Where AI Translation Still Fails in Marketing

Machine translation got good enough to change the economics of global content. It is still unreliable for exactly the pages that carry your argument, and the failure mode is that it reads perfectly fine.

By Pierre Subeh, published June 25, 2026, 6 minute read

Machine translation crossed a threshold in the last few years. For a large share of content it is now genuinely good, fast, and cheap enough to change what a small team can attempt internationally.

It is also still wrong for the specific content that matters most, and the failure mode is dangerous precisely because the output reads perfectly fine. A bad translation used to be obvious. A modern one is fluent, confident, and subtly off in ways that only a native speaker with commercial judgment will catch.

What It Handles Well

Be clear about the wins, because refusing to use these tools is now its own kind of mistake.

Documentation, help articles, knowledge bases, and support macros translate well. The language is functional, the terminology is consistent, and the reader wants information rather than persuasion. Machine translation with a light editing pass is the correct approach here, and paying human rates for it is a waste of money that should be spent on the pages that need it.

Internal content, user generated content, and long tail archive material fall in the same category. So does the first draft of almost anything, as raw material for a human writer who will rebuild it.

That is most of the volume on a typical site, which is why the economics changed.

Where It Fails, and Why the Failure Is Invisible

Persuasion. The output is accurate and inert. The argument survives, the reason anyone should care does not. Homepage copy, value propositions, and campaign lines come back correct and flat, and because they are correct, reviewers approve them.

Humor, wordplay, and idiom. These do not translate, they have to be reinvented. A model will produce a literal rendering of a pun that reads as a mistake, or worse, an unintended second meaning nobody checked.

Register. The distance between formal and casual differs by language, and several languages encode social relationship grammatically. A model asked for the local equivalent of friendly American brand voice will often land somewhere that reads as either presumptuous or oddly stiff. It will not tell you it made that choice.

Terminology with commercial weight. Regulated language, category names that are contested, and terms where the obvious translation carries a connotation the original does not. In financial and health contexts this is a compliance issue rather than a style one.

Cultural references. Sports metaphors, holidays, historical allusions, and anything assuming shared media. These translate into confusion at best.

The common thread is that every one of these produces output that looks right. There is no error signal. Teams discover the problem through underperformance months later, if at all.

A Staffing Model That Reflects This

The useful move is to stop treating translation as one activity with one budget and split it explicitly.

Machine translated, spot checked. Help content, documentation, archives, internal material. Review a sample rather than everything. Accept that some imperfection will ship, because it will and it does not matter much here.

Machine translated, fully edited by a native speaker. Product pages, most blog content, email sequences. The model does the first pass, a human with commercial judgment rewrites for register and clarity. This is where most of your human budget should go, because it covers the most pages that carry any persuasive load.

Written from scratch by a native speaker. Homepage, primary campaign copy, pricing, taglines, and anything a journalist might quote. Do not translate these at all. Brief a local writer from intent, the same way you would brief the original.

The mistake is not using machine translation. It is using one tier for everything, which either wastes money on documentation or ships a machine written homepage.

Prompting Improves the Middle Tier

For content in the editing tier, how you brief the model changes how much editing is needed.

Give it the audience, the register, and the purpose rather than only the text. Telling a model that this is homepage copy for small business owners in Mexico, that the tone should be warm and direct, and that the goal is a free trial signup produces materially better raw material than passing the sentences alone.

Provide a glossary of terms that must not change, including product names, and terms that must change in a specific way. Models are inconsistent about terminology across a long document, and inconsistency is one of the clearest signals of machine output to a reader.

Ask for options on anything short and load bearing. Three candidate headlines with an explanation of the connotation of each gives a human reviewer something to choose between, which is a much easier task than judging a single output in isolation.

And ask the model to flag what it was unsure about. Modern models will identify idioms and ambiguous phrasings if you ask them to, which turns an invisible failure into a review queue.

You Still Need a Native Reviewer

Every market that matters commercially needs at least one native speaker with commercial judgment who reads the important pages before they ship. Not a translator checking accuracy. A marketer checking whether this works.

The distinction matters. A translator will confirm the meaning was preserved. That is not the question. The question is whether this page makes someone in this market want to buy, and only a marketer from that market can answer it.

If you cannot fund that for a market, you probably should not be entering that market yet. Reviewing the pages that carry the argument is a small commitment against the cost of the traffic you are about to send at them.

The Line, Stated Simply

Use machine translation for content people read to learn something. Use humans for content people read to decide something.

That line will keep moving, and I expect the models to keep taking territory from the middle tier. It has not yet moved past persuasion, and the reason is not really a language problem. Persuasion requires knowing what a specific group of people already believes and what would change their mind. That is a research problem wearing a translation costume, and it is the part worth staffing properly.

About the author

Pierre Subeh is Co-President of AMA Orlando and the founder and CEO of X Network, an SEO and paid marketing firm whose work spans Apple Music, Häagen-Dazs, and Pepsi. He is a Forbes 30 Under 30 honoree in marketing and advertising, a TEDx speaker, and a member of the PR and Media Council at the Council of Global Change, an independent international policy council that convenes at the United Nations. He writes about search, brand, and global growth at pierresubeh.com.

Topics: ai translation, localization, machine translation, global marketing, ai workflow