Visual Memory Index methodology

Source populations and definitions

The archive contains several distinct populations that are easy to conflate. The Index never mixes them, and each published metric names the population it counts.

Duration is the sum of published clip runtimes ( 299 hours). Because multiple clips can be cut from overlapping stretches of the same master, and the catalog does not store source-timeline offsets, deduplicated source duration cannot be computed responsibly. The Index therefore never describes summed clip runtime as unique archive hours.

These definitions are shared with the rest of the site through one canonical archive-facts layer, so the homepage, Archive Integrity, the machine-readable llms.txt surfaces, and this Index report the same populations under the same names.

Demand data: search log, filtering, normalization

Demand comes from the on-site search log: 5,145 searches between 2026-02-27 and 2026-07-14. What is and is not in that log:

• Only same-origin browser search events are recorded, which excludes most automated traffic by construction. The endpoint is rate limited. There is no session identifier, IP address, or user agent in the log, so per-user analysis is impossible by design.
• Queries are normalized before aggregation: lowercased, whitespace collapsed, length capped. The Index aggregates normalized queries; it never publishes them.
• Click-through events exist in the log but are excluded from this edition entirely: a distributed scraping incident in March 2026 contaminated the click stream, and there is no clean way to separate human clicks retroactively. Click analysis waits for a clean window.
• Web, API, and agent searches cannot currently be separated; the logged population is effectively the website audience. This is documented rather than estimated.
• A search counts as zero-result when the engine reported no matches at query time.

Subject taxonomy and query mapping

Version 1.0 of the taxonomy defines 20 subject themes. Each theme has two rule sets: word-boundary tokens that classify search queries, and a full-text query that defines the theme's matching supply in the catalog. Both are deterministic string rules, published in the codebase, applied offline when a release snapshot is built. No machine-learning model classifies queries, at build time or at request time.

• A query can match more than one theme; ambiguity is kept, not resolved.
• 19.2% of searches map to at least one theme. The rest, mostly place names and years, are reported as unmapped rather than forced into a category.
• Year detection uses four-digit years 1900 through 1999 plus decade words. Place detection matches query n-grams against the catalog's own country, region, and city vocabulary.

Privacy suppression

• Raw query text is never published, in any file, figure, or feed.
• Any published demand cell (theme, decade, or derived figure) requires at least 10 searches; smaller cells are suppressed and the count of suppressed themes is reported (1 in this release).
• The pipeline contains no customer identities, emails, IP addresses, user agents, session identifiers, board contents, or transaction identifiers at any stage.
• Licensing metrics publish clip counts and rates, not revenue. See the privacy policy for the site-wide rules this inherits.

Coverage Gap Index

For each theme: demand share is the theme's searches divided by all searches in the window; supply share is the theme's matching clips divided by all searchable clips. The Coverage Gap Index is demand share divided by supply share.

• Above 1: people ask for the subject more often than the archive can show it.
• Below 1: the archive runs deeper than expressed demand.
• Both inputs are published beside every index value, so a reader can always see the two shares behind the ratio. Themes below the suppression threshold get no index.

The index compares shares within one archive and one audience. It does not measure absolute market demand, and a small demand sample moves individual values; treat bands, not decimals, as the finding.

Licensing Index

The licensing basis is lifetime marketplace outcomes: a clip either has or has not ever been licensed. 3,606 clips (2.2% of the catalog) have at least one lifetime license. Per segment, the Licensing Index (reported as lift) is the segment's ever-licensed rate divided by that base rate.

• Lifetime data cannot be time-windowed and predates per-transaction records, so refunds and promotional sales cannot be filtered out of it; the figure is treated as a demand-proof signal, not accounting.
• Stockfilm launched direct on-site licensing in 2026. The program has completed fewer transactions than this report's minimum publishable cell size, so direct-license outcomes are not segmented in this edition.
• Revenue amounts are not published in the Index.

Confidence grades and missing data

Date and location coverage is high but machine-asserted: 98.7% of searchable clips carry a shot year and 98.9% carry a country, but most of those assertions are catalog-pipeline output that has not passed individual human review. The Index therefore publishes coverage shares (how many clips carry a value) and never accuracy claims (how many values are certainly correct). Stockfilm's clip-level confidence grading, including why nothing is auto-labeled "confirmed," is documented on Archive Integrity.

• Missing values are excluded from the relevant denominator and the coverage share is shown, rather than imputing values.
• Country and city are display-normalized strings, not ISO codes; "England" and "United Kingdom" style overlaps are resolved by the catalog's geo synonym layer for search, but counts of distinct names remain counts of names.
• 46.1% of searchable clips trace to a cataloged physical reel; the rest derive from masters whose reel-level paper trail is weaker. Reel provenance never uses synthetic grouping records.

Known biases

• Collection bias. This is one archive, acquired reel by reel, and 77% of located clips are American. The Index measures this archive, not the world's surviving amateur film.
• Survival bias. What was filmed, kept, and sold decades later skews toward prosperous families, celebratory occasions, and the physical durability of the film stock itself.
• Search-demand bias. Visitors search for what they already expect a home-movie archive to hold, and the site's own navigation shapes queries. Demand shares describe this audience, not the documentary market.
• Sample size. 5,145 searches over 138 days is enough for directional theme shares with the suppression rules above, and not enough for fine-grained trend claims. Ns are printed on the figures.

Corrections and version history

Releases are immutable. A correction produces a new version (for example v1.1) with a change-log entry stating what changed and why; the corrected date appears in the release metadata and the previous version's figures are superseded, not deleted. Errors can be reported through the metadata correction workflow or the contact page.