GEO-084·Issue 07·Measurement & Growth·84 / 132
Localization ROI: When to Stop
Localization is a budget with a payback period. Four inputs, a three-tier stop line, and the German-site downgrade that beats another launch.
Issue focusAcross Markets: Multilingual, Localization, and Regional Semantics
A composite scene: a US pet-tech brand ships a smart feeder into five languages by month seven. The German storefront is live and the invoice is paid, yet a Berlin renter asking an assistant, in German, for a feeder that fits a small flat gets two local brands with checkable reasons. Hers does not appear. At the review, nobody asks what went wrong in German; they ask about Italian — a quote is easier than an explanation.
Localization is an investment with a payback period, not a coverage contest: price each language’s cost, ceiling, and reuse, and write the downgrade terms before you need them.
Cost is an invoice; return is a probability
Localization spend is front-loaded and countable — translation, review, templates, hreflang upkeep, support, compliance copy. Return is deferred and diffuse: one more German mention, one shortlist seat, one local review quoted back. None of it lands in a cost center’s report, and when AI referrals count as Direct, it may not land anywhere.
So teams count pages shipped, the only column with data, and expansion becomes the default. No ledger asks whether a language repays its cost.
Return falls as languages accumulate. The first language reuses master data, evidence, and templates; by the fifth, what remains is long tail, while price, stock, and policy sync must hold everywhere. Each language is its own retrieval market: an American review wall does not vouch for a German answer.
Marginal cost per page is roughly flat; marginal return is not. A twenty-language plan at thirty pages each reads as coverage, yet no single language answers a full purchase chain — a list, not a position.
Four inputs, and one is usually blank
An ROI model that misses usually has one empty input.
| Input | Question | Where it comes from | If blank |
|---|---|---|---|
| Marginal cost | What this language costs per year | Translation, review, hreflang, support, compliance, tagged by language | Budget is averaged and small languages look cheap |
| Ceiling | Upper bound of AI-reachable demand | Category volume, listing counts, rivals, as a range | Slow growth reads as poor execution, so you add pages |
| Payback | By when shortlist share or orders appear | Quarterly checkpoints, fixed before launch | The project lives in “give it another quarter” |
| Reuse | How much evidence and template carries over | One spec sheet or certificate serving several languages | Every language becomes a new project |
Ceiling is skipped most because it needs judgment, not a report; without it, payback has no denominator. The common substitution is content completeness: pages finished are not pages cited.
Write the stop line before you need it
A stop line is not “close the site when it loses money” but three actions agreed in advance: invest, maintain, downgrade.
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Ledger | Cost and return per language by content, technical, ops | Finance, growth | One row per language, annual cost computable |
| 2 Ceiling | Estimate the limit; record assumptions and range | Growth, category | A range, not a single number |
| 3 Checkpoints | First, second, fourth quarter; shortlist share and local evidence, not sessions | Growth | Each review ends in invest, maintain, or downgrade |
| 4 Downgrade | Freeze long tail; keep core pages, local evidence, price and policy sync | Content, technical | The action exists before the stop line |
Step 4 decides the outcome. Most teams hold the “should we close it” conversation the day the numbers turn, when sunk cost takes over; written in advance, the discussion is execution.
Downgrade is not teardown. Deleting pages, collapsing hreflang, and redirecting old URLs zero out the local signals you built; a small citable core keeps evidence a model can still use. If you merge languages, handle redirects and alternates through multilingual URLs and hreflang.
Example: downgrade the German site, do not close it
Example — brand, market, and page counts are invented; only the decision logic is reusable.
A home-projector brand’s German site is nine months old and the second most expensive of five languages. Mentions in German answers rose for a quarter, then flattened. German orders skew to branded search with few new customers; the team is ready to add Austria and Switzerland. The review finds three things: most long-tail pages never appeared in sampled answers; local evidence is the brand’s own claims; and support’s return wording contradicts the policy page, so the model quotes both and argues with itself.
The decision: maintain, not close. Stop long-tail pages; keep twelve core pages — category, comparison, policy, certification. Move the budget into German third-party review and community evidence. Fix the policy wording before adding a language. If shortlist share has not moved in three quarters, drop to a five-page citable set.
The stop point is not “German does not matter” but “expanding without local evidence copies the contradiction onto more pages.”
Where this breaks
The model fits optional markets, not mandatory ones.
Compliance languages. Customs, invoices, and consumer notices may legally require the local language. That is admission cost, not investment; ranking it by ROI trades a saving for a larger penalty.
Platform-validated markets. If listings must be local-language, judge the site’s return after the platform’s. Do not use a weak site metric to cancel a market the platform already sells in.
Early positioning. In a thin corpus, early numbers look bad. If you believe the market matters in a year, bet on the ceiling — and write the assumptions and exit condition, or the bet becomes an indefinite subsidy.
Regions inside a language. Spanish is not one market: units, payment, returns, and vocabulary differ between Spain and Mexico. One ROI unit overstates return and understates upkeep; the five traps in local buying language are where this surfaces.
One misuse: a page that cannot be crawled, or an entity that does not resolve, is not a localization problem; counting it as one turns a technical failure into a market verdict. Fix visibility first.
For market order, start with how to choose the first country. For metrics and cadence, see the measurement and growth hub.
This month, split localization spend into one row per language. Whatever will not split is what you cannot yet decide on. For the most expensive, least legible language, fill the four inputs and write two conditions: what moves it to maintain, and what moves it to a minimum citable set.