Varley Germany · Internal planning · Prepared by SearchFlex
One question changes every sentence: Du or Sie? Here is how we get a German copy of the current Varley site live in about seven weeks, and what we need decided to start.
The goal is a German copy of the current Varley site, live before Black Friday on 27 November 2026, with enough lead time for Google to index it.
A complete German version of the current site, not a partial one. Black Friday is about 7.5 weeks away.
Flexpress can draft and check at speed. A native German speaker still has to approve the glossary and review the pages that matter.
A new language version takes time to be picked up. Every day we launch earlier is a day of indexing before the sale.
Nine inputs. Tick them off as Varley confirms; the bar shows how close we are to kickoff. The three tagged items are the ones that can stall everything else.
Ticks are for use in the room and reset when the page reloads.
Click any bar for the detail. Dates assume we start the week of 5 October. Pink bars are Flexpress, yellow is the freelancer, grey is the team.
Going live at the end of week 6 means about 15 November, which leaves 12 days before Black Friday. That is not much time for indexing.
Proposal for discussion: overlap the later stages and ship the top pages first, so the pages that earn the most get the longest time in the index. The alternative is to treat Black Friday as a soft launch for the German site.
Flexpress has no translation agents today because the EU store is English-only. These would be new agent definitions that reuse the existing pipeline pattern. Each stage reads a file and writes a file for the next one, so no text is pasted around. Click a stage.
Anything a check cannot pass is flagged for the freelancer. It is never shipped.
A page can read beautifully and still have lost a sentence. That is why fluency and fidelity are checked by two separate agents.
Judges fluency
Judges fidelity
Running the reviewer on a different model from the translator helps catch what the translator misses.
Freelancer edits feed back through the existing learning loop, so the same mistake is not repeated.
Translation must not add or strengthen a product claim. The accuracy checker is there to hold that line.
Full review where the money is, a sample everywhere else.
A sample of the remaining pages. If the sample shows a systemic problem, we fix it at the source and rerun, instead of patching page by page.
Five areas where a straight translation goes wrong or gets missed.
Toggle the options. The estimate assumes about 1,000 pages at roughly 3k tokens in and 3k out per pass, across four model passes: translate, blind review, accuracy check and revise.
Translator model
Options
About 1,000 pages
about $0.10–$0.15 per page
Estimates from pricing cached on 2026-09-25, not yet re-checked. The pilot gives us real numbers to replace these.
It changes every sentence, so it has to be fixed before the glossary and before the pilot.
They come from a German e-commerce legal source and are not a translation job. Ask Varley on day one.
About 10 representative pages through the full pipeline, to show real quality and real cost before we commit.