Multilingual SEO is sold as translation work, and that is the first thing our own audit disproved. Over two days we pointed an AI at the one site we could break without apologising to a client: ours. An English and German WordPress install of roughly 330 pages, ten years of accumulated content, and every shortcut a growing translation company takes along the way.
What follows are seven findings from that rebuild — including the ones that argue against letting AI run unsupervised.
The seven findings at a glance
- The biggest wins were technical, not linguistic
- Machine-generated metadata is a draft, not a deliverable
- Keyword work does not survive translation, and AI will hide that
- Structured data is where AI genuinely saves days
- Internal links are the cheapest ranking asset most sites ignore
- CMS and cache behaviour breaks more multilingual SEO than translation quality does
- Site performance is a translation problem too
1. The biggest wins were not linguistic at all
Before anyone looked at a single keyword, the audit surfaced problems no translator would ever see.
A sitemap returning 404 to crawlers. A retired language version still generating dead URLs. Animated statistics that rendered invisible to search engines because the numbers only appeared after JavaScript ran. Duplicate pages competing against each other in the same language.
Multilingual SEO gets scoped as a content job. A large share of the value sits in the plumbing.
2. Machine-generated metadata is a draft, not a deliverable
Generating titles and meta descriptions for dozens of pages across two languages takes minutes. Checking them does not.
Ours needed pruning for length, for claims we could no longer support, and for figures that contradicted each other between pages. Generation is the cheap part. Deciding what a company may claim about itself is the expensive part, and it cannot be delegated to a model.
3. Keyword work does not survive translation — and AI will happily hide that
We found one of our main English service pages carrying a focus keyword in Romanian. It had been sitting there for years.
An AI asked to “optimise” that page would have written beautifully around the wrong term. Multilingual SEO starts with what people actually search for in each market, not with the source market’s keywords rendered into another language. That is why our language pages are built per market rather than translated from one template.
4. Structured data is where AI genuinely saves days
Hand-writing schema markup is tedious and error-prone. Generating it and then validating it against live output took an afternoon: organisation details, address and contact data, and service markup across sixteen service and language pages, in both languages.
This is repetitive, well-specified work with a machine-checkable result. Automation is simply better at it than a person is.
5. Internal links are the cheapest ranking asset most sites ignore
Our highest-traffic pages were consumer blog posts that linked nowhere useful. The pages that earn money — technical translation among them — had almost no internal links pointing at them.
Adding one relevant, language-matched paragraph to the most-read posts took an hour. AI is good at proposing those links. It is not good at judging which one a reader would actually click, so every one was written and placed by hand.
6. CMS and cache behaviour breaks more multilingual SEO than translation quality does
Two of the hardest problems had nothing to do with language.
An abandoned caching plugin kept serving stale pages that showed none of the new metadata. And the site’s llms.txt file had generated itself from the German side of a WPML install, listing German URLs for an English-first business.
If your localisation vendor cannot see the content management system, they cannot deliver multilingual SEO. They can only deliver text.
7. Site performance is a translation problem too
Two megabytes of settings were being loaded from the database on every single page view, most of it one cache entry that should never have been there. Trimming it, together with proper page caching, cut that load by about 88%.
Nothing about this is linguistic. All of it affects how the site ranks, in every language it is published in.
Where AI helped, and where we stopped it
It helped with breadth: reading a whole site quickly, spotting patterns across hundreds of pages, generating and validating structured data, drafting metadata in two languages, and writing tedious redirect rules.
It was stopped, deliberately, at anything irreversible — deleting files, publishing claims, changing hosting settings, submitting forms. Every change went in with a way back, and every one was verified on the live site afterwards.
The failures were the instructive part. Edits were saved but invisible to visitors for an hour, because of a caching layer nobody had disabled. A settings change looked applied in the interface and had not been written. Anyone running AI-driven SEO at scale needs the discipline of a review desk: verify against the live artefact, not against the tool’s own report.
What this means for a multilingual SEO project
Multilingual SEO is not translated SEO. If you are scoping one, budget for three things the translation line item usually hides:
- Keyword research in each market, not keyword translation.
- Metadata written rather than converted, with someone accountable for the claims it makes.
- A technical audit, which in our experience finds the largest single wins.
Use AI for breadth and for repetitive, structured work. Keep a human on claims, on link relevance, and on anything that cannot be undone. And make sure whoever does the work can see both the source language and the CMS, because those two problems are rarely separable.
Frequently asked questions
Is multilingual SEO the same as translating my SEO content?
No. Search behaviour differs by market, so the terms people use in German are rarely the translated equivalent of the English ones. Translating optimised English copy produces fluent pages that target terms nobody searches for.
Can AI do multilingual SEO on its own?
It can do the breadth work — auditing, generating structured data, drafting metadata, proposing internal links. It cannot decide what your company is allowed to claim, which links a reader would follow, or whether a change is safe to make. In our own rebuild, every irreversible step was stopped and done by hand.
What usually gives the biggest improvement first?
In most multilingual sites we look at, the technical layer: crawlable sitemaps, dead language directories, duplicate pages and caching behaviour. These are unglamorous and they almost always outrank content tweaks in impact.
Working on a multilingual site?
We are a London-based translation company working mainly with language service providers on technical, engineering, life sciences, legal and marketing content. If you are planning a multilingual SEO project and want someone who can read both the source language and the CMS, get in touch or look at what we do.
A shorter version of this article first appeared on LinkedIn.
