Until recently, expanding into a new market meant one of two things for a small business. Either you ran the same advert in English everywhere and hoped for the best, or you paid thousands of dollars to an agency to translate and re-record every piece of content.<\/p>\n
Neither option worked for a team without a marketing department. This gap is rapidly narrowing: artificial intelligence tools now handle translation, dubbing and the adaptation of visuals at a fraction of the cost of the past.<\/p>\n
A survey conducted by CSA Research, published on 7 July 2020 under the title \u2018Can\u2019t Read, Won\u2019t Buy\u2019, involved 8,709 consumers across 29 countries. The survey found that 76 per cent prefer to buy products with information in their own language, whilst 40 per cent would never buy from websites in other languages. Ignoring this preference quickly becomes costly.<\/p>\n
The cost difference between the old and new approaches is considerable. The cost analysis published by Localize in 2026 puts the cost of full human translation at between 0.08 and 0.30 dollars per word, i.e. between 80,000 and 300,000 dollars per million words, whilst raw machine translation costs as little as $0.001 to $0.005 per word, or between $1,000 and $5,000 per million words. The intermediate option, machine translation with human post-editing, overlaps with the lower end of the human translation range, costing between $50,000 and $150,000 per million words.<\/p>\n
This is the difference between localisation reserved for major brands and a realistic option for a team of five people. The steps below show how small businesses can effectively use artificial intelligence to localise advertising campaigns, without a team of translators and without a large budget.<\/p>\n
Literal translation is the quickest way to make an advert seem out of place. Idiomatic expressions, humour and calls to action rarely survive a word-for-word translation. A punchy English slogan can become confusing in another language.<\/p>\n
Modern machine translation tools go beyond simple dictionary-based substitution. Platforms such as DeepL and workflows based on model training can adapt the tone and expressions to the target audience.<\/p>\n
For a small business, this means writing a single good advert, followed by using the model to generate culturally adapted versions for each market. A simple starting point: draft the core advert in your main language, then provide the tool with context about the target market before asking it to produce a version. Specify the audience, platform and desired tone, as this context significantly affects how natural the result sounds.<\/p>\n
You don\u2019t need a separate tool for each market if the platform you\u2019re using already includes this feature. On 9 October 2025, Meta announced the automatic translation and dubbing of Reels videos using artificial intelligence, with the option of lip-syncing, starting with English, Spanish, Portuguese and Hindi. The announcement has been updated several times since then: in January 2026, Indian languages such as Bengali, Tamil, Telugu, Marathi and Kannada were added, and by the summer of 2026, French, German, Italian, Japanese and Korean on Instagram, and Arabic, Indonesian, Thai and Vietnamese on Facebook.<\/p>\n
An important clarification: the feature described by Meta applies to clips published organically by creators, not to paid campaigns, so for adverts, the translated material must be uploaded separately. The figures circulating regarding the impact of similar features on costs \u2013 such as reductions in CPM in localised markets or cuts in production costs \u2013 stem from third-party analyses, which present them as best-case scenarios, rather than from Meta\u2019s own press releases. They should be treated as such, not as planning benchmarks.<\/p>\n
Google Ads takes a different approach. The documentation recommends separate resource groups by language in Performance Max campaigns, and the automated text features generate additional headlines and descriptions based on your page, rather than translating existing content. For a small business, enabling the platform\u2019s native features remains the first step requiring the least effort, as it costs nothing on top of the advertising budget already allocated.<\/p>\n
Check your advertising account settings before building anything custom. Many small businesses jump straight to hiring an external contractor, without realising that the platform already offers a version of this. Native features don\u2019t cover every nuance, but they\u2019re a quick way to test whether a market is worth a more serious investment.<\/p>\n
Text and voice are only half the story. Colours, gestures and images carry different meanings from one culture to another. A gesture that seems friendly in one country may be offensive in another.<\/p>\n
Image generators can quickly produce or adjust visual variations: they can change backgrounds, alter the demographic profile of the people in the image, or rearrange the overlay text for languages read from right to left. A single piece of source material, run through an AI-powered editor, can yield several locally tailored versions in a single afternoon.<\/p>\n
This used to require a photo shoot for each market, plus a local creative director to approve it. Now it involves a set of instructions and a few rounds of checking \u2013 a difference that makes it realistic for a small team to test three or four markets simultaneously.<\/p>\n
This step is easy to overlook, as it is less obvious than translation. Omitting it is also the quickest way for a localised advert to still feel out of place.<\/p>\n
Even small adjustments help. Changing currency symbols, local landmarks or seasonal references shows genuine effort, and viewers notice the difference, even if they can\u2019t quite put their finger on why.<\/p>\n
Subtitles are cheap, but they yield poor results in feed-based formats such as Reels or TikTok. Dubbing keeps the viewer\u2019s gaze on the image, rather than drawing it to the bottom of the screen.<\/p>\n
The difficulty with dubbing has always been lip-syncing. The translated audio rarely matches the speaker\u2019s lip movements on screen, which makes the dubbing look obviously artificial. Lip-syncing solves this problem by adjusting the visible movement of the mouth to match the audio in the new language.<\/p>\n
Tools such as lip sync API<\/a> from VEED, i.e. the application programming interface for lip-syncing, allow developers and content creators to integrate this feature directly into their video production workflow. This is useful for small businesses that already produce their video adverts in-house or with a partner, as it adds multilingual dubbing without having to rebuild the entire workflow.<\/p>\n For a business without development resources, there is a simpler way to achieve the same result. Dubbing apps designed for the general public, combined with the native features of the platforms mentioned in step 2, cover most needs without a single line of code.<\/p>\n Either of these two approaches is better than the old alternative: reshooting the footage in every target language. This used to involve bringing in local actors or hiring a production team for each market, whereas lip-synced dubbing delivers a version that sounds native at a fraction of that cost.<\/p>\n Artificial intelligence makes it cheap to produce multiple variants, which creates a new problem: knowing which ones work. Run a small test batch, with two or three variants, before increasing your spend in a new market.<\/p>\n Let the platform\u2019s optimisation algorithm identify the winner. A sentiment analysis tool can also scan the comments on the first few posts to spot cultural missteps, as wording that passes internal testing may be received very differently when presented to a real audience.<\/p>\n Allocate the budget for this testing phase from the outset, rather than treating it as an afterthought. Even $50 to $100 spent on the market will give you real data on interaction, rather than just an assumption, and that data can determine where the bulk of the budget goes.<\/p>\n An advert is just the first point of contact. If a shopper clicks on a Spanish-language advert and lands on a page that\u2019s only in English, most of the localisation work is wasted. The same CSA research shows why: 40 per cent of its respondents would never buy from a website in a language other than their own.<\/p>\n Website localisation tools can translate landing pages, checkout flows and product descriptions. This doesn\u2019t mean a fully translated website from day one: even a single localised landing page for each campaign goes a long way towards bridging the gap between click and conversion.<\/p>\n Prioritise the pages that carry the most weight in the purchasing decision. A localised product page and a localised checkout process matter more than a translated footer or \u2018About Us\u2019 page. Start there and expand the rest of the site once the market has proven its value.<\/p>\n None of these steps require an in-house localisation department. A small business can start with Meta\u2019s and Google\u2019s translation and dubbing features, then add automatically generated visuals and lip-synced video dubbing for campaigns that warrant a greater effort.<\/p>\n Tools have advanced quickly enough that cost is no longer the main constraint. The trickier question is which markets deserve priority and how to establish a simple routine for checking the model\u2019s output.<\/p>\n Start by choosing the market with the greatest potential, not necessarily the largest one. A smaller market, where there is already organic demand for your product, will teach you more than a large, unfamiliar one. Follow the steps above there, measure what happens, and expand once you\u2019ve identified what actually converts.<\/p>\n The observation that opens the first step \u2013 that word-for-word translation makes the advert sound foreign \u2013 is underpinned by one of the most frequently cited demonstrations in the history of computational linguistics. In 1960, the logician Yehoshua Bar-Hillel published, in the first volume of the Advances in Computers series, a report on the state of machine translation, accompanied by an appendix entitled \u2018A Demonstration of the Infattibility of High-Quality Full Machine Translation\u2019. The proof relies on a trivial sentence, set in an equally trivial context: a child is looking for their toy box, eventually finds it, and the text states that \u2018the box was in the pen\u2019. The word \u2018pen\u2019 in English means both a pen and a playpen; the human reader instantly chooses the second option, because they know what fits inside what.<\/p>\n Bar-Hillel argued that no existing or conceivable programme would be able to make this choice, as it stems from knowledge of the relative sizes of objects in the world, rather than from language; the suggestion of equipping the machine, in addition to a dictionary, with a universal encyclopaedia struck him as downright chimerical. Half of his conclusion has been disproved: contemporary models effortlessly solve the pen example, having statistically absorbed a considerable portion of the knowledge of the world that he believed impossible to codify. The other half remains valid, and it is this that matters here: disambiguation depends on knowledge of the world being discussed, not on knowledge of the language in which it is discussed. From this it follows directly why a model translates a technical product description flawlessly yet stumbles over a slogan: the slogan refers to a world that the model knows only unevenly, and adverts thrive precisely on such references.<\/p>\n It is worth noting the limitations of this interpretative framework. Bar-Hillel wrote before the advent of statistical translation and long before neural translation, and his pessimism regarding quality has been disproved by the facts; the use of his argument today is an interpretative extension, legitimate insofar as we maintain the distinction he introduced between linguistic competence and knowledge of the world, and abandon his verdict of impossibility. The distinction also applies to steps three and five: visual adaptation and audience testing supplement the model\u2019s lack of world knowledge regarding that particular market, as the language aspect has already been resolved.<\/p>\n Once the production of variants becomes inexpensive, the constraint shifts to their verification, and here the small business encounters a difficulty that has been touched upon without being named, when a simple verification routine is recommended. The person commissioning the translation does not, as a rule, speak the language into which it has been translated; they cannot judge whether the tone is appropriate, whether a phrase sounds businesslike or obsequious, or whether a joke has come across. The \u2018sentiment test\u2019 on comments, proposed in step five, helps, but it comes after publication \u2013 that is, after the mistake has already been seen by the public.<\/p>\n There is, however, a category of errors that can be spotted without knowledge of the language, and it is worth listing, as it is also the category with the most costly consequences: the company\u2019s proper name, written identically throughout; the address and legal form; the telephone number in international format; the currency, date format and decimal separator; units of measurement; and the mandatory information regarding price, delivery and the right of withdrawal, where required by the law of the target market. All of these are verified by comparison with a reference document, not by linguistic judgement. A translation model sometimes alters them without flagging the changes, in an effort to make the text sound natural. The mechanism is the same when applied to company data: models predict patterns, and when a detail is missing from the context, the plausible form takes the place of the correct one. Comparison with a stable record therefore remains the most practical defence, the logic underpinning company directories used as reliable anchors against hallucinations<\/a>, with the caveat, acknowledged by their author, that coverage of small and new companies is poor, and updates are delayed.<\/p>\n The right verifier is neither the translator who produced the text nor a relative who speaks the language, but someone who lives in that market and buys from that category, because the purpose of verification is commercial plausibility, not accuracy. The questions worth asking this reader are few and precise: what impression does the tone leave on them, which phrasing strikes them as unnatural, which promise seems exaggerated given how products are sold there, and what is missing from what they would expect to see. The last question is the most productive, as the model cannot flag the absence of something it does not know is customary.<\/p>\n This leads to a working rule that requires no budget: anything that can be compared to a written document is checked before publication, whilst anything relating to register and nuance is given to a native speaker to read through, just once. An hour\u2019s proofreading paid for on the relevant market costs less than a day\u2019s campaign budget and fills precisely the gap that Bar-Hillel identified.<\/p>\n Prioritising the product page and the payment flow is correct in terms of spending order, but incomplete as a description of buyer behaviour. A visitor from a market where the company is unknown makes, before payment, a move that no localisation tool monitors: they search for the company\u2019s name separately, to see what appears about it outside its own website. In the home market, this action encounters years of accumulated traces; in a new market, it usually encounters an advertising page and nothing else, and this absence is interpreted as a signal.<\/p>\nStep 5: test the message before committing the budget<\/h2>\n
Step 6: identify what happens after the click<\/h2>\n
How the steps fit together<\/h2>\n
Why literal translation fails: a problem dating back over sixty years<\/h2>\n
The verifier\u2019s problem<\/h2>\n
What the buyer is looking for after clicking, beyond the landing page<\/h2>\n