Guide
How to run a multilingual blog with AI without getting it deindexed
Google does not penalise AI writing. It penalises writing produced at scale that helps nobody, and it applies that judgement across the whole site rather than page by page. The difference between the two is entirely in what you put into the post before a word is written.
Short answer
Write from real material: your own site, your own customers, named sources and real figures. Give each post citations a reader can follow and at least one thing only you could supply. For other languages, translate the whole page rather than the navigation, self-canonical every locale, and make hreflang reciprocal with an x-default. Publish fewer locales than you can genuinely fill.
What Google actually objects to
The confusion here costs people a lot of work. Google stated plainly in 2023 that it rewards helpful content however it is produced, and reiterated it since: automation is not the problem.
The policy that does bite is scaled content abuse: producing many pages whose main purpose is to rank rather than to help. The March 2024 update rewrote that policy so it applies regardless of how the content was made, closing the gap where "a human lightly edited it" was a defence. In 2025 and 2026 the enforcement has been about volume and value, not authorship.
Two consequences worth internalising:
- It is assessed site-wide. The helpful content guidance is explicit that unhelpful material anywhere on a site can affect how the whole site performs. A hundred filler posts do not sit quietly beside your ten good ones; they pull them down.
- Nobody is checking whether a model wrote it. They are checking whether the page contains anything that was not already on the first page of results.
The input decides the output
A post written from nothing but a keyword can only paraphrase what already ranks. A post written from material says something. The material is the work.
| Input | What it gives the post |
|---|---|
| Your own site, read properly | What you actually sell, to whom, in your own vocabulary |
| Real customer questions | The sub-questions nobody else answers, in the words people use |
| Named sources with figures | Claims a reader can check, and something to cite |
| Your own photographs | Images nobody else has, at print resolution rather than model resolution |
| A video you or somebody else made | A reason to stay on the page, and a second schema type |
| Your own numbers | The only thing on the page that cannot be copied from anywhere |
If a post has none of these, it should not be published. That test alone removes most of what gets people into trouble.
Citations, and how to place them
Cite the source next to the claim, not in a list at the bottom that nobody reads. Use the page title as the anchor text rather than the bare domain, so the link says what the reader will find. Prefer primary sources: a regulator, a standards body, the company that published the figure, the study itself rather than the article about the study.
Two or three strong citations beat a dozen weak ones. A link to a competitor blog post that itself cites nothing adds no credibility to yours.
Images, video and structured data
Every post wants at least one image that is genuinely yours. A customer photograph at 4000 pixels across is worth more, on the page and in print, than anything a model produces at 832 by 1216, which is under three inches wide at print resolution.
Mark the post up as Article or BlogPosting with a real author, a published date and a modified date that you actually update. If the post embeds a video, add VideoObject with its duration and thumbnail: it is one of the few structured data types that still changes how a result looks.
Doing other languages properly
This is where multilingual blogs fail, and the failures are mechanical rather than editorial.
- Translate the whole page. Title, description, headings, body, alt text, captions. Translating only the navigation while the article stays in English produces a page Google reads as an English duplicate.
- Self-canonical every locale. The French page canonicals to the French page. A cross-locale canonical removes that locale from the index, and it overrides hreflang when the two disagree.
- Make hreflang reciprocal, and include self. Every page lists every alternate including itself. If A points to B, B must point back to A, or the pair is discarded.
- Use correct codes. ISO 639-1 language, optional ISO 3166-1 Alpha 2 region.
en-GB, neveren-UK. - Declare an x-default pointing at the language selector or your primary locale.
- Keep one signal. If you annotate in both the HTML head and the sitemap, they must agree exactly; conflicting pairs are dropped.
Google own documentation on localised versions is the authority here and worth reading in full before you open a second language.
Then the editorial rule, which matters more than any of the above: open fewer locales than you are able to. Five languages you genuinely serve will beat twenty you have machine-translated and never read. Thin locale pages are counted in the same site-wide assessment as thin English ones.
A publishable post, as a checklist
- One question a real person asks, answered in the first hundred words.
- At least one fact that came from you and exists nowhere else.
- Two or more citations to primary sources, placed at the claim.
- An image you own, at a size that would survive printing.
- Structured data that matches what is on the page.
- An internal link to the thing you actually sell, where it is genuinely relevant.
- A named author and a date you will keep current.
Seven things. A post that has them is worth publishing in every language you can support. A post missing four of them is what the spam policy was written for.
Neubook blog studio reads your site, pulls out what your business does and who it serves, takes your own notes, media and links as material, and writes posts with the citations, schema and translations already in place.
Open the blog studioSources
- Google: Search and AI content
- Google Search spam policies, including scaled content abuse
- Google: creating helpful, reliable, people-first content
- Google: localised versions of your pages, and hreflang
- Google: Article structured data
- Google: Video structured data
Common questions
Does Google penalise AI-written content?
Not for being AI-written. Google has said since 2023 that it rewards helpful content however it is produced. What it acts on is scaled content abuse: generating many pages primarily to rank, with no original value. That policy makes no distinction between automated, human and hybrid production, so a thin page written by a person is treated the same as a thin page written by a model.
Is AI translation safe for SEO?
Yes in itself, and it is how most large multilingual sites operate. The risk is volume without value: hundreds of machine-translated pages nobody has checked, for languages where nobody is searching. Translate fully, check the ones that matter, and only open locales you can actually serve.
What is the most common hreflang mistake?
Leaving out the self-referencing entry. Every page in a hreflang set must list itself alongside its alternates; if it does not, the entire set is ignored. The next most common is a missing return link: if the English page points to the French one but the French page does not point back, Google drops the pair.
Should each language version canonical to the English one?
No. That is the mistake that silently removes every other locale from the index. Each locale page must canonical to itself. A cross-locale canonical tells Google the other language is a duplicate to be discarded, and it will be, no matter what your hreflang says.
How many posts should a blog publish?
As many as you can make genuinely useful and no more. The helpful content assessment is site-wide, so a hundred thin posts will drag down the ten good ones. Ten strong posts with real sources outrank a hundred paraphrases of the same search results.