Where every Atlas number comes from
Each dataset with its edition and refresh cadence, then the model constants and what the numbers leave out.
Sources
Two datasets are fetched and cached on disk; the rest are curated tables in the repo. When a fetch fails, the Atlas serves the last good cache.
| Dataset | Vintage | Refresh | What it feeds |
|---|---|---|---|
Unicode Consortium | cldr-json main branch (latest published CLDR) | Fetched, 7-day cache | Population, literacy, long-run GDP, and per-territory L1+L2 language shares: the base layer of every Atlas view. On a network error the last cached copy is served. |
World Bank Open Data | Most recent reported year per country (varies by country, year stored per record) | Fetched, 30-day cache | Indicator IT.NET.USER.ZS. Turns population into online population, the addressable-market figure the reach calculator and planner quote. Shown with the derived online population on every country page. |
World Bank Open Data | Most recent reported year per country (varies by country) | Fetched, 30-day cache | Indicator NY.GDP.PCAP.CD. Country pages (beside CLDR's long-run GDP, never blended), the compare table, and the GDP-per-capita lens in the explorer, coverage map, and opportunity matrix. A CLDR-derived figure fills in where the World Bank has none. |
World Bank Open Data | Most recent reported year per country (varies by country) | Fetched, 30-day cache | Indicator IT.CEL.SETS.P2, shown on country pages as subscriptions per 100 people. Values above 100 mean multiple SIMs per person, not full coverage. |
World Bank Global Findex | Most recent survey wave per country (the Findex runs every ~3 years, wave year stored per record) | Fetched, 30-day cache | Indicator FX.OWN.TOTL.ZS: adults with an account at a financial institution or a mobile-money provider, an upper bound on who can pay by card or mobile money. Shown on country pages and raised as a briefing risk below 50%. Never scored. |
Ethnologue (SIL International) | 2023 edition, rounded published estimates | Static, updated by hand | De-duplicated global speaker totals for the major languages, the cross-border figure CLDR's per-territory shares cannot give. Good for ordering, not for citation. |
W3Techs | Content-language survey, 2024 | Static, updated by hand | Share of surveyed websites per content language: the supply side, set against Ethnologue's audience side. |
EF Education First (Signum International AG) | 2025 edition (released November 2025, EF SET test data from 2024). 116 of 123 published countries, cross-checked against two sources | Static, updated by hand | English-proficiency score and band per country. Context only; never scored. Countries whose score differs between the official PDF and the Wikipedia table of the same edition are left out. |
CSA Research | 2020 survey of 8,709 consumers across 29 countries | Static, updated by hand | 76% of respondents preferred products with information in their own language; 40% said they will not buy in other languages. Background only: no Atlas number comes from it. |
Curated (industry-typical defaults) | 2025 defaults, repriced through the budget estimator's CSV export | Static, updated by hand | Five tiers from raw MT to transcreation, used by the budget estimator and the launch-cost model. A starting point for a budget conversation, not a vendor quote. |
Natural Earth | 1:110m admin-0 countries (177 features), bundled as a static file | Static, updated by hand | The country outlines behind both world maps. Geometry only: the Atlas reads no statistics from it. |
The EF EPI table ships 116 of the 123 countries in the 2025 edition: only those where the official report PDF and an independent transcription agree on the score. The seven where they disagree are left out.
The launch score
The launch planner scores each candidate market as a weighted blend of three components, each 0–100: incremental reach, GDP, and engineering ease. The product profile sets the weights:
| Profile | Reach | GDP | Ease | Rationale |
|---|---|---|---|---|
| B2C | 55% | 15% | 30% | Weights audience size highest: consumer products scale with user count. |
| B2B | 20% | 55% | 25% | Weights GDP highest: deal size tracks purchasing power more than headcount. |
| SaaS | 35% | 40% | 25% | Splits revenue and reach: self-serve products need both an audience and buyers in it. |
| Balanced | 34% | 33% | 33% | A flat prior for when the product profile is undecided. |
Reach (0–100)
Net-new speakers gained worldwide by adding the market's primary language, under the overlap model below; a half-covered market earns half credit. The scale is linear: 200,000,000 net-new speakers scores 100, and more saturates. The 200M anchor is a judgment that lets the largest realistic additions fill the scale; no market fact says 200M rather than 250M. "% of world" divides by a fixed 8.0 billion.
GDP (0–100)
The country's CLDR long-run nominal GDP, normalized so the largest economy in the dataset scores 100. It orders markets and does not forecast revenue. Scores compare within a run but not across datasets, because the maximum comes from the country list in play.
Engineering ease (0–100)
A checklist for the candidate language: start at 70 and apply each adder that fires, clamped to 0–100. The adder sizes are judgment, chosen to order markets by integration effort; no project data measured them.
| Check | Adder | Why it moves the score |
|---|---|---|
| Script already shipped | +12 | Fonts, fallback stacks, and rendering QA for the script already exist. |
| Left-to-right layout | +8 | No mirroring work. |
| Right-to-left layout | −10 | Mirrored layout, bidirectional text handling (UAX #9), and icon review. |
| No custom webfont needed | +6 | OS-bundled fonts cover the glyphs. No font licensing or payload cost. |
| Custom webfont needed | −4 | Font selection, subsetting, and load-performance work. |
| 2 or fewer plural categories | +4 | Every count-bearing string needs at most two message variants. |
| 5 or more plural categories | −6 | Arabic-class plural systems multiply message variants and test cases. |
Plural-category counts are CLDR plural rules, checked against Intl.PluralRules. A candidate language is attributed only to countries where it holds at least a 30% L1+L2 share, so Spanish surfaces against Mexico and Spain, not the US.
The overlap model
CLDR reports each language's share of a territory as L1 plus L2 speakers, so a bilingual person counts once per language and one country's shares routinely sum past 100%. Every Atlas aggregate sums the selected-language shares within each country and caps them at 100% of the population. Incremental reach applies the cap before and after, so adding a minority language to a saturated bilingual market contributes only the uncovered remainder.
The cap bounds the error without resolving it: within the capped total, a trilingual speaker may be credited to whichever of their languages was selected first. The data gives share totals, never which individuals overlap.
How speaker shares are printed
Every share of a territory's population prints to one decimal place, rounded up, through one formatter. Russian is spoken by 0.24% of the United States (820,000 speakers), which whole-number rounding printed as 0%. Rounding up is safe because CLDR's territory data is read at a 0.1% floor, so nothing rounds up from zero, and the largest overstatement is 0.09 points on a survey estimate.
How Chinese is counted
Simplified and Traditional Chinese are separate languages in every reach and coverage figure: different written standards, character sets, and terminology. CLDR's territory data keeps their audiences apart. Mainland China and Singapore carry a zh row; Taiwan, Hong Kong, Macau, and Traditional-reading communities in a dozen more territories carry zh_Hant. Shipping one earns no credit for the other, so a product with only Simplified Chinese shows Taiwan as uncovered. When a source lists Chinese without saying which (the App Store search API does this), the model counts it as Simplified, the majority case, and says so where it happens.
CLDR also records spoken Sinitic varieties as their own rows: Cantonese, Wu, Xiang, Hakka, Min Nan, Gan. They stay out of coverage maths, because their speakers read their territory's written standard, which the standard Chinese row already counts. Country and language pages still list them.
The cost model
Costs start from per-word rates times word counts, using this site's curated 2025 planning defaults. Three adjustments move each estimate toward a real quote: a per-language-pair rate index, a translation-memory leverage discount, and a one-time engineering-setup line. One LQA round prices at 20% of the tier's per-word rate, charged on every word regardless of leverage.
| Tier | $/word | Words/day | Appropriate for |
|---|---|---|---|
| Raw machine translation | $0.0001 | 250,000 | Internal tools, search indexes, and content where the gist is enough. |
| MT + light post-edit | $0.04 | 5,000 | Support articles, docs, and FAQs: content read once for the answer. |
| MT + full post-edit | $0.08 | 3,500 | Marketing pages, onboarding flows, anything the user reads carefully. |
| Standard human translation | $0.15 | 2,500 | UI copy, legal text, brand-facing surfaces, contracts. |
| Premium / creative | $0.25 | 1,500 | Taglines, hero copy, and campaigns, where a literal translation falls flat. |
Per-language-pair rate index
A word into Japanese, Icelandic, or a Nordic language costs more than the same word into Spanish, because of translator scarcity and the cost of living where the translators work, not script difficulty. The estimator multiplies each locale's per-word rate by an index; a mainstream pair is 1.0, the default for anything unlisted. These are planning indices, not sourced rates.
Translation-memory leverage
Words already in the translation memory bill below the new-word rate: repetitions and exact matches at 25%, fuzzy 75–99% matches at 60%, new words at full price. A first launch has no memory (0% leverage); a mature product reuses much of its prior wording. Leverage applies to translation cost and schedule, never to LQA.
Engineering setup
Work a per-word rate does not cover, priced once as its own line so an RTL or new-script launch is not quoted like a Latin one. Planning figures, from the locale set:
| i18n framework | $6,000 | One-time: externalize strings, wire ICU, pseudoloc CI gate, TMS hookup. Skipped if already internationalized. |
| RTL enablement | $4,000 | One-time, charged once if any target locale is RTL: bidi isolation, logical-property CSS, icon mirroring. |
| Webfont (per script) | $1,200 | Per distinct non-Latin script that needs a bundled font: subsetting and glyph QA. |
| Per-locale integration | $500 | Per locale: file wiring and an in-context sign-off pass. |
The launch planner prices three launch depths by assigning each content bucket (UI strings, docs, support content) a tier from the table above. Even the cheapest depth keeps a light human review of the UI strings.
| Depth | What ships | Buckets priced | LQA rounds |
|---|---|---|---|
| Machine-translation pilot | Ships UI strings only, machine-translated with a light human review pass. Docs and support stay in the source language. | ui · MT + light post-edit | 0 |
| UI at publication quality | Ships UI strings edited to publication quality with one round of LQA (a human quality-review pass). Docs and support stay in the source language. | ui · MT + full post-edit | 1 |
| Full launch | Ships UI at human-translation quality, docs at full post-edit, and support content at light post-edit, with one round of LQA (a human quality-review pass). | ui · Standard human translation; docs · MT + full post-edit; support · MT + light post-edit | 1 |
Those figures price a one-time launch; upkeep is estimated separately. The model assumes 40% of the source words are retranslated each year (new features, edited strings, refreshed docs), priced at the same depth: tier rates, LQA rounds, and PM overhead. The churn rate is a planning assumption; replace it with a measured figure when one exists.
The briefing's derived numbers
The briefing combines the models above into a few numbers of its own, each with a fixed convention.
Budget ask
- Auto-fills from the recommended market's launch cost (the first row of the cost appendix) plus that locale's one-time engineering setup, itemized in the appendix.
- Setup credits what the portfolio already has: a second RTL language does not pay for RTL enablement again, a shipped script does not re-buy its webfont, and a portfolio of more than one language is assumed to have an i18n framework.
- Leaves out the all-markets total, because the decision approves one launch; the appendix prints that total on its own row.
- A typed override always replaces the auto-filled figure. With the cost appendix off or unpriced, the drafted ask has no dollar amount.
Proposed sequence
The ranking table scores every candidate against the same fixed portfolio, so its reach numbers must never be summed: two overlapping languages would double-count shared markets. The briefing's Now / Next / Later tiles order launches in time, and each tile's net-new count is recomputed as if the earlier tiles had shipped, so those three numbers do sum.
Payback scenario
- Four terms, each labelled on the card: online headroom (an estimate: the anchor market's online population minus its reported MAU), ARPU (your data: revenue ÷ MAU), capture (your assumption), and upkeep (40% annual string churn, an assumption).
- users = headroom × capture; monthly revenue = users × ARPU; months to recover = launch cost ÷ monthly revenue net of upkeep, rounded up.
- Shown only when every term exists: a priced cost appendix and an uploaded row for the anchor market with revenue and an online-population estimate. Otherwise the card is left out, never shown as zero.
- A scenario, not a forecast.
Custom weights
The launch planner's Custom profile has three sliders that renormalize to integers summing to 100. The briefing takes the same override as a URL parameter, weights=R-G-E (three non-negative integers, normalized by their sum): weights=55-15-30 weighs reach 55%, GDP 15%, ease 30%. A malformed value falls back to the selected profile. While an override is active, the memo header says "custom weights" and the methodology footer prints the split and its parameter. Picking any profile clears it.
Snapshot deltas
Comparing two uploads of one product, per-market penetration deltas divide both snapshots' MAU by the same (current) online-population estimate, so the delta measures MAU movement, not drift between estimate vintages. Markets in only one snapshot appear as added or removed, never as ±100%, and totals cover only markets in both. The diff shows what moved, not why.
What the numbers are not
- Double-counting is capped but not resolved. L1+L2 shares still overlap inside the 100% cap, so the model can credit a multilingual person to the wrong selected language.
- Two GDP figures coexist. CLDR's long-run GDP estimate drives the launch score and the nominal-GDP metric. World Bank GDP per capita (current US$, latest year) appears on country pages and drives the per-capita lenses; CLDR long-run GDP ÷ population fills in where the World Bank has no figure. The two are different vintages and can disagree. Each number comes from one source, never an average of both.
- No revenue estimates. The Atlas sizes audiences and costs, never revenue. The only usage numbers are uploaded ones (Your data), and they are display-only, never scored.
- The ease model has a coverage edge. Its metadata table covers 80 language codes. Outside it, the model assumes the cheapest shape (Latin script, LTR, 2 plural categories, no webfont), so the planner flags those rows "No engineering metadata".
- EF EPI is context, not an input. The index samples self-selected online test takers, skewed young and toward people already learning English. It never enters planner scoring.