Most multilingual sites stop at translation. The pages exist in target languages, hreflang routes the right user to the right page, and the assumption is that visibility follows. It usually doesn't — because translation produces grammatically clean copy in the target language while leaving the keyword targeting, the cultural references, and the intent matching tied to the source market. A truly localized program does the harder work and earns the rankings; a translated program ships content and wonders why the international visibility never materialized.

“Translation gets you past the language barrier. Localization gets you into the target market's search behavior. The two are sometimes confused as the same step but they produce wildly different outcomes — and the difference is usually invisible until the rankings show the gap.”

What translation produces

A translated page reads the source-language content in the target language. Modern machine translation handles this well — DeepL, Google Translate, and LLMs produce grammatically clean output for most language pairs. The output:

• Grammatical and idiomatic in the target language.
• Faithful to the source's meaning and structure.
• Same H1, same H2 structure, same calls-to-action, same examples.
• Same keyword targets — translated literally from the source.
• Same length, same flow, same tone.

What it doesn't produce: rankings against local search behavior. The translated H1 is the literal translation of the English head term, not the term that local users actually search for. The keyword targeting is therefore wrong by default.

What localization adds

• Local keyword research. The Mexican Spanish translation of "best CRM for sales teams" is "mejor CRM para equipos de ventas" — but Mexican users may search "CRM ventas" or "software CRM" or other patterns the literal translation misses. Local keyword research surfaces these.
• Local examples. A US-focused page that mentions "Black Friday" needs to surface "Buen Fin" for Mexican readers. A New York city example becomes a Mexico City example. Examples that resonate locally drive engagement signals that translation alone can't produce.
• Local calls-to-action. "Sign up free" in English maps to "Regístrate gratis" — but the Mexican market often expects a clearer commitment phrase ("Empieza gratis" or "Prueba gratis"). The CTA language matters for conversion, which feeds engagement signals.
• Local pricing and currency. A page that quotes USD prices to Mexican users is friction; localization shows MXN. Format differences (date formats, number formats) also matter.
• Local trust signals. Testimonials from Mexican customers, case studies in the local market, references to local press or local partners. These feed the prominence signal beyond the linguistic layer.
• Tone and register. Mexican Spanish uses different formality conventions than Spain Spanish; British English uses different conventions than American English. The right register for the market is part of the localization, not optional polish.

The keyword research mistake

The single biggest localization mistake is using the source-language keyword list translated. The example that recurs in every market:

• Source: "best CRM software for sales teams" (US English).
• Translated: "mejor software CRM para equipos de ventas" (es-MX literal).
• Local-research alternatives: "CRM ventas" (higher volume in MX), "mejor CRM en México" (geo-modifier pattern), "CRM para PyMEs" (segment match for the Mexican market where small/medium business is the dominant segment).

The literal translation is a valid keyword in Spanish — it just isn't the highest-volume or highest-converting target in the local market. Local keyword research, run against Google's Keyword Planner with the geographic filter set to the target market, surfaces the right targets. The chapter on keyword research in the content cluster covers methodology; it applies once per market, not once per project.

The content tiers

Localization isn't all-or-nothing. The pragmatic budget split:

• Tier 1 — full localization. Homepage, top-3 product pages, top-3 SEO landing pages, top-3 conversion paths. ~10–15 pages per market. Per-page cost is high; the ranking and conversion lift justifies it.
• Tier 2 — translation + localization edit. Long-tail SEO content, FAQs, comparison pages, case studies. The bulk of the content. Machine-translate, then a local editor reviews for keyword targeting, calls-to-action, and cultural fit. Per-page cost moderate.
• Tier 3 — translation only. Terms of service, privacy policy, technical documentation, archived content. Machine translation is fine; no localization editing required. Per-page cost minimal.

Mistake to avoid: spending localization budget evenly across all content. Tier 1 over-investment compounds; Tier 3 over-investment doesn't.

The bilingual editor as a quality gate

Localization quality is hard to judge from outside the target market. The right control is a native-speaker editor with marketing or SEO experience in the local market, not just a translator. The editor's checklist:

• Does the keyword appear in the target-market vernacular, not just the literal translation?
• Are the examples local? (Cities, brands, holidays, cultural references.)
• Is the tone right for the market? (Formal vs informal, regional register.)
• Are the CTAs phrased in a way the local audience expects?
• Are pricing, currency, dates, and number formats local?
• Are idioms replaced with local equivalents (not literal translations)?
• Does the page read as if originally written in the target language, not translated into it?

AI translation in 2026

LLM-based translation (Claude, GPT, Gemini) has narrowed the gap between machine and human translation for most language pairs. The current state:

• Grammatical quality: excellent. The output reads naturally in most language pairs.
• Tone preservation: good with prompting; the model can be instructed on register.
• Local keyword matching: still weak. The model translates the keyword literally rather than substituting the local equivalent.
• Cultural reference adaptation: mixed. With explicit prompting, the model can substitute local examples; without, it translates verbatim.
• Idiom handling: good with explicit instruction; mediocre by default.

The pragmatic 2026 workflow: machine-translate the bulk of the content, run a localization editing pass with explicit instructions about keyword targeting, examples, and CTAs. The chapter on the content-editor agent in agency tooling automates the editor's checklist for both EN ↔ es-MX and EN-AU contexts.

The "untranslated brand asset" pattern

Some brand assets shouldn't be translated:

• Brand name (almost always — exceptions for brands that have established localized variants).
• Product names (usually — exceptions for product names that are deliberately localized).
• Slogans (depends — sometimes a localized slogan beats a translated one; sometimes the English slogan carries brand equity).
• Trademarks and registered marks (always preserved verbatim).

The decision per asset is a brand decision more than a translation decision; document it once per market and apply consistently across all localized content.

What localization doesn't fix

• Wrong URL architecture. If the structure is broken (no hreflang, language-only directories where you needed language-region), localization fills the wrong containers.
• Local-link absence. Localized content needs local backlinks to rank against local competitors; the next chapter, international link building, covers this.
• Local trust deficit. A new entrant to a market starts with no local authority; localization plus link profile plus PR plus time builds it.
• Bad source content. Localizing a poorly-targeted page produces a poorly-targeted page in another language. Fix the original first.

The next chapter, international link building, covers the off-page work that turns localized content into ranked content in the target market.