Discussion starter: an international SEO question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- International SEO
Machine translation with post-editing versus full human localisation: have you seen a ranking difference, or only a conversion difference?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
No controlled public test exists that I can point to which isolates translation method as a ranking variable, so confident claims in either direction are anecdotes. The mechanism suggests translation method is rarely what moves rankings on its own. Rankings move when the localised page matches how people in that market search, and when it clears the quality bar that stops Google treating it as low-value content produced at scale. Well post-edited machine translation built on local keyword research can plausibly rank as well as human translation, while a careful human translation of an unlocalised page can underperform both.
Conversion is different, and the mechanism there is direct. Awkward phrasing, the wrong level of formality, untranslated legal terms and foreign-feeling prices erode trust at the moment someone decides whether to buy. So the more useful question is which pages need human localisation for conversion value, and which need local search research regardless of who or what does the translating.
The comparison bundles three separate jobs
“Machine translation with post-editing (MTPE) versus full human localisation” compares two packages, not two variables. Each package includes a different amount of three jobs:
Linguistic accuracy
Correct, fluent text with consistent terminology. Competent MTPE with a maintained glossary covers most of this. Mainly affects trust and conversion.
Search-language fit
Using the terms and intent local searchers actually use, researched per market. Neither machine nor human translation does this by default. Mainly affects rankings.
Market fit
Local currency, pricing, shipping, payment methods, contact details, reviews, examples and regulation. Needs business decisions, not translators. Affects both.
Most reported ranking differences between the two approaches are probably differences in the second and third jobs. An agency that does local keyword research as part of “full localisation” will often beat plain MTPE on rankings, but the cause is the keyword research, not the absence of a machine. That is why the test design further down separates them.
What Google’s documentation actually says about translated content
Google’s spam policies do not ban machine translation. The relevant text sits under scaled content abuse, which includes “Scraping feeds, search results, or other content to generate many pages (including through automated transformations like synonymizing, translating, or other obfuscation techniques), where little value is provided to users”. The operative condition is many pages with little value to users, not the tool. Google’s broader guidance on AI-generated content takes the same position: it evaluates the quality of the result rather than how it was produced.
Google’s international documentation adds two relevant points. It warns that translating only boilerplate while leaving the main content in one language “can create a bad user experience”, and it lists signals that identify a site’s intended audience, including “local addresses and phone numbers on the pages, the use of local language and currency, links from other local sites”. Its localised versions guidance states that “Localized versions of a page are only considered duplicates if the main content of the page remains untranslated”, so a properly translated page is not a duplicate-content problem.
The practical complication is that raw or lightly edited machine translation at scale tends to fail in ways that look like quality problems: literal keyword choices, mixed-language templates, untranslated titles and alt text, identical boilerplate across markets. A pattern practitioners describe often is large sets of machine-translated pages sitting in “Crawled - currently not indexed” or indexing slowly, with performance improving once the key templates are edited and localised. That observation does not contradict the guidance. The pages were thin or mismatched, and unreviewed translation was simply the route to that outcome.
Market fit also sets a ceiling that translation cannot lift. Aleyda Solis’s analysis of the May 2026 core update found visibility shifts driven by intent, market fit and source type, with UK results moving from US.com retailers towards local-market domains. That is a large-scale observation rather than a controlled test, and it says nothing about translation quality directly. It is a reminder that a linguistically flawless page from a site that is obviously not built for the market still competes at a disadvantage.
Where a ranking difference plausibly comes from
Mechanism How MT or light MTPE fails How to check Literal keyword choice Translates the source term rather than using the local one: “used car” becomes “voiture utilisée” where French searchers say “voiture d’occasion”; “mobile phone” becomes “Mobiltelefon” where Germans commonly say “Handy” Compare titles and H1s with local keyword research and with the wording on ranking local competitors Intent mismatch The translated head term triggers a different intent in that market (informational rather than transactional, or comparison rather than product) Manual SERP review per market for the head terms Half-translated pages Titles, meta descriptions, alt text, slugs, structured data or UI strings left in the source language Crawl each locale and flag source-language strings Scale without value Thousands of pages translated because they exist in the source language, not because the market needs them Page indexing report filtered by locale folder or sitemap; share of “Crawled - currently not indexed” One version for several dialects A single Spanish version serving Spain and Latin America, where vocabulary differs (“móvil” versus “celular”) Search Console query and CTR data split by country Missing market signals Foreign currency, no local contact details, reviews from another country Template audit per market Rule out the technical causes first:
Before blaming translation quality for weak rankings, check that each locale’s pages carry a self-referencing canonical rather than pointing at the source-language version, that hreflang annotations are reciprocal, and that no staging-era noindex or robots.txt rule survived the launch. These mistakes are common on machine-translation-driven launches and suppress rankings regardless of translation quality.
Why the conversion difference is easier to see
Conversion effects have shorter causal chains. A visitor who meets an informal “du” on a German checkout page that should say “Sie”, a returns policy that reads like a legal translation exercise, or a product specification with an obvious error does not need an algorithm to react. Trust drops on the page itself. Conversion is also easier to measure, because you can serve two versions of a page to the same traffic and compare, which is not possible for rankings. That measurability is probably part of why conversion differences get reported more often than ranking differences: they are simply easier to attribute.
Matching translation effort to page value
Page type Suggested approach Reason Home, key category and landing pages Full localisation with local keyword research and native review Highest traffic and conversion value; head terms need market-specific wording Product and listing templates, top sellers MTPE with a glossary, plus human localisation of template strings, titles and top items Template errors multiply across thousands of pages Long-tail catalogue MTPE with a glossary and sampled quality checks Cost scales with volume; each page matters less Checkout, pricing, returns, legal Human translation plus local legal or compliance review Conversion and liability Help centre and documentation MTPE with terminology management and fast update cycles Freshness and consistency matter more than style Blog and thought leadership Translate only topics with local demand; adapt rather than translate Translating everything creates scale without value How to test it on your own site
A two-arm test (MTPE versus full localisation) will tell you whether the packages differ, not why. A three-arm test separates the variables:
- Arm A: MTPE as currently produced.
- Arm B: MTPE plus local keyword research applied to titles, H1s, headings and key terms.
- Arm C: full human localisation, including the same keyword research.
Hypothetical example:
A retailer with a machine-translated German site takes 450 comparable category and product-group pages, matches them on pre-period impressions and clicks, and randomly assigns 150 to each arm. The primary metric is non-brand clicks per page from Germany, using Search Console’s country and page filters. Secondary metrics are the number of distinct queries per page (coverage), CTR, conversion rate and revenue per session. The test runs for at least 8 to 12 weeks after the changed pages are recrawled, avoids launching into a core update rollout, and annotates one if it lands mid-test.
How to read the result:
- B matches C on rankings, C beats B on conversion: the search value sits in keyword research. Pay for human localisation where conversion value justifies it, and roll keyword research into the MTPE workflow everywhere else.
- C beats B on rankings as well: something beyond vocabulary (intent handling, content quality, tone) is affecting performance in that market. Extend human localisation to more templates.
- A, B and C perform about the same: rankings in that market are being decided by other factors, such as links, brand strength or market fit. Spend there rather than on translation.
Confounders to control: internal linking changes that ride along with the rewrite, seasonality that differs by market, hreflang errors, and page age. If the site is too small for three arms of meaningful size, run B versus C on the highest-value templates only and accept a directional answer.
If this were my budget, I would not frame the decision as machine versus human. I would pay for local search research on every market’s important templates, use human localisation where conversion value or legal exposure justifies it, run MTPE with a glossary on the long tail, and refuse to publish pages in a market just because they exist in the source language. Then I would run the three-arm test on one locale, because it separates the question asked (does translation method affect rankings?) from the one that usually decides the outcome (did anyone research how that market searches?).
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.