International SEO: hreflang, domains and the mistakes that cost years

Most SEO work is reversible. A weak title tag is a five minute fix, a thin page can be rewritten. International architecture is different. Once your markets are mapped onto domains and URLs, and search engines have spent two years learning that structure, unwinding it means a full redirect map, a fresh crawl and indexation cycle in every market, hreflang rebuilt from scratch and authority that has to settle again across new addresses. Teams that migrate an international estate usually plan for one quarter of disruption and live through two.

The expense is not only technical. Domain choice drags legal, financial and operational commitments behind it: an entity in the market, local payment methods, customer service in the language, returns handled locally, sometimes a physical address. Those commitments are why architecture debates stall in meetings. An SEO can say which structure is easiest to rank; only the business can say which structure it will still staff in five years. The right answer sits at that intersection, and it is usually decided far too early, by whoever set up the hosting.

What follows is a working map: the three architecture options and what each really costs, how hreflang behaves and how clusters break without warning, why language and country are different targets, where machine translation is now acceptable and where it still destroys trust, how markets diverge in search behaviour, and how to run several at once without a translation bottleneck. A pass through a technical SEO audit checklist beforehand is time well spent, because international problems are usually domestic problems multiplied by the number of markets.

Country-code domains (example.de, example.fr) give the clearest signal of local intent and the strongest trust with users, who in several European markets read a national extension as proof of real presence. The price is that every domain starts from zero. Authority does not flow between them, so links, mentions and digital PR must be earned five times for five markets. For a group with real local operations and local budgets, that cost is worth paying. For a company testing a market, it buys five weak sites instead of one strong one.

Subfolders, meaning example.com plus a language directory, sit at the other extreme. Everything you earn anywhere strengthens everything else, migrations are cheap, one CMS and one analytics property serve every market, and a new language goes live in weeks. The trade-off is that one technical incident, manual action or botched migration hits all markets at once, and some audiences trust a national domain more than a language folder. Subdomains are the compromise nobody quite loves: cleaner separation for hosting and teams, but weaker on both signals they borrow.

In practice the default for most companies below enterprise scale is subfolders, and the exceptions are worth naming. Choose country domains when the brand name differs by market, when legal or regulatory separation is required, when national players dominate a vertical and a foreign address reads as an outsider, or when you already hold strong national domains worth keeping. Partners used to multi-market work push back on the fashionable answer and ask about your operations first; habits differ: compare how the German agency market approaches this against agencies in Spain.

hreflang is not a ranking factor. It is a disambiguation instruction telling a search engine that several URLs carry the same content prepared for different audiences, so the right variant is served instead of the wrong one. Declare it in the head of each page, in HTTP headers for files like PDFs, or in the XML sitemap, usually the sanest choice on large sites because it centralises maintenance. What hreflang never does is push a page into results where it does not belong. Weak content is not rescued by perfect annotations.

Two mechanics cause most of the damage. The first is reciprocity: every page in a cluster must reference every other, itself included, and a German page pointing at a French one without a return pointer may see both discarded. The second is x-default, the fallback for visitors matching nothing in the set. It is optional, but omitting it on a site with a global selector or generic English version means a fallback is chosen for you. Both fail quietly: nothing in your traffic report announces that hreflang broke last Tuesday.

The errors are boringly consistent across audits: invalid codes such as en-UK where en-GB is meant, or a country code sitting in the language position; annotations pointing at redirected or canonicalised URLs; hreflang combined with a canonical naming a different variant; and clusters truncated when a CMS republishes only part of the set. Symptoms surface as the wrong country version ranking, or two variants trading places weekly. Search Console flags some of it, but the fastest diagnosis is crawling each market and comparing the declared cluster against the one that exists.

A language target and a country target answer different questions, and conflating them is the second most expensive mistake after architecture. Spanish is spoken by far more people outside Spain than inside it: a plain Spanish page serves all of them, while country variants serve specific markets with their own vocabulary, currency, shipping and competitors. English is the same problem at larger scale. Country granularity without country-specific content produces near duplicates that dilute each other. Language granularity, when your offer differs by country, promises a service you cannot deliver at that address.

Multilingual countries expose the confusion fastest. Switzerland needs German, French and Italian variants all targeted at the same country; Belgium needs Dutch and French; Canada needs English and French. A single country target forces one language on half the audience; a single language target sends Swiss buyers to German pricing and delivery terms. The rule in practice: split by country only when something material differs, meaning price, stock, legal terms, payment or contact details, and split by language whenever a reader would otherwise land on a page they cannot read.

Geo-targeting settings and hosting location are weaker signals than most people assume. A country domain carries its target implicitly; for everything else the signals that work are the content itself, the currency, the address and phone formats, the internal linking and the links you earn from sites in that country. Automatic redirection by IP address is the recurring error: it hides variants from crawlers, traps travellers on the wrong version and interacts badly with hreflang. Offer a visible, persistent language selector instead, remember the choice, and keep every version directly reachable.

Machine translation in 2026 is genuinely good, and acceptability now depends on the stakes rather than on the technology. Raw output is fine for content whose job is coverage: help centre entries, specification tables, long-tail informational pages. It is not fine for anything that must persuade or carries risk: landing pages, pricing, legal terms, medical or financial content, anything where a slightly wrong verb makes a brand sound foreign. The failure is rarely grammatical, it is a register subtly off in a way native readers notice but cannot articulate.

The workable middle is machine translation followed by a native editor, usually called post-editing, with the editorial brief attached, not only the source text. It costs a fraction of full human translation and removes most of the damage, provided the editor may rewrite rather than correct. Google treats automated translation published without human review as low value, but the practical test is simpler: if a native speaker there would not have written the sentence that way, the page is not ready. The same applies to AI-assisted content in general.

Duplicate content across languages is mostly a myth: the same page in French and German is localisation, not duplication, and hreflang exists to describe that relationship. The real risk is duplication within one language: three English variants for Britain, the United States and Australia differing only by a currency symbol. Those compete, split link signals and confuse selection. If you cannot justify a country variant with content a local reader would recognise, target the language instead. Sustained localisation is a content marketing discipline more than a translation task.

Markets do not merely speak differently, they search differently. Query length and structure vary: German compounds pack a whole phrase into a single token, French searchers use more prepositions and full questions, and transliteration matters wherever a script boundary exists. Seasonality moves, sale periods differ, and intent behind an apparently identical keyword can flip, so a term that is commercial in one country is purely informational in another. Keyword research translated from your home market is the commonest cause of a content plan that ranks for terms nobody uses to buy anything.

The competitor set changes too, usually more than teams expect. In many European markets the top of the results is held by national comparison sites, marketplaces and press brands no global tool flags as competitors, and in some verticals the effective search engine is a marketplace or a local aggregator. Result features differ too: shopping surfaces, local packs, video and AI answers with different citation habits per language. Two markets with identical volume can therefore need different plans, which is why catalogue-driven sites rarely scale one template across borders.

So every market needs its own research pass before its own content plan: keywords collected in the language, not translated into it, intent verified on the actual results page, competitors identified from that page rather than a global database, and a check of what sits above the organic links. That is at most a week per market, and it decides whether the next six months produce anything. Buyers comparing SEO agencies in France with agencies in the Netherlands often find their strengths are not comparable, because the markets are not.

The bottleneck in most multi-market programmes is not writing, it is waiting. A central team writes in the source language, hands finished pages to translation, and every market advances at the speed of the slowest queue. The alternative is to brief rather than translate: define the topic, the intent, the structure and the non-negotiable claims centrally, then let each market produce its own version, borrowing from the source where it fits and departing where the market differs. Output rises, review load falls, and pages stop reading like documents that arrived from head office.

Two artefacts make that model work. A terminology glossary fixes product names, feature names, legal wording and the handful of terms that must never drift, which also keeps entity signals consistent for search engines and AI systems. A shared brief template forces the same thinking in every market without forcing the same sentences. Around those, a small bench of native editors, often independent specialists retained a few days a month, is usually cheaper and faster than a translation agency, and better than an internal generalist working through a dictionary.

Choosing a partner for a market nobody internally can read is the last hard problem. Replace language judgement with evidence of results. Ask for verified Search Console data from that market, not global portfolio figures. Ask who on the team is native, and speak to that person. Commission one paid pilot page, reviewed blind by an unrelated native speaker, before committing. Rankings built on verified performance, such as an agency leaderboard, help here, and the method in our guide to choosing an SEO agency holds across borders. Start from a country-filtered agency directory, not a search you cannot evaluate.