Translation Memory and CAT Tools: Where the Savings Come From

Localization is easy to budget badly. The instinct is to treat it as a per-word cost that multiplies cleanly: so many words, times so many languages, times a rate. By that logic a product with twice the content, or twice the target markets, costs twice as much to localize, and every update restarts the meter from zero. Teams that run multilingual content this way watch the bill climb with each release and conclude, reasonably enough, that going global is simply expensive.

A well-run localization programme does not behave that way, and the reason is infrastructure that is set up before the first sentence is translated: translation memory, a managed glossary, and the CAT tools that tie them together. None of this is glamorous, but it is where the economics of localization are actually decided. This article explains what each piece is, the mechanics of how it reduces cost, and the consistency it buys along the way. The broader question of when to translate, localize, or transcreate is a separate decision covered in our strategy guide at /blog/localization-vs-translation-strategy/; the focus here is narrower and more practical.

CAT stands for computer-assisted translation, and the phrase is worth taking literally. A CAT tool does not translate; a human linguist does. What the tool provides is an environment that makes a translator faster, more consistent, and cheaper to employ across large and repetitive bodies of content. It is the difference between a linguist working in a word processor, retyping similar sentences and trying to recall last month's terminology, and a linguist working in a system that remembers for them.

Mechanically, a CAT tool breaks a document into segments, usually sentences or sentence-like units, and presents each one with its source on one side and an editable target on the other. As the linguist works, the tool consults two databases in the background: a translation memory of everything translated before, and a termbase, or glossary, of approved vocabulary. It also runs automated checks, flagging mismatched numbers, missing tags, inconsistent terms and untranslated segments before a file moves on. Native-speaking linguists do the actual translation with CAT tools; the software removes the mechanical drag around them rather than replacing their judgement.

Translation memory, usually shortened to TM, is the single biggest lever on localization cost. It is a database of segment pairs: every source sentence the programme has translated, stored alongside its approved translation in each language. When new content arrives, the CAT tool compares each incoming segment against the memory and classifies it by how closely it matches something already there.

Those match categories are the heart of the savings model. An exact match is identical to a stored segment and can be reused verbatim. A fuzzy match is similar but not identical, so the linguist edits an existing translation rather than writing a new one. Repetitions are segments that recur within the same project, translated once and propagated. Everything else is new and is translated from scratch. Because reused and partially reused content costs far less to produce than fresh translation, vendors price these categories at reduced rates, and genuinely repeated content is, in the words of our own localization process, never paid for twice.

The first time content is translated, translation memory saves little, because the memory starts empty and most segments are new. This is the point buyers most often misread, expecting a discount on a first project that has nothing to leverage yet. That first translation is an investment: it populates the memory. The return arrives afterwards, and it compounds in three predictable places.

The first is internal repetition. Long manuals, product catalogues and technical documents repeat themselves more than their authors realize, so a single project can carry a meaningful share of repeated and near-repeated segments before any history exists. The second is versioning: when a manual reaches its second edition or an app ships its next release, only the changed segments are new, and the rest match the memory and flow through at reduced cost. The third, and the largest over time, is continuous content such as help centres, knowledge bases and product strings that change a little every month.

A concrete illustration from our work: a B2B SaaS company maintained roughly 1,400 help-centre articles across twelve languages while shipping updates every month. Rather than re-quote each update from scratch, the programme translated only the segments that had changed and let the memory cover the rest; across the ongoing volume, around 38% was handled through translation memory at reduced rates, flattening what would otherwise have been a linear monthly bill. Two caveats keep expectations honest. Leverage is real only where content genuinely repeats or recurs, so one-off prose gains little; and creative or marketing copy, where the aim is effect rather than fidelity, is better treated as transcreation.

If translation memory is mainly about cost, the glossary is mainly about consistency, though the two are closely linked. A termbase records the approved translation of every term that must not vary: product names, technical vocabulary, terms that should be left in English, and terms that are forbidden, often with a note on usage or context. As the linguist works, the CAT tool surfaces the approved term in place and flags any deviation, so the same concept is rendered the same way on page one and page nine hundred, by every linguist on the project and in every later update.

The value of this is easy to underrate until it is missing. Without a managed glossary, a single feature can acquire three different names across a help centre, a manual and an interface, and that inconsistency reads as carelessness to customers while creating confusion for support staff in the field. In a regulated or safety-critical context it is worse than untidy. This is why a reviewer's job, in our process, is not only to confirm a translation is correct but to check that terminology, numbers, formatting and tone match the glossary, with a second linguist reviewing every file before sign-off.

Consistency is also a quiet cost saving in its own right. Terminology settled once, in the glossary, is terminology not re-argued in every file and not corrected in every review cycle. The rework that an unmanaged programme spends chasing terminology churn often dwarfs the per-word difference between one vendor and another.

Because the memory and the glossary do the saving, the highest-leverage moment in a localization programme is the setup before translation begins. In our localization process this is a distinct first step: review the files, agree terminology and tone, and stand up the translation memory so that repeat content is identified from the outset rather than discovered halfway through. Skipping it does not make a project cheaper; it simply moves the cost downstream, into rework and inconsistency.

Clients can improve their own leverage at this stage by bringing existing assets to the table. Previous translations, even if produced elsewhere, can often be aligned and imported to seed the memory before a single new word is translated. A list of approved terms, a few reference documents, and a short style note give the glossary a head start. The same preparation reduces cost across every target language at once.

Understanding the machinery changes how a localization quote reads. A quote built on CAT tooling will usually present a weighted word count: the total volume broken into match categories, each with its own rate, rather than a single price per word. A flat per-word number with no analysis is a sign that no leverage is being applied, which often means paying full price for content that has already been translated once.

The same understanding sharpens the questions worth putting to any localization vendor; the answers separate a programme that compounds in your favour from one that simply bills per word forever.

The throughline is simple. The cheapest word to localize is the one you never pay to translate a second time, and the discipline that makes that possible — memory, glossary, and the tooling that enforces both — is set up long before the first deadline. Our localization service line at /localization/ is built around that machinery, in more than seventy languages with native-speaking linguists and a second-linguist review on every file.