The worst outcome of a website translation is not a bad translation. It is no translation at all, with the site giving no sign of it. The English page opens, the layout is right, the URL carries /en/, and the content is in Persian. To an English-speaking visitor the site is broken. To you it looks perfectly fine.
This site serves four languages and we had exactly that problem. What follows comes from that, not from a generic guide.

One URL or four
We keep one post with one slug, and the same address is served under a language prefix: Persian with no prefix, the others under /en/, /ar/ and /tr/. The translations live on that same post rather than in separate posts.
The benefit is that versions never drift apart. Change a figure in the article and all four sit in one place, so you cannot forget to update one. The cost is that editing is more complicated and no off-the-shelf plugin will do it for you.
Three things to look for in the page source
- An
htmltag with correctlanganddir. On this site the Persian version islang="fa-IR" dir="rtl"and the Turkish one islang="tr" dir="ltr". hreflanglinks for all four languages plusx-default. Our pages carry five.- A canonical pointing at that same language, not at the Persian version. The English page should have an
/en/canonical. - A language switcher that links to the same page in the other language, not to that language's home page. Dropping people on the front door is the commonest lazy implementation.
The third breaks more often than the others, and the result is that Google treats your other languages as duplicates of the Persian. The rules are written down by Google.
The worst bug we hit
The page cache was writing the output of all four languages into a single slot. If a Turkish visitor arrived first, the Turkish version was stored as the Persian one, and the next person opening the Persian page got Turkish.
The cause was simple and it took a day to find. The language prefix is stripped from the request URI before routing, so that every existing WordPress rule keeps working untouched, and the cache layer read the URI after that strip. As far as the cache was concerned, all four languages were the same address.
If you run a multilingual site with a page cache, test this today: open one page in two different languages in a private window and check that the second one comes back in the right language. This failure does not announce itself.
A rule specific to Persian and Arabic
Arabic and Persian share a script and their rules run in opposite directions. Arabic yeh, Arabic kaf, teh marbuta and the hamza forms are correct letters in Arabic and wrong in Persian. Run a Persian text normaliser over Arabic and it converts every Arabic yeh to a Persian one and every teh marbuta to a heh, producing text that is neither language.
We have a number for this. Running the Persian rule set over one Arabic article reported 928 false failures. Had the auto-fix run, the article would have been corrupted and nobody would have noticed until an Arabic reader said so.
So the checking tool has to take its rules from the language, not from the site. It is the smallest thing to get wrong on a multilingual project and it does the most damage.
Where machine translation is enough and where it is not
Our position: for a page meant to sell or to rank, it is not enough. Not because words come out wrong, but because the argument and the rhythm of the sentences arrive from the source language, and a native reader senses it even when they cannot name it.
For internal panel screens and error messages? Machine translation with one human pass over it is usually fine. You can see the same language and place effects at work in the piece on how location shifts rankings.
Do this arithmetic before commissioning anything
Every new language means that many pages again, and more importantly that many pages to keep current from now on. Three extra languages is four times the maintenance, not four times the traffic.
If you are unsure the market exists, start with one extra language and watch it for six months. Adding a fourth language to a site with three live ones is easy. Keeping four dead ones alive is not.
If you decide to go ahead, translation is work we do, and its technical structure belongs inside building the site rather than being bolted on later.
Questions we get
Subdomain or language folder?
Both work and Google prefers neither. A language folder is simpler to manage and does not split domain authority, and it is what we chose.
What happens if a page exists in only one language?
On this site it falls back to Persian and renders perfectly, which is the most dangerous outcome because it looks healthy. The only way to catch it is a tool that checks completeness per language.
Do I need separate keyword research per language?
Ideally yes. What a Turkish buyer types is not a word-for-word rendering of the Persian phrase. We have not done that work for all three of our own languages yet, and we are saying so here.
Does translating a site affect speed?
The translation itself does not, but caching it does. Each language is a separate entry in the cache, so memory use and warm-up time multiply.
Comments & Questions
Have a question about this article? Ask, we'll answer.
No comments yet; be the first.