Methodology

74 active sources
Every number on this site is built from the pipeline described below. If anything looks off - or if you want to feed Navisca a new source - drop us a line via the link at the bottom.

Navisca aggregates public job postings from creative-tech studios - VFX, games, motion, tech art, arch viz - and turns the raw text into queryable signals: skills, roles, salaries, AI-tool mentions, geographic mix.

We don't infer anything we can't show. If the parsers can't extract a salary band, the cell stays blank. If a category has fewer than 5 postings in the window, the chart is hidden, not estimated. The whole approach is "read what's there, never invent."

Jump to: Sources Cadence Parsing Dedup Salary Translation Thresholds What we don't claim
01 · Where the data comes from

Sources

We collect only from public sources - studios' own careers pages and the public job boards they post to. Nothing behind a login wall, no bought or leaked lists, no re-resale. Every page we read is one the studio has chosen to make publicly visible.

We collect respectfully: we honour robots.txt, keep our request rate well below ordinary human browsing, and stop immediately if a studio asks us to. No leverage, no negotiation.

02 · How often the data refreshes

Update cadence

The data refreshes weekly. Every Monday an automated run:

The "this week" labels you see across the site always refer to the calendar ISO week of the latest pipeline run. Postings older than 6 months drop out of trend windows; they're still in the per-studio dossiers as historical context.

03 · Turning text into structured data

Parsing pipeline

Each posting is read by rules-based parsers, not AI guesswork - the only AI step is translating non-English postings (see below). We pull out:

Every value is tied back to the exact words in the original posting, so any number on the site can be traced to its source. When something is mis-read we fix the rules, not the individual number - so the logic stays consistent.

04 · One posting counted once

Deduplication

Studios often post the same opening in several places at once (their own careers page and a job board). Counting that as three postings would inflate every chart, so we match duplicates and count each opening once - by the board's own job id where there is one, otherwise by a fingerprint of the studio, role, date and description. Studio name variants ("Industrial Light & Magic", "ILM", "ILM (Lucasfilm)") all resolve to one studio.

Trade-off: we err on the side of under-merging - in rare edge cases we'd rather count one posting twice than collapse two genuinely different roles at the same studio into one. The drift is small and we re-check it regularly.
05 · Apples to apples on pay

Salary normalisation

All salaries are converted to USD per year for cross-country comparison:

Caveat: these are advertised salaries, not paid salaries. Many markets systematically under-report - Tokyo posts "Negotiable", London posts "Competitive", a lot of US states still don't require disclosure. We surface what was published; treat the matrix as a directional signal, not a contract.
06 · Reading non-English postings

Translation (Japanese / Chinese / Korean)

CJK postings are machine-translated to English before parsing. We translate only the job-description text - never invent or paraphrase responsibilities the posting didn't include.

The original-language source is preserved on each posting record (toggle the flag in the per-posting popover to swap between the English translation and the source). Skill / role parsers run against the English text but match a multilingual alias dictionary, so a posting that says "Houdini経験者歓迎" still resolves to Houdini even if the translator phrased it slightly differently each run.

Translation runs only on postings new to that week's batch; nothing is re-translated unless the source text changed.

07 · When we stay quiet

Statistical thresholds

Stats are hidden rather than shown small when the sample is too thin to be meaningful:

These thresholds are tuned conservatively and only loosened as data volume grows.

08 · What you should NOT read into this

What we don't claim

Found a number that doesn't match what you see elsewhere? Want to suggest a source we're missing?
We read every note.

Send us feedback