This stopped being a binary question some time ago. Almost every site now machine translates most of its pages and has a person edit the ones that carry risk, which the trade calls post-editing, or mtpe. The useful question is not which is better but which pages get which treatment, because that split moves the cost of a translation project more than the per-word rate does.
What machine translation is now good enough for
Descriptive content where a clumsy sentence costs you nothing: blog archives, help articles, product descriptions in a catalogue nobody reads twice. Automated website translation serves these acceptably in the major European language pairs and increasingly beyond, and the alternative is usually that they are not translated at all.
What still needs a person
Anything that makes a promise or carries a risk: the pages that sell, the legal pages, the checkout flow, safety information, and anything where a mistranslation is a claim you did not intend to make. Also anything with brand voice in it, because machine output is fluent and characterless, and that is visible on a home page.
Where post-editing sits
Between the two, and it is what most agencies now quote for. A translator edits machine output rather than typing from scratch, which is faster and quoted lower per word. The saving is real but it is not the whole rate: budget post-editing as a discount on translation, not as machine translation with a rubber stamp.
Making the split a number
Decide the percentage of your words that will be machine translated and left alone, and the rest is your human word count at your quoted rate. The scope sheet on this site takes that share as an input for exactly this reason: it is the lever, and raising that share by a third changes the invoice more than any negotiation will.
The pages, sorted
Legal pages, the checkout, pricing, safety and compliance information, the home page and the top landing pages: a person, from scratch or by post-editing, every time the source changes. Product and feature pages, the about page, onboarding emails: post-editing, because a fluent mistake here costs a sale rather than a claim. Help articles, release notes, the blog archive, long-tail category pages: machine output, published as such, reviewed when a reader complains or a page starts to matter. Written as three lists against the site's own pages, that sort is the machine-translated share the scope sheet asks for, and it is a decision the company can defend rather than a percentage somebody guessed.
Questions people ask about human translation vs machine translation
Is machine translation good enough for a company website?
For most of it, with a person on the pages that carry risk. Publishing raw machine output on a checkout or a legal page is where sites get into trouble.
What is mtpe?
Machine translation post-editing: a translator edits machine output instead of translating from scratch. It is quoted lower per word than translation and is what most agencies now sell.
Does machine translation hurt search rankings?
Unreviewed, low-quality output at scale is the risk, not machine translation as such. The pages that matter are the ones worth a person's time anyway.
Should machine-translated pages be marked as such?
For readers, a short note on pages that are machine output and unreviewed is honest and cheap. For search, the tags do not change; what matters is that the unreviewed share is not the pages that carry a promise. The record keeps which pages are which, so the note can be true.