GEO-127·Issue 11·Technical Foundation·127 / 132
A Multilingual Publishing Pipeline: One Brief, Two Markets
One brief, two localizations, one release state: a publishing pipeline that keeps shared facts frozen, local wording free, and rollback paired by ID.
Issue focusScale It: SOPs, Organization, and the 2027 Roadmap
Early November: a US smart-lock brand’s German storefront enters its promo week. English ships Monday; German is queued for the next batch. Over three days, operations changes two facts — battery life 12 to 18 months, cells two to four — on the English page only. hreflang already points at /de/, which has not been generated, so a crawler takes a 404. Then support forwards a screenshot: a chat assistant told a Munich buyer the lock runs 12 months on a set of batteries and carries a three-year warranty. One number came from German copy, one from English. No page says both. Composite scene.
The unit of multilingual publishing is not an article but two drafts under one ID — one brief, two localizations, one batch, one rollback — because a staggered schedule lets a model assemble a product that exists on neither page.
A staggered schedule is two sites in one logo
Queuing translation after publication assumes facts freeze at handoff. They do not. In the two weeks a translation waits, someone changes a price, a return window, a spec.
A page is a file; a fact is a paragraph. Write the same number twice and there is no single source. The editor touching the visible draft makes a locally correct decision that becomes a second truth.
The model does not arbitrate. Retrieval returns spans, synthesis joins them. Every number in that screenshot traces to a real page; the combination traces to none.
Crawling pays too: hreflang claims an alternate that exists, the sitemap lists one side, and budget goes into 404s. Crawl consistency across language paths covers access.
The friendly counterexample, “English next week,” republishes stale facts with fresh confidence. Whole-page machine translation reads well while units, voltage, and warranty framing stay in the source market.
One brief, two checklists
A brief is not a translation memory. It is one page: which questions the page answers, which facts are frozen, which must be rebuilt locally.
| Layer | Fields | Test |
|---|---|---|
| Shared, frozen | Model string, price basis, specs and test conditions, certification IDs, warranty and return windows | Reconcile to master data; zero differences |
| Local, must change | Units, voltage, compliance and warranty wording, scene language, delivery and tax display | Sample real questions per market |
Read shared values from the source, never the previous version. Price, stock, and policy in body copy get missed; generate them from one SKU record and let prose explain.
Two differences, two checks: the same fact in two sentence shapes is localization; the same fact as two numbers is an incident. Review that conflates them polishes translationese and waves a price through.
The local question list is generated locally. Translating a question tests the translator, not demand. Sample per market, as per-market measurement argues.
Five steps, then a state machine
The pipeline lets two people work at once on one version of the facts.
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Brief | Pull facts from master data; list 20 local questions | Category ops | Every question maps to a page |
| 2 Dual-track draft | Two native writers draft the same skeleton blind | Local writers | Each draft answers alone |
| 3 Cross-review | Swap; check numbers, policy, citations; polish last | Other market’s writer, editor | Shared differences zero |
| 4 Publish as a batch | Both pass before sitemap and hreflang | Release owner | One code, two URLs logged |
| 5 Roll back by ID | Withdraw both, or keep both | Engineering | No dangling hreflang |
Step three is the one teams invert. Polish first and every corrected number forces a second polish. Facts first.
Step two costs a second native writer, not a second budget: one brief, one review. Re-translation is the expensive option — stale facts again, plus another crawl and sample.
Publishing is a state machine. A draft ships nothing and carries no hreflang; single-sided goes live without alternates, or is noindexed; paired ships reciprocal hreflang and x-default together. hreflang means “already exists,” not “intended.”
Rollback uses the same unit. The log holds both URLs, the price snapshot, the approver. Pull one version and the pair falls back to single-sided with a dangling alternate, so the survivor becomes the only source. Rollback granularity equals publish granularity.
Example: one lock, two drafts
Example — model, specs, and figures invented; no real product described. Shared facts: L-20, four AA cells, 18 months at eight opens a day, tested at 77°F, a two-year statutory German warranty plus one year extended, three business days from a German warehouse.
| Field | English | German | Class |
|---|---|---|---|
| Runtime | 18 months | 18 Monate, same conditions | Shared |
| Cells | 4× AA | 4× AA (Mignon) | Shared fact, local notation |
| Warranty | 2-year statutory + 1-year extended | 2 Jahre Gewährleistung + 1 Jahr Garantie | Same terms, local wording |
| Delivery | 3 business days | 3 Werktage ab DE-Lager | Shared lead time |
Sentences may differ; facts may not. The row needing judgment is warranty: statutory and extended are two commitments, and a German buyer needs to know which years are which. Blur it and the model quotes “three years” as one brand promise.
Where this breaks
One market, one language. A release checklist, not a state machine.
Long-tail and policy pages. Size charts and shipping terms stay in the primary language with a scoped local summary; absolute pairing jams the site.
No native writer. One draft, several languages, local review; lower the visibility expectation, and machine-translate only non-decision pages.
Regulated categories. Radio equipment, children’s products, medical: regulatory review is part of the gate.
Marketplace listings. Owned by the platform and feed pipeline, out of reach.
A pair is not demand. Two versions prove readable, not cited; sample per market.
This line belongs to the technical foundation hub: whether two drafts point at one set of facts, not whether the market is worth entering. Not this quarter? Put it after visibility work on the 90-day GEO roadmap. For which fields must match at all, the multilingual parity audit goes deeper.
This month, pick one live bilingual SKU and open both bodies side by side. Check the shared fields, one at a time. Where a row disagrees, do not edit the sentence — fix the source it reads from.