Machine Translation and AI Translation for Websites: When Translated Pages Help and When They Become Scaled Content

A small outdoor shop sells tents and rain jackets in English. Sales in France and Germany are growing, so the owner installs a translation plugin, connects it to an AI translation service and publishes French and German versions of all 1,800 pages in one afternoon. A few weeks later the new sections have barely any impressions. Some product pages say "fleece" where they mean a different fabric, the delivery page still promises next-day shipping in the UK, and every translated page tells search engines it's written in English as well as in French or German.

Nothing about this is unusual. Pushing a whole site through a translation API in an afternoon is low-key the most tempting growth shortcut around. Translation has never been cheaper, and modern models read far better than the tools of ten years ago. The question in 2026 isn't whether you're allowed to use machine translation. It's whether the translated pages are worth showing to someone searching in that language, and whether your setup lets search engines understand what you've built. This article covers both, and what DeepSpider.ai can and can't check for you.

In short
Google doesn't treat translated pages as spam because a machine helped. Its scaled content abuse policy targets many pages with little value, and it lists automated translation as one way such pages get made. Translate pages that are already useful, have a fluent speaker review them, add local facts, give each language its own URL with correct hreflang, and don't force visitors into a language. Keep unreviewed translations out of search until someone has checked them.

What Google says about translated pages

Older advice still circulates, so start with the current text. The spam policies (last updated 28 August 2026) define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", adding that it's "typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created." One of the listed examples is "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."

Translation is one of several ways to turn content into many pages, and the problem the policy names is the result, not the method. Google confirmed that reading in a 2025 statement to Glenn Gabe, who had asked about Reddit's millions of AI-translated pages: "our policies do not strictly define content that has been translated by AI as spam. Our scaled content abuse policy mentions automated transformations, including translations, as part of the overall warning against creating large amounts of unoriginal content that provides little to no value to users."

That's softer than Google's older guidelines, which listed "text translated by an automated tool without human review or curation before publishing" as auto-generated content (Search Engine Journal quoted that wording in 2022). The direction of travel is clear: Google cares whether people find the result useful, not which tool produced the first draft. Its spokespeople have said much the same over time:

• In 2018 John Mueller said translated content isn't automatically a problem: "it's more a matter of the intent and kind of the bigger picture what what they're doing. If they're essentially just spinning content and hoping that it ranks, then that that would be more of a problem for us."
• In 2019 he said he didn't think auto-translating would trigger manual actions, "but if the translations are bad, then it's bad content in general", and suggested an option worth remembering: "noindex until reviewed by local users / speakers."
• In a 2022 office-hours answer about machine translations that human editors only lightly corrected, Lizzi Sassman said it was "totally fine for Google as long as there's a human involved in the review process. That's the key."

Google's current guidance on generative AI content adds a detail that applies directly to translation. It calls it "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing", and says that review "also applies to metadata like" titles, meta descriptions, structured data and image alt text. A translated page with a reviewed body and a raw machine-translated title hasn't been fully reviewed.

What happened when translation went big

The clearest public example is Reddit. In March 2025 Glenn Gabe documented that machine-translated Reddit threads, served on URLs with a tl= parameter such as ?tl=fr, were surging in Google across several countries after the March 2025 core update. By checking the hreflang annotations he found some threads translated into as many as 22 languages. In a follow-up he reported about 2.3 million translated URLs ranking in France and 2.4 million in Spain.

For a while that looked like proof that mass AI translation works. Gabe kept warning that it didn't generalise: he wrote that companies had contacted him after receiving manual actions or being hit by spam updates for auto-translated content, and that "most sites cannot get away with mass-translating content via AI (or machine translation tools) without getting hammered by Google."

In July 2026 he reported that Reddit's AI translations had dropped heavily in Google after the May 2026 core update and the June 2026 spam update, including in AI Overviews and AI Mode. Their ChatGPT citations had dropped too, a little earlier. Nobody outside Google knows exactly which system caused it, and Reddit is not a typical site. But the arc is worth keeping in mind: a strategy that rides high for a year can still unwind, and visibility in AI answers tends to go with it.

Gabe's practical advice: judge the source content first, and be honest about it, because "you could be too close to your site and content." He suggests getting a group of unbiased people to review the content before translating it at scale, and warns against stacking one shortcut on another: scaling AI-generated content into other languages with AI translation is, in his view, asking for serious trouble with Google. If that sounds familiar, our guide to AI content and AI slop covers how to audit the source pages first.

When it's not a problem
A company that translates its 40 best support articles and product pages, has a fluent speaker check each one and keeps them up to date is not who the scaled content policy is about. Neither is a news site that translates its own reporting. The risk grows with volume, thin source pages and a lack of review.

What makes a translated page useful

Aleyda Solis put the minimum bluntly in 2023, commenting on a poll about AI translation: "The answer is no without validation of native translators." In a 2025 conversation with Gianluca Fiorelli, published by Advanced Web Ranking, Michael Bonfils separated translation (converting words) from localisation, which he described as being "about intent. It's about emotion. It's about formatting and adaptation." In practice there are three levels.

Raw machine translation

Plausible text with no human check. Fine for internal drafts and for understanding a document, risky as an indexable page, because nobody has confirmed that it's accurate, that product terms are right or that it makes sense in that market.

Machine translation plus post-editing

A fluent speaker, ideally someone who knows the subject, checks and corrects the output. This is the workflow Google's staff have described as acceptable, and it should cover more than the body text:

• Titles, meta descriptions, headings, alt text and structured data.
• Product names, sizes, units and technical terms, using a glossary so the same term is translated the same way every time.
• Links, so translated pages point to translated pages rather than back to the English section.
• Forms, error messages and buttons on the path to buying or contacting you.

Localisation

The page is adapted to the market, not just the language:

• Local prices, currency, taxes, delivery times, returns and payment methods.
• Local contact details, regulations, certifications and size charts.
• Examples, units and use cases that make sense there.
• Keyword research in the target language, because people don't search with translated English phrases. The French page for rain jackets should target what French shoppers type, not a literal rendering of "waterproof shell jacket".

Fiorelli goes further in a February 2026 article for Advanced Web Ranking. He argues that direct translations "tend to produce highly similar representations" in AI retrieval systems, so a translated page may be treated as interchangeable with the original, and recommends adding local entities, such as places, regulations or climate, to make each version distinct. That's his model of how these systems work, not something Google has confirmed, but the advice suits human readers anyway.

Glenn Gabe's December 2025 tests point the same way. Searching in French and Italian for content that existed in those languages with correct hreflang, he found Google and Bing returned the right language version, while ChatGPT, Perplexity and Claude often cited the English one. A page that only restates the English version gives those systems little reason to prefer it.

Where machine translation is fine on its own

Low-risk, low-traffic pages, such as an archive of old announcements, can be translated cheaply for visitors and kept out of search with noindex. Large user-generated sections are a judgement call. Google's multilingual guidance warns that translating only the "boilerplate" around content that stays in one language "can create a bad user experience if the same content appears multiple times in search results with various boilerplate languages."

How to set up translated pages so search engines understand them

A good translation can still fail if the technical setup confuses search engines. Most of these rules come from Google's multilingual documentation.

Give every language its own URL

Google recommends "using different URLs for each language version of a page rather than using cookies or browser settings to adjust the content language on the page." Subfolders (/fr/), subdomains (fr.example.com) and country domains all work. URL parameters are "not recommended".

One language per page

Google uses "the visible content of your page to determine its language" and doesn't use code-level information such as the lang attribute or the URL for that. It recommends "a single language for content and navigation on each page" and "avoiding side-by-side translations". A French page with an English sidebar, English reviews and an English footer sends mixed signals.

Hreflang that lists every version, including the page itself

Google's rules: "Each language version must list itself as well as all other language versions", URLs must be fully qualified, and "If two pages don't both point to each other, the tags will be ignored." You can use HTML link elements, HTTP headers or the XML sitemap. Google says the three are equivalent and there's no benefit in using more than one.

Joost de Valk's hreflang guide on Yoast adds a rule that translation plugins sometimes break: hreflang and canonical tags "can and should be used together", and every language version should have a canonical pointing to itself. A French page whose canonical points to the English original tells Google the English URL is the one to index, which breaks the hreflang setup.

Hreflang doesn't make a page rank. It helps Google show the right version to the right searcher once the versions are worth showing.

Set the html lang attribute correctly

Google ignores <html lang> for language detection. Bing doesn't. Fabrice Canel of Microsoft Bing said in 2020 that "hreflang is indeed a far weaker signal than content-language at Bing." Set lang (and, if you use it, the Content-Language header) to the actual language of each page. It's also an accessibility basic: screen readers use it to choose pronunciation.

Let people switch, and don't force them

Google is direct about redirects: "Avoid automatically redirecting users from one language version of a site to a different language version of a site." It also says "Don't use IP analysis to adapt your content", and warns that Googlebot usually crawls from the USA and sends requests "without setting Accept-Language in the request header". A site that sends every US visitor to English, or picks the language from the browser and hides the rest, may never have its other versions crawled.

Add plain, crawlable links to the other language versions of each page instead. A banner that suggests another version is fine. A redirect that takes away the choice isn't.

Check what Google translates on its own

Google may translate a result's title and snippet itself when people search in a language you don't cover. Its documentation lists 21 languages, including Spanish, French, German, Hindi and Turkish, and says you can see these clicks with the Search Appearance filter in the Search Console Performance report. Clicking such a result opens a machine-translated view of your page, much like opening it in Google Translate.

Erik Sarissky of Ahrefs wrote in June 2025 that these proxied visits can show up as translate.google.com referrals rather than organic search, so your reporting undercounts them. His suggested fix is to "Ship even a 300-word native-language version" of pages that get this traffic. If you'd rather opt out entirely, Google supports a notranslate rule as a meta tag or HTTP header. Treat this traffic as market research: it shows where people already want your content in another language.

Signs a translated section is underperforming

Give a new language section several weeks before judging it; a site new to a market will grow slowly whatever the quality. These signs are worth investigating:

• Many translated URLs not indexed. Lots of "Crawled, currently not indexed" or "Discovered, currently not indexed" pages in one language folder of Search Console's Page indexing report suggest Google doesn't see enough value there yet.
• Impressions in the wrong market. If the English pages still collect most impressions from France while the French pages get few, Google may not be connecting the versions, or may not prefer the French pages for French searchers.
• Translated pages showing as duplicates. If Google picks the English page as canonical for a French URL, the main content may still be in English, or the canonical tags point the wrong way.
• Poor engagement compared with the original. Short visits and few conversions, against the same pages in the source language, often point to wording, prices or delivery details that don't fit the market.
• Traffic arriving through Google Translate. There's demand in a language you haven't covered properly yet.
• A drop after a core or spam update that hits one language folder much harder than the rest. That's the pattern to take seriously.

If a large translated section was a quick win at launch and is now in decline, it's giving the Reddit trajectory, on a smaller scale.

What to do when it isn't working

Start with the pages that matter, not with the whole section.

• Separate setup problems from quality problems. Broken hreflang, wrong canonicals and forced redirects can hide good translations. Fix those first; they're cheap and measurable.
• Group translated URLs by type. Product pages, category pages, guides, support articles and archive pages need different decisions.
• Read samples like a local visitor. Better still, ask someone who is, and treat it as a proper vibe check by someone who lives there. Check facts, prices, delivery, terms and metadata.
• Decide per group. Improve and localise pages that have demand. Keep low-value translations for visitors but noindex them until reviewed. Remove pages nobody in that market needs, redirecting to the closest useful version if one exists.
• Fix the source pages if they're the problem. If the English pages are thin, better translation won't help. Our guide to deciding whether to update, merge or remove pages works the same way in every language.
• Measure in months. Gabe notes that recovery from a spam update can take months, and isn't guaranteed.

When you use noindex as a holding pen, also remove those pages from the other versions' hreflang until they're reviewed. Otherwise one tag says "show this version to French searchers" and the other says "don't show it at all".

How DeepSpider.ai helps

DeepSpider can't tell you whether a translation is good. It doesn't compare languages, check terminology or judge whether a page reads naturally. It checks the setup that translation plugins often get wrong, and points you at pages that need a person to look at them.

These hreflang and language checks are in the Technical category:

• Hreflang implementation issues (Warning): invalid language or region codes, relative alternate URLs, missing self-references, missing return links, and alternates that redirect, fail, are set to noindex or are canonicalised elsewhere.
• Hreflang URL declared as multiple languages (Warning): one URL assigned to two or more languages by its own annotations, other audited pages, HTTP headers or the XML sitemap. This is the plugin bug in the opening example, and across a whole language folder it usually comes down to one template or plugin setting.
• Incomplete hreflang language set (Warning): versions that are only connected through another page, for example a German page that lists the English page but not the French one.
• Hreflang language differs from page language (Warning): the page's own hreflang says one language while <html lang> says another, which usually means the theme sets the default language on every translation.
• Canonical or hreflang outside `<head>` (Warning): tags printed in the body, where Google ignores them.
• Noindex page with self-referencing hreflang (Needs review): a noindexed page that still lists itself as a language version. Useful when you hold unreviewed translations back with noindex.
• Language-dependent content accessibility (Needs review): compares responses with and without a language preference and reports redirects, noindex or canonical differences, or a large loss of text. It runs on the homepage plus a sample of up to 100 other pages, doesn't test IP-based redirects, and is skipped on sites served through Cloudflare, where it shows as not checked.
• Missing or invalid page language (Needs review, in the Accessibility topic): pages without a valid lang attribute.

Each check comes with How to fix it guidance. The Warnings count towards Website Health Score. The Needs review checks don't.

Duplicate page content finds pages with identical main text, so an English page published again under /fr/ without translation shows up next to the original. Near-duplicate main content only compares pages that declare the same language, so a lightly edited English copy marked lang="fr" slips past it. A partly translated page won't show up in either, and DeepSpider has no dedicated check for untranslated language versions. AI Content Detection only looks for repeated paragraphs and a few English assistant phrases, and doesn't estimate whether a machine wrote or translated a page.

After a plugin update or a migration, Compare Audits → Important changes shows URLs whose full hreflang set changed between two saved audits. With Google Search Console connected, Audited pages Google has not indexed groups inspected pages by Google's reason and Indexable pages without Google impressions lists audited pages with no impressions, which helps when you're reviewing a language folder.

DeepSpider reads the source HTML without running JavaScript, so a language switcher or hreflang added only by JavaScript isn't seen.

Frequently asked questions

Is machine-translated content against Google's guidelines?

No, not by itself. Google told Glenn Gabe in 2025 that its policies "do not strictly define content that has been translated by AI as spam". The risk is many translated pages with little value.

Do I need a human to review every translated page?

For pages you want indexed, yes, that's the safe standard, and Google staff have described human review as "the key". For low-value pages you can translate for visitors and keep them noindexed until someone has checked them.

Will translated pages be treated as duplicate content?

Not if the main content is translated. Google says localised versions are "only considered duplicates if the main content of the page remains untranslated." Hreflang tells Google they're versions of each other.

Should I redirect visitors to their language automatically?

No. Google recommends against automatic redirects by language or IP address. Suggest the right version, and let people choose.

Is AI translation better than older machine translation for SEO?

The output is usually more fluent, which makes errors harder to spot, not less important. Fluent text with wrong prices or terms is still a poor page.

The short version

Machine translation is a fine starting point and a poor finishing point. Translate pages that are already useful, have a fluent speaker review them, including the metadata, and localise the facts that differ by market. Give each language its own URL, one language per page, hreflang in both directions, a self-referencing canonical and a correct lang attribute, and don't force anyone into a language. Keep unreviewed translations noindexed, watch indexing and impressions per language folder, and fix the setup before you judge the content.

Sources