How to Choose a Translation Company

Search for a translation company, and you get several million results, all saying roughly the same four things: Fast, accurate, certified, global.

None of that helps, because the differences that matter are rarely the ones anyone advertises. They show up eighteen months in, when a resource file comes back with the variables translated, when Legal asks who approved the German version of an indemnity clause, or when you want to change providers and discover that nobody documented who owns the translation memory.

The real test is not what a translation company promises on day one. It is how well the relationship holds up eighteen months later.

This guide covers the six areas worth evaluating, a weighted scorecard for comparing providers on evidence, and a pilot design that finds problems before a contract does.

How should you build a shortlist?

Around business risk, not price. A safety manual, an ecommerce product page, a software string, and an internal memo do not need the same workflow or the same review depth. Price them identically, and you will overpay for one while under-protecting another.

Before you ask anyone for a number, group your work by consequence. German “Gift” means poison. Spanish “embarazada” means pregnant. Both are simple examples, but they illustrate why domain knowledge matters more than a generic bilingual match.

For high-impact content, ask each company to walk you through its process from discovery to final quality control. If the work touches customer acquisition, ask about localized SEO and analytics. If it involves software, ask how they handle string files, pseudo-localization, and issue reporting for engineers.

The point is not to hear that they have a process. It is to find out whether that process is designed around the consequences of getting the translation wrong.

A good provider will welcome this, because it lets them price the work properly. A weak one quotes everything as standard translation and leaves risk discovery until after kickoff, which is the point at which it becomes your problem.

What kind of company are you actually comparing?

Most shortlists end up mixing three or four different business models. Comparing them with a per-word rate produces a ranking that means nothing.

The archetype tells you which risks to probe hardest. It does not tell you which provider is better. A technology-first platform and a specialist LSP can both be right, for different content. What should not change is the evidence you ask for: Documented process, qualified linguists, secure handling, technical competence, and asset ownership in writing.

Language coverage deserves the same skepticism. Breadth and depth are different claims. A provider advertising 200 languages may have a thin bench on the three pairs you actually need. The useful question is how they handle a pair they have never quoted for you before. Who sources the linguist, what qualification evidence gets kept, and who revises the first delivery.

How do you verify a quality system?

Check the system before you look at samples. ISO 17100:2015 is the clearest baseline for whether a company has a documented process or an informal pool of bilingual contractors.

Ask whether the workflow includes translation by a qualified translator and revision by a second qualified person. That separation is the control, and it is the thing most likely to be missing.

Samples mislead. A company can assign its best linguist to a test and a different team after you sign. Stronger evidence looks like anonymized quality reports, terminology approval records, reviewer instructions, escalation logs, and examples of what they changed after an error.

Ask how errors are categorized. A critical terminology error in a medical device instruction is not the same as a style preference in a social post. If they use an error typology, ask how defects feed back into retraining, linguist assignment, and future QA checks. That loop is the difference between a quality system and a quality report.

Four questions about linguists

• Who qualifies linguists for each subject area, and what evidence stays on file?
• Does a second linguist revise customer-facing content before delivery?
• How is approved terminology maintained across recurring projects?
• When an in-country reviewer disagrees with the translator, who decides?

For recurring work, ask for a named project manager and a documented communication path. Quality degrades when feedback lives in email threads instead of a termbase, a style guide, or a translation memory.

How do you check technical fit?

A company handling websites or software in 2026 should be able to explain BCP 47 language tags from RFC 5646, because tags such as en-GB, fr-CA, and es-MX affect routing, rendering, and search signals.

This is where most evaluations go shallow. A linguistically strong company can still slow you down if it cannot process JSON, XLIFF, XML, HTML, Markdown, YAML, PO, or design-tool exports without damaging code. Do not ask whether they support a format. Ask which parser, connector, or translation management system they use, and how locked text is protected.

One small attribute causes visible problems. The HTML lang attribute tells browsers and assistive technologies what language a page is in. W3C guidance on language tags explains how those declarations affect processing and accessibility behaviour.

For software, ask how they protect placeholders. In ICU MessageFormat, a string such as {count, plural, one{# file} other{# files}} has to survive translation with its braces, variable names, and plural categories intact. A company that treats it as ordinary prose will introduce runtime defects, and they will surface in production rather than in review.

Three things to test before signing

• Pilot one real file type, not a copied paragraph. Parser behaviour decides whether tags and variables survive the round trip.
• Review a screenshot-based QA sample. Truncation and gender-agreement defects only appear in context.
• Confirm ownership of translation memory and termbase assets in writing. Switching providers later get expensive when the assets are locked away.

If you use a CMS, PIM, DAM, Git repository, or translation management system, ask whether they work through native connectors or controlled file exchange. Connector workflows remove the manual copy-and-paste step where most file-integrity defects begin.

How should you evaluate AI translation and data protection?

Separately from linguistic quality. ISO 18587:2017 gives you a reference point for machine translation post-editing, and ISO/IEC 27001:2022 gives you one for how information should be controlled.

AI translation can cut time and cost on the right content, but it changes what you need to ask. A provider should state which content types are eligible, which are excluded, and what human review applies to each. Marketing headlines, regulated claims, legal obligations, and safety instructions usually need tighter control than knowledge-base drafts.

Ask whether your content trains public models, and get the answer in the contract rather than the conversation. If they use machine translation engines, ask about the deployment model, retention policy, access controls, and data segregation.

For EU personal data, the GDPR has applied since 25 May 2018. Under Article 83 of Regulation (EU) 2016/679, certain infringements can carry fines up to 20 million euros or four percent of total worldwide annual turnover, whichever is higher. Translation workflows routinely carry names, employee records, customer tickets, medical information, contracts, and unpublished financials. Security belongs in the selection criteria, not in a review six months later.

Pricing should reflect these choices. Raw output, a light post-edit, a full post-edit, and human translation with independent revision are four different services. A single blended rate hides which one you are buying.

What multilingual SEO evidence should you ask for?

Process, not capability claims. Google Search Central requires each language or regional page to identify its alternates, and hreflang annotations have to be reciprocal to be interpreted correctly.

A company working on web content should know the difference between translating keywords and researching search demand in the target market. A US English product category may not map cleanly to French in Canada or Spanish in Mexico. Ask for localized keyword research, metadata recommendations, URL handling guidance, and content adaptation where a literal translation would miss buyer intent.

Ask how they QA hreflang. A common defect is declaring hreflang=”fr-ca” on the Canadian French page with no return tag from the English Canadian page. Another is an invalid code such as en-UK where the valid BCP 47 tag is en-GB. Small strings, large consequences, and neither shows up in a translation review.

Ask about governance too. Who approves translated title tags? Who checks meta descriptions for length and call-to-action fit? Who confirms localized pages are indexable after launch? A company offering multilingual SEO should answer those without pushing every decision back to your web team.

How should you compare cost?

Across a realistic twelve months, not per word. The unit price is the least informative number in a translation proposal, because it hides what service you are actually buying.

Ask for pricing that separates human translation with revision, post-editing, review-only work, desktop publishing, localization engineering, and rush handling. Then ask three questions that will otherwise surface later.

• How does translation memory reuse apply, and at what discount bands? A repetitive programme should get cheaper by year two. If it does not, the memory is either not applied or not maintained.
• What is billable outside translation? Project management, connector maintenance, minimum charges, file preparation, change requests. All legitimate. The question is whether they are disclosed before signature.
• Who owns the translation memory, termbase and bilingual files? If that is unclear, the cost of leaving is unbounded, and it will shape every renewal conversation you have.

Be sceptical of universal multipliers. No reliable public benchmark gives a fixed ratio between workflow tiers that holds across languages, industries and content types. Model your own mix instead. Take a representative month of real content, price it under each candidate workflow, compare the totals.

How should you score and pilot the finalists?

Use a weighted scorecard after the discovery call, then pilot with real files. Weight process evidence, security, technical capability, subject-matter fit, and total cost rather than comparing rates.

The same provider might be excellent for marketing transcreation and weak for software strings, or strong in European languages and thin in Japanese life sciences. Weigh the criteria to your content mix rather than using the defaults unchanged.

Design the pilot to fail informatively

Send every finalist the same four assets: One high-value web page, one software resource file, one legal or regulated excerpt, and one terminology-heavy product asset. Same brief, same deadline. Ask each to submit their queries as well as their output. The query log tells you whether they noticed the ambiguities you left in deliberately, and it is usually more revealing than the translation.

During reference checks, ask about responsiveness after delivery. Translation programmes rarely fail over an invoice line. They fail because feedback goes unmanaged, terminology changes do not propagate, or a launch team finds broken files too late. If a provider cannot show how they learn from corrections, treat it as a serious gap.

How does GPI score against this?

GPI by the numbers

Operating since 2001. Over 200 languages. More than 500 enterprise clients, including Fortune 1000 companies. 164,000 completed projects informing the ARTEE 1000 engine. Four ISO certifications with certificates published for download: ISO 17100:2015, ISO 18587:2017, ISO/IEC 27001:2022 and ISO/IEC 27017:2015, the last with all 37 cloud controls implemented. Fourteen native CMS and DXP connectors plus a Translation Services API, free to configure.

It would be a poor guide that told you to demand evidence and offered none. Here is GPI against the same five criteria. Every certificate is published, so you can check before we ever speak.

We work across professional translation, website localization, software localization, multilingual SEO, and AI and machine translation. Whether you choose us or someone else, apply the same tests to everyone.

Frequently asked questions

1- What is the best way to choose a translation company?

Define your content risk first. Then evaluate against ISO 17100:2015 quality practices, security controls, technical file handling, subject-matter expertise, and reporting. Use a weighted scorecard and a real pilot file before signing anything long-term.

2- How much should price matter?

After the workflow is clear. A low per-word rate gets expensive when it excludes revision, terminology management, desktop publishing, CMS handling, SEO adaptation or post-launch fixes. Compare total programme cost over twelve months.

3- Should we use AI translation or human translation?

AI where speed, cost, and controlled post-editing suit the risk profile. Human translation with independent revision for regulated, brand-sensitive, legal, technical or revenue-critical content. ISO 18587:2017 is the reference point for post-editing.

4- What certifications should a translation company have?

ISO 17100:2015 for translation service requirements, ISO 18587:2017 where post-editing is offered, and ISO/IEC 27001:2022 for information security. Certification proves an audited management system. It does not prove it is fit for your content, which is what the pilot is for.

5- How do we know if they can handle website localization?

Ask for evidence of multilingual SEO, hreflang QA, CMS workflows, HTML language tagging, metadata localization, and launch testing. They should understand Google localized-version guidance and BCP 47 tags such as fr-CA and en-GB.

6- How long should an evaluation take?

Six to twelve weeks is typical. Two to three weeks to define requirements and issue the RFP, two to three for responses and scoring, two to four for a pilot with real files, and a final week for references and contracting. Compressing the pilot is the most common shortcut and the most expensive.

7- How many translation companies should we work with?

Many organizations do best with one primary provider and carefully managed specialist backups. A single provider improves terminology consistency and reporting. Multiple providers help with niche domains, regulated markets, or surge capacity. Let content risk decide.

8- Who should be involved in the decision?

Localization or content operations owns the process. The decision needs IT security for data handling, legal for regulated content and contract terms, engineering for file formats and integrations, procurement for commercial terms, and in-country stakeholders for locale judgment. Bringing security and engineering in afterwards is how a signed contract becomes a stalled implementation.

Where to start

The right company is the one that can prove fit before the first large project: Quality discipline, secure data handling, domain-qualified linguists, technical capability, and transparent workflow-based pricing.

If you do one thing this week, take the scorecard above and apply it to your current provider. Not to replace them, but to find out which of the areas you have never actually asked about. That gap is usually where your next problem comes from.