出典SOURCES

Sources — godzillarank.com data compilation

All figures retrieved 2026-08-31 unless a different date is given on the row. Aggregate scores drift. Nothing here should be treated as permanent.

1. Film metadata (titles, dates, directors, budgets, box office)

2. Critic aggregate scores

3. User / audience aggregate scores

Every row in scores.csv carries a region column (global-en / jp) so the two audiences can be weighted deliberately rather than blended by accident.

Letterboxd retrieval

Letterboxd never puts the average on the film page — it is served from letterboxd.com/csi/film/{slug}/rating-histogram/, which 403s to automated clients. Retrieved through an authenticated browser session, parsing the literal "Weighted average of X based on N ratings" string. This supersedes the ten second-hand values previously taken from a Collider article; an earlier direct fetch appeared to return 4.5 for the 1954 film, which was a misread of the fan count — the real figure is 4.07.

Douban: identified but not harvested

Douban rate-limited its search after ~15 queries (error_info: 搜索访问太频繁) and then escalated to a full anti-bot interstitial redirecting subject pages to sec.douban.com. The block is on the session, not the URL pattern, and did not clear.

The hard part is solved. data/douban_ids.csv holds the Douban subject ID for 39 of 40 films, resolved through Wikidata property P4529 — Douban's own search is a dead end and is not needed. harvest_douban.md has the finishing procedure: serial requests ≥2s apart, with a guard that stops on the first sec.douban.com redirect. Roughly two minutes of wall clock.

This matters more than any weighting choice. Douban rates the 1998 Emmerich film 7.2/10 on 88,657 ratings — above where it places Shin Godzilla, Godzilla 2014 and every MonsterVerse entry, against 20% from RT critics and 2.31/5 on Letterboxd. That divergence currently has zero influence on the ranking.

Filmarks retrieval

Matched on Japanese title + release date + runtime via filmarks.com search cards. 4K-remaster re-releases carry separate Filmarks entries with their own, self-selected and usually higher scores (the GMK 4K remaster reads 4.1 against 3.5 for the original release) — those were excluded. Counts come from each page's own og:description (レビュー数：N件 ／ 平均スコア：★★★★X点), which IS in the fetched HTML — the .c-content__count element is a Vue placeholder, which is what defeated the first attempt. Trap: Filmarks exposes two different counts and labels both レビュー. JSON-LD's aggregateRating.reviewCount is written reviews only and is far smaller (Shin Godzilla 33,260) than the rating count the average is computed over (241,284). The og:description figure is the correct sample size.

3b. Awards and contemporaneous critical placement (data/awards.csv)

The Japanese critical split on Godzilla Minus One

Three Japanese critic polls, one film, one year:

Eiga Geijutsu is famously contrarian — the same poll placed Drive My Car 4th-worst in 2021 — so this is a house position as much as a verdict. But it is a real one, from real Japanese critics, and any model that reduces "the Japanese critical view" to one number will erase it.

Kinema Junpo's own database (kinenote.com) returns HTTP 403 on its search path, so Best Ten placements were sourced from Kinema Junpo's announcement pages and secondary Japanese coverage.

4. Single-source journalist rankings (data/rankings.csv)

Eight of these are full-franchise lists and feed critics_list_percentile_mean in master_table.csv: slant, indiewire, awardswatch, grunge, variety, paste, tomsguide, screenrant. Den of Geek (Showa-only) and Collider (partial) are carried in rankings.csv but excluded from that mean.

Retrieval note on Variety and IndieWire

Both are Penske Media properties and redirect automated clients to a TollBit paywall (HTTP 402) — direct fetch, headless browser, and the in-app browser all get gated. They were retrieved through the user's own signed-in Chrome, which is served the real page. Variety's article is a single page that rewrites the URL per entry via pushState (e.g. /lists/godzilla-movies-ranked/all-monsters-attack-1969/), so every entry is in the DOM at once. If you need to re-scrape either, use a real browser session; automated fetching will not work.

Ranking lists identified but still NOT retrievable

A Japanese critic list of the franchise does not appear to exist

Searched across Japanese film media, magazines and aggregators. Japanese film culture produces polls (ねとらぼ, みんなのランキング, 日本映画専門チャンネル's 2014 ゴジラ総選挙) and annual best-tens (キネマ旬報, 映画秘宝, 映画芸術) — but not the Anglophone "one writer ranks all 38 films" listicle, which is a format of the English-language film web. The nearest equivalents found:

Den of Geek MonsterVerse — ordering resolved

Previously excluded because two extraction passes disagreed on direction. Settled by reading the standfirst directly: "our staff votes on the definitive ranking of the American MonsterVerse films, from worst to best", with headings running 4 > 3 > 2 > 1 down the page. So their #1 is Godzilla: King of the Monsters (2019) and their worst is Godzilla (2014). Kong: Skull Island held rank 2 and is omitted from rankings.csv as a non-Godzilla film, but list_size stays 4 so the percentile maths remain correct.

5. Novelty / obscure statistics (data/stats_novelty.csv)

6. Caveats

Generated from gzconfig.py by build_caveats.py — edit the caveats there, not here.

Read the two lists differently. Defects are a worklist and should trend to empty. Constraints are permanent properties of the sources — they are instructions to the ranking algorithm, not unfinished work. A non-empty constraint list is not a sign the dataset is incomplete.

Open defects (0)

None. Every known error or unknown has been closed — see Resolved below.

Permanent constraints (5)

Nothing closes these. Design the model around them.

1. Two films have no Tomatometer at all. RT does not compute one below 5 critic reviews. The Return of Godzilla has 1 review and Godzilla Against Mechagodzilla has 4, so both carry a Popcornmeter and no critic score at all — tomatometer and critic_average_rating are both null. Never impute a zero: an absent Tomatometer is an absence of critics, not a negative verdict.
2. Two different RT audience populations, and RT publishes only one per film. The four MonsterVerse-era films carry RT's VERIFIED audience score; all 36 others carry ALL. Verified skews higher because it counts only confirmed ticket buyers. Confirmed unresolvable: RT exposes no ALL-type score for a film that has a VERIFIED one — the /reviews/all-audience page lists reviews but publishes no second aggregate. Treat the four MonsterVerse audience scores as a separate series, not as points on the same scale.
3. Metacritic drifts under you. Shin Godzilla reads 68/14 on Wikipedia, 67 in Screen Rant's 2022 piece and 66/15 on the live page today. Every score in scores.csv carries a retrieved date for this reason.
4. Showa-era yen figures are sparse and one is self-contradictory. data/japan_boxoffice.csv now labels every Japanese revenue figure as either 配給収入 (rentals, 1954–1995) or 興行収入 (gross, 1999–), so the two are no longer silently mixed. But only 4 of the 15 Showa films have a published yen figure at all, and Mothra vs. Godzilla's 15.5億 cannot be right — King Kong vs. Godzilla took 3.5億 on 1,120万 admissions, so 351万 admissions cannot yield 15.5億. It is flagged SUSPECT. For cross-film comparison of the Toho films use attendance, not yen — attendance is complete, internally consistent and cross-verified.
5. Filmarks n scales with recency, not quality. Filmarks launched in 2012, so Showa entries sit at 2,000–5,000 ratings against Shin Godzilla's 241,284. Godzilla 1985 rests on 194. eiga.com is worse: Tokyo S.O.S. is a 3.2 built from twenty reviews.

Resolved

• RT audience counts were banded — Resolved. likedCount + notLikedCount from the scorecard JSON IS the Popcornmeter's denominator — verified across 8 films, it recomputes every published score to within 0.5pp. sample_size now carries that exact number. RT's displayed band is a different, larger population and is badly misleading: Son of Godzilla displays “5,000+” against a real denominator of 676, and The Return of Godzilla's Popcornmeter rests on 28 ratings.
• Attendance was initial-run with no cumulative counterpart — Resolved. Cross-verified every figure against ja.wikipedia — all match, confirming they are initial-run. Cumulative figures for the 7 re-released films are now in data/japan_boxoffice.csv alongside them. Final Wars corrected to 1.04M; Minus One added at 5.03M.
• Japanese box office mixed rentals and gross unlabelled — Resolved into data/japan_boxoffice.csv, which labels all 30 films as 配給収入 or 興行収入. What remains is a sparsity problem, not a labelling one — see the open caveat.
• GMK's duplicate Rotten Tomatoes page — Resolved. RT has two entries; the hyphenated slug is the real one (65% off 17 critics, avg 5.60; audience 79% off 500+). The underscored slug is a near-empty duplicate with 0 critic reviews whose 96% was its AUDIENCE score. The dataset now uses the hyphenated page.
• King Kong vs. Godzilla's “disputed” attendance — Not a dispute. 11.2M is the original 1962 run; 12.55M is the official figure including re-releases. Both are correct and measure different things.
• Mothra vs. Godzilla's “wrong” gross — Not an error. ¥1.55B (ja.wikipedia) is 配給収入, distributor rentals; ¥3.2B (en.wikipedia) is estimated gross. Rentals ran roughly half of gross, so the two reconcile. Generalised into the open caveat above.
• Godzilla vs. Kong's two release dates — Confirmed: 2021-03-24 international, 2021-03-31 US theatres + HBO Max. films.csv uses the US date and the international date is noted.
• Letterboxd averages were second-hand — Resolved. All 38 canonical films plus 5 re-edits now carry Letterboxd's own weighted average and exact rating count from the rating-histogram endpoint.
• Filmarks had no sample sizes — Resolved. Every Filmarks and eiga.com row now carries an n, taken from each page's og:description.