Is Automatic Website Translation Good?

It depends on what you're translating. Automatic translation reaches 85-92% quality on simple, factual copy — product specs, FAQs, blog body, T&Cs. It collapses to 40-60% on persuasive marketing language. The 2026 stack of DeepL, ChatGPT-4, and Google Translate has narrowed the gap enough for automatic to handle the bottom 80% of website pages — but the conversion-critical 20% still requires human editing.

Where automatic translation has genuinely caught up

The 2024-2026 generation of neural and large language model translation has closed measurable ground on factual, structured content. DeepL hits 92% accuracy on product specifications. ChatGPT-4 with brand-voice prompting hits 88% on FAQ content. Google's PaLM-2 translation API hits 87% on T&Cs.

Categories where automatic translation now performs at native-acceptable quality:

• Product specifications — Dimensions, weights, materials, compatibility lists. The structured nature plays to neural translation strengths.
• FAQ body content — Questions and factual answers. When the source is well-written, automatic output is consistently usable.
• Blog body paragraphs — Informational, expository writing without heavy idiomatic content. Output is often indistinguishable from human translation.
• Legal terms (with caveats) — T&Cs, privacy policies, returns. Quality is high but legal-precision risk means most brands still want a native lawyer to review.

Slator's 2025 translation quality benchmark shows the gap between top-tier neural output and professional human translation has narrowed from 18 BLEU points in 2020 to 6 BLEU points in 2026 on European pairs.

Where automatic translation still fails badly

The 40% quality figure on persuasive copy isn't soft — it's measurable. Native speakers rank automatic-translated hero copy and CTAs as foreign-flavoured 60% of the time, and rejection rises to 70%+ on culturally-specific brand voice.

The four failure modes on persuasive copy:

1. Brand voice flattening — A challenger DTC tone becomes neutral catalogue prose. The verbs lose their punch.
2. CTA register collapse — "Shop now" translated literally lands as either childish or aggressive. Native CTAs use different verbs.
3. Cultural reference failure — US-cultural references (holidays, sports, pop culture) translate literally and confuse target-market readers.
4. Idiomatic compression — English marketing prose packs meaning into idioms. Automatic translation expands these to literal explanations, losing the punch.

These failures don't show up as obvious mistranslations — the output is grammatically correct. They show up as conversion damage. CSA Research data shows 73% of EU shoppers can identify foreign-flavoured copy on a first read and rate the brand 18-31% lower on trust.

The 80/20 split that actually works

Use automatic translation as the engine for the 80% of website copy that's factual and structural, with native editing on the 20% that drives conversion. That split has held stable since 2022 because the failure boundary tracks copy type rather than absolute quality threshold.

The practical 80/20 split:

Copy type	Volume share	Automatic alone	Needs human
Product specs	25%	Yes	No
FAQ + help	15%	Yes	No
Blog body	20%	Yes	No
T&Cs + policies	10%	Yes	Lawyer review
Footer + nav	5%	Yes	No
Headlines + hero	5%	No	Yes
Category pages	8%	No	Yes
CTAs + buttons	3%	No	Yes
Checkout copy	5%	No	Yes
Email capture	4%	No	Yes

Automatic alone covers 75% of word volume. Human editing covers 25% — but it's the 25% doing the conversion work. The economic logic is that the editor's £400-£800 per language sweeps over only the conversion-critical pages. Our best AI website translation guide walks through engine selection.

How to deploy automatic translation safely

The deployment side matters as much as the engine side. Automatic translation deployed without proper hreflang, without a language switcher, and without GDPR cookie banners can rank worse than the source-language version alone. The engine gives you raw text — deployment turns it into a working multilingual site.

The seven-point safe deployment checklist:

1. Hreflang on every page — Both directions plus x-default fallback. Google's hreflang documentation is the canonical reference.
2. Native URL slugs — /de/produkte/ not /de/products/. Slug translation matters for SEO and trust.
3. Translated meta titles and descriptions — These appear in SERPs and need conversion-quality translation.
4. Localised structured data — JSON-LD product schema with localised currency, language attribute, offers.
5. GDPR cookie banner in target language — Cookiebot or Iubenda at £8-£15/month per language.
6. Language switcher in header — Not buried in footer. Visible globe icon or country flag.
7. Sitemap segregation — Separate sitemap-de.xml, sitemap-fr.xml files in a sitemap index.

Brands that ship 1-3 of these see partial benefit. Brands that ship all 7 see full uplift. Our make website multilingual guide walks through implementation.

What we ship at Eldris vs DIY automatic stacks

Our Growth tier at £997 activation ships the full 7-point deployment automatically — DIY automatic stacks typically ship 3-5 of the 7. The cost of the difference is what determines whether automatic translation is good in your specific case. Shopify's international ecommerce guide consistently flags incomplete deployment as the top reason translated sites underperform projected ROI.

Concrete shipped scope at Growth tier for 4 languages:

• Engine layer — DeepL Pro API for bulk translation, ChatGPT-4 with brand-voice prompting for hero pages, native editor sweep on conversion-critical copy.
• Deployment layer — Site migrated to Cloudflare Pages static export, hreflang on every page, localised JSON-LD product schema, native URL slugs, translated meta titles, sitemap index.
• Compliance layer — GDPR cookie banner per language, German Impressum/AGB/Widerrufsbelehrung where DE is selected, French/Italian/Spanish equivalent legal pages.
• Ongoing layer — New product translations within 48 hours, blog post translation within 7 days, monthly hreflang QA.

For sibling EU compliance work that runs alongside translation, epr.eldris.ai handles LUCID and ElektroG registrations FBA EU sellers need, or contact us for a scoped quote.

Frequently asked questions

Is DeepL automatic translation better than Google Translate?

Yes, on European pairs. DeepL scores 31.1 BLEU on English-to-German versus Google's 28.8. The gap is larger on commercial product copy (6 BLEU points). For Italian and French, DeepL leads by 3-5 BLEU points. For Asian languages, Google leads on Mandarin and Japanese. Our DeepL vs Google Translate post covers head-to-head benchmarks.

Can ChatGPT replace automatic translation engines?

Yes for persuasive copy with prompting; no for bulk efficiency. ChatGPT-4 produces 88-94% quality on hero copy with a brand-voice prompt. But it's expensive at scale (3-5x DeepL API costs) and slower (seconds versus DeepL's millisecond response). The 2026 stack uses ChatGPT for the conversion-critical 20% and DeepL for the bulk 80%.

What's the worst-case quality on automatic translation?

Roughly 40% native-acceptable on heavy idiomatic marketing copy without prompting or editing. That manifests as output that's grammatically correct but unmistakably foreign to native readers — a German shopper reads it and rates the brand as untrustworthy. The fix is the human editing layer; raw automatic output deployed at this quality level damages brand more than it helps reach.

Does automatic translation work for legal pages?

For first-pass drafting, yes. For final published versions on consumer-facing legal pages, you want a native lawyer's review. The risk is that wrong words become legally enforceable consumer rights. A mistranslated 14-day return becomes a 28-day return because the German says "Wochen" instead of "Tagen". For B2C ecommerce in Germany, France, Italy, Spain, plan £200-£500 per language for legal review.

How much does fully automatic translation cost per language?

DeepL Pro API for a 50-page site runs €30-€80 per language one-off. Add ChatGPT-4 brand-voice prompting for hero pages and you're at €200-€400 per language. Add native editor sweep on the conversion-critical 20% and you're at £600-£1,200 per language. That's the realistic cost of good automatic translation in 2026.

What's the difference between automatic and machine translation?

In 2026 they're effectively the same thing. Machine translation is the academic term. Automatic translation is the marketing term in commercial contexts. Both refer to neural and LLM translation — older statistical machine translation (SMT) has been deprecated. When a SaaS tool advertises automatic translation, it's running DeepL, Google Translate, Microsoft Translator, or an LLM under the hood.

When should I avoid automatic translation entirely?

Three cases. If your brand voice is highly idiomatic (heritage brands, challenger DTC) — automatic flattens it. If you're translating into a low-resource language where engines are weaker (Hungarian, Estonian) — quality drops below safe deployment threshold. If you're translating high-stakes legal or medical content where errors carry liability.