DeepL vs Google Translate: An English-to-Dutch Workflow Test

Quick answer: DeepL and Google Translate preserved tested filenames, template tokens and numerical values in our English-to-Dutch workflow study. Context was less consistent: Google's initial bird-specific translation changed when the sentence pair was repeated; both identified the isolated bird. Actual outputs and screenshots support task-specific checks, not an accuracy winner. Human Dutch-language review is pending.

A support translation can look polished while changing the one thing that matters: a condition, an amount or a word with two meanings. This study compares DeepL and Google Translate on those details. The evidence is the actual text returned by their public interfaces, with the full inputs available for inspection.

How we tested: On September 8, 2026, a Codex agent operated both products through Chrome using synthetic English text and Dutch as the target language. The founder did not personally perform or score this test. The article and analysis are AI-assisted. A qualified human review of the Dutch translations has not been completed; observations here are not a professional linguistic assessment. The later-session screenshots below show actual output. They are cropped to product content to exclude browser chrome and account identifiers; no text was retouched or synthetic interface created.

About the cover: The cover is an AI-generated editorial illustration of translation ambiguity. It is separate from the actual product screenshots below and is not official DeepL or Google artwork.

Jump to: Test method · Context results and screenshots · Additional workflow checks · Practical checklist · FAQs

What exactly did the test cover?

The first session used free desktop web translation in a fresh Incognito window, without signing in. Six numbered cases were pasted as one request per product, followed by an exploratory context request. Later the same day, the operator tested 12 additional cases in three predefined batches and four context requests per product.

That is nine primary requests per product across the two sessions, not 18 independent sentence trials. Batch context can influence individual sentences. Two additional DeepL requests restored earlier inputs for screenshots; their outputs matched the recorded ones and are documented separately, not counted as independent tests. No paid feature, API, document upload or response latency was evaluated.

The initial input/output JSON and extension input/output JSON preserve the full requests and responses. Different account conditions prevent treating the later session as a controlled replication. No account setting or translation history was examined, and we cannot attribute output changes to signing in.

The excerpts below reproduce returned text. Differences in paragraph spacing are retained in the downloads but are not counted as language-quality failures. Screenshots establish what appeared on screen for the illustrated cases; they do not independently validate the translations.

Did filenames and template tokens survive translation?

Both products left the exact filename and placeholder unchanged. That is a useful result for this sentence: translating a filename can make an instruction unusable, and modifying a token can break a template. It does not prove that either service will preserve nested markup, every placeholder convention or an uploaded document.

Input, case 3:

Keep the file name budget_Q4.csv and the placeholder {{first_name}} unchanged.

DeepL:

Laat de bestandsnaam budget_Q4.csv en de plaatshouder {{first_name}} ongewijzigd.

Google Translate:

Laat de bestandsnaam budget_Q4.csv en de placeholder {{first_name}} ongewijzigd.

The surrounding word differs: DeepL returned plaatshouder, Google returned placeholder. We are not awarding a style point for either. The measurable check here is that budget_Q4.csv and {{first_name}} match the originals character for character.

For your own trial, choose the exact token syntax your system uses and check it after translation. Protecting one example should be an acceptance check, not a product-wide guarantee.

What happened to the invoice amount and date?

Both outputs retained the numerical value and the October 9, 2026 date, while changing number punctuation to the Dutch format. Google replaced EUR with the euro symbol; DeepL kept the currency abbreviation after the amount. A workflow that requires literal currency codes needs a stricter check than a workflow that only requires the same value.

Input, case 2:

The invoice total is EUR 1,234.56 and payment is due on 9 October 2026.

The complete sentences use different wording around the payment deadline. A Dutch reviewer still needs to assess whether that wording suits an actual invoice. This fixture also spells out the English month; it does not test ambiguous dates such as 09/10/2026.

Did the instructions keep their conditions?

The backup instruction produced different Dutch constructions, and the renewal instruction produced the same sentence in both services. No apparent reversal was identified in this initial inspection, but that observation awaits human semantic review. Negation deserves a separate check because an otherwise fluent instruction can become harmful if its condition changes.

Input, case 1:

Do not delete the backup until the new copy has been verified.

DeepL:

Verwijder de back-up pas nadat de nieuwe kopie is gecontroleerd.

Google Translate:

Verwijder de back-up niet voordat de nieuwe kopie is geverifieerd.

Input, case 6:

Your subscription will not renew unless you turn automatic renewal back on.

Both products:

Uw abonnement wordt niet verlengd, tenzij u automatische verlenging weer inschakelt.

The identical renewal output gives no basis to place one product above the other on this case. For an actual support workflow, ask the reviewer to state what action triggers renewal, and compare that with the source instruction.

Why did the crane example need a follow-up?

The initial crane sentence did not establish that the first crane was a bird. A construction crane can stand beside a river. Both products chose kraan for both occurrences, so calling that a translation failure would overstate the evidence. The follow-up added wings and a building site to make the intended contrast explicit.

Initial input, case 5:

The crane stood beside the river. The crane lifted a steel beam.

Both products:

De kraan stond naast de rivier. De kraan tilde een stalen balk op.

After viewing that output, the operator submitted this new request to both products:

The crane spread its wings beside the river. The crane lifted a steel beam at the building site.

DeepL:

De kraan spreidde zijn vleugels uit naast de rivier. De kraan tilde een stalen balk op op de bouwplaats.

Google Translate:

De kraanvogel spreidde zijn vleugels naast de rivier. De kraan heeft op de bouwplaats een stalen balk gehesen.

That first result appeared to favour Google on this example. The later session made the conclusion less simple. When the same two-sentence input was repeated, Google returned kraan for both occurrences, while DeepL again returned kraan for both. The original output remains in the dataset rather than being replaced with the later result.

For the isolated machine sentence, both returned kraan. The table is a record of word choices, not a blinded linguistic score. We have not established why the paired and isolated versions differed, or why Google's original-pair output changed between sessions.

Google Translate, later session on September 8. The output uses kraan in both sentences. Google was signed in; the crop excludes account identifiers and browser chrome.

DeepL, later session on September 8, signed out. The same input was restored for this capture and the output matched the recorded result. This restoration is not an additional independent test.

The reader-facing implication is narrower and more useful than a winner badge: review the text in the context in which you will actually use it. Splitting a sentence out can produce a different translation. This study gives examples to investigate; it does not establish that removing context generally improves translation.

What did the 12 additional support-workflow cases reveal?

The later batches tested more literal tokens, dates, conditions, units and formatting. They also exposed differences that a support team should inspect even when the numbers survive. These cases were written before running the extension. They broaden the task coverage, but do not create a representative Dutch translation benchmark.

In batch E, both translated the English button labels “Save draft”, “Preview” and “Publish” into Dutch. That is visible behaviour, not necessarily an error. If the actual application still displays English labels, the translated support instruction may be harder to follow. Decide whether labels should be translated before sending them through either tool.

Both also translated the words inside **bold text** while retaining the asterisks. The placeholder and code identifier stayed literal. That distinction matters: preserving markup syntax and preserving all text inside the markup are different requirements.

Google Translate's later-session batch E output. The screenshot records plain text with markup characters; it does not demonstrate rendering in a document or support tool.

DeepL's batch E output, restored for capture after the recorded run. The words and tokens matched that run; extra blank lines are retained in the raw data. Content-only crop, with no retouching.

Which conditions still need a human decision?

One source instruction was: “Refund the delivery fee unless the customer selected express shipping.” DeepL returned an imperative, while Google's output described the fee being refunded:

DeepL: Geef de verzendkosten terug, tenzij de klant voor expresverzending heeft gekozen.
Google: De verzendkosten worden terugbetaald, tenzij de klant voor expresverzending heeft gekozen.

The exception remains visible in both, but the change in grammatical form deserves review if this is an instruction to an employee rather than a message to a customer. We have not rated its severity.

The dispatch example said: “We will dispatch the parcel by Friday. This does not mean it will arrive by Friday.” DeepL used uiterlijk vrijdag in both sentences; Google used it for dispatch, then op vrijdag for arrival. A Dutch reviewer should assess the deadline distinction in the full message. Preserving “Friday” alone does not settle that question.

Finally, the gender-neutral source “Each customer should check their settings before they continue” became zijn and hij in DeepL, and zijn of haar and hij of zij in Google. Neither output retained the English source's neutral wording in the same way. House style and audience should guide the edit; this pilot cannot award a general inclusivity score.

What did the polite attachment request show?

Both outputs began with Kunt u and asked for the missing attachment, but their wording differed. DeepL included alstublieft; Google did not. This observation does not settle which wording a Dutch support team should use. Tone depends on the audience and house style, which the short source sentence does not specify.

The source was: “Could you please send the missing attachment when you have a moment?”

DeepL showed disabled Pro controls for formality and style in this session. Those controls were not tested. Their presence is not evidence that a paid option would fix a particular sentence. Consult DeepL's plan documentation for the service you are considering.

Can these results justify translating customer records?

No. The study used invented text and did not assess either provider's handling of confidential information. Translation quality and permission to process a record are separate decisions. A European company's free service should not be assumed to have the same data terms as its paid plans.

DeepL's privacy policy distinguishes free translation, where submitted content can be used to improve its systems, from its listed Pro services. It prohibits personal data in the free translator. Pro has different terms and requirements; buying a subscription is not by itself a complete processing assessment.

For Google Translate, review its privacy and account-control guidance alongside your organisation's rules. We did not test a business translation API or verify a retention configuration. Use synthetic examples when investigating quality before approving a service for real work.

How should a support team use this study?

Use it as a starting checklist for your own acceptance test. Literal tokens survived the tested inputs, while context and operational wording still needed inspection. A small set of tasks with explicit expectations is more useful for that decision than a fluent-looking paragraph and an unexplained score.

1. Choose representative synthetic messages. Include actual placeholder syntax, conditions and ambiguous terms without real customer details.
2. Write expectations before translation. Identify values and tokens that must survive, the action the recipient should take, and the intended register.
3. Run the same inputs in both products. Record language direction, plan, settings and date. Keep the complete output, including awkward sentences.
4. Have a Dutch-speaking reviewer assess blind copies. Separate meaning, terminology and tone; record disagreements rather than hiding them in an average.
5. Check the workflow separately. A successful text example does not establish document formatting, integration reliability or data-policy suitability.

Start with the DeepL directory entry for product context and the broader AI category for related tools. If this is part of replacing several Google services, our de-Googling guide covers the wider transition, and the Proton Mail vs Gmail comparison distinguishes mailbox migration from changing the tools around it.

The evidence supports testing those details carefully. It does not support a star rating, a claim that either translator is best for Dutch, or a purchase recommendation for an untested paid plan.

Frequently asked questions

Was DeepL or Google Translate more accurate in this Dutch test?

This small study does not establish an accuracy winner. Both products used kraanvogel for the isolated bird sentence. In paired bird/machine sentences, outputs differed between sessions and both used kraan in the later session. Account conditions also differed. Human Dutch-language adjudication is pending.

Did the translators preserve the template placeholder?

Both preserved the literal placeholder and filename in the tested sentence. That result applies to those exact tokens in this input, not every template syntax or document format.

Were these actual products or AI-generated example translations?

A Codex agent operated the actual DeepL and Google Translate public web interfaces in Chrome using identical synthetic English inputs. The recorded Dutch outputs came from those interfaces. The article and analysis were prepared with AI assistance; no claim is made that the founder personally performed the test.

Were DeepL Pro and Google Cloud Translation tested?

No. Both sessions used public desktop web translation. The first session was signed out in both products; the later session used signed-in Google Translate and signed-out DeepL. No paid feature, API, glossary, document upload or alternative tone setting was tested.

Can I paste customer emails into the free translators?

Do not treat this synthetic-text pilot as approval to submit customer information. DeepL's free-service policy prohibits texts containing personal data. Check the exact service terms and your organisation's rules before processing real records with either provider.

Can I reproduce the comparison?

The article links two JSON files containing the complete source inputs and observed outputs. Use the same language direction and record your date, plan and settings. Outputs can change; matching our result is not guaranteed.