prefix and position, topical-mismatch flags) as part of the Monday workflow.
- v0.2.4 (2026-07-21): two search-presence inputs join Demand & Mentions at modest weights. (a) youtube_search_index — Google Trends YouTube-search property (gprop="youtube") fetched alongside web trends with the identical anchor batching (Kai Cenat in every payload, anchor = 100). (b) autocomplete_rank — weekly probe of YouTube's public suggest endpoint: type the person's query_terms one character at a time, record the shortest prefix at which they surface and their position; score = 100 * prefix_share * position_share (surfacing at 1 char in slot 1 = 100, never surfacing = explicit 0; endpoint failure = loud skip, no row). Weights: youtube_search_index 0.06, autocomplete_rank 0.04; the seven existing inputs rebalanced x0.90: trends 0.1625, wikipedia 0.0976, mentions 0.1951, appearances 0.1301, sound_clip_reuse 0.0650, youtube_clip_views 0.1147, press_mentions_tiered 0.1350.
- v0.2.3 (2026-07-21): three live Demand & Mentions ingests join the weekly workflow. (a) wikipedia_views — Wikimedia REST daily pageviews summed over the chart week; name -> article resolved once (exact-title match only, null when no article exists, never guessed) and stored in roster.json as wikipedia_article. (b) NEW INPUT press_mentions_tiered — Google News RSS search per person using roster query_terms; articles published in the chart window, deduped by (domain, title), tiered by the whitelist in ingest/news_sources.json (major 3x, mid-tier 1x, unknown 0.5x). Weight 0.15 within the pillar; existing six inputs rebalanced proportionally (x0.85): trends 0.1806, wikipedia 0.1084, mentions 0.2168, appearances 0.1445, sound_clip_reuse 0.0722, youtube_clip_views 0.1275. (c) google_trends_index — now live via pytrends with anchor batching: every payload of 4 people includes the fixed anchor (Kai Cenat) and scores are rescaled relative to the anchor's mean for cross-batch comparability (anchor = 100); Trends topic entities used where the top suggestion exactly matches the person's name, roster query_terms otherwise; retries + loud graceful skips for pytrends fragility. All name-based ingestion reads roster query_terms (disambiguated search strings, founder-reviewed).
- v0.2.2 (2026-07-21): new Demand & Mentions input `youtube_clip_views` — weekly YouTube search per person by canonical roster name; sums views of videos published inside the chart window that are about the person (name must appear in title/description), excluding all channel IDs owned by roster members (resolved from roster youtube handles), deduped by channel with a per-channel contribution cap of 5,000,000 views/week. Log-scaled like all inputs (ln(1+v)). Weight 0.15 within the pillar; the five existing demand inputs were rebalanced proportionally (x0.85): google_trends 0.2125, wikipedia 0.1275, mentions 0.255, appearances 0.17, sound_clip_reuse 0.085. Ingest now fails loudly: resolved-but-quiet channels write explicit zero rows; unresolved handles emit a prominent error block in the Actions log instead of being silently skipped.
- v0.2.1a: pillar ratios use additive smoothing (delta = 10% of reference) and a 0.2 floor, so a zero pillar penalizes proportionally instead of annihilating the geometric mean. Momentum ratio clamped to [0.5, 2.0].
Think your number is wrong? The Score runs on public data — claiming your page will let you connect verified sources. (Coming soon.)