Gherkin/Cucumber's big thing is that it lets the people who can't read code (i.e. project managers, suits, and customers) understand what's being tested. Usually along the lines of:

"As a user, I want to buy things, so I should be able to put things in the cart and get an accurate total price"

or

"As a user, I want to view the site in my own language, given I have come from a German network address, I should get the German language site"

This frees you from having to explain the finer points of cookies, or geoip, or any number of concepts to people who do not operate on a level that close to the bare metal.

The translation isn't really error prone either and isn't hard to implement from the programmer's side. Say that language example above? The actual test would be written like:

    As a user,
    I want to vew the site in my own language
    Given I have come from a German network address,
    I should get the german language site.

Those last two things are recognized via a verbatim regex (basically /^I have come from a German network address$/),

and the actual implementation code does something based on that, say, setting a test IP to a known German netblock, and the last line would actually load the site and look for the right localized strings.

So they're rather closely coupled together. If you change the wording of the test for whatever reason, the old code won't work anymore (actually, I think Cucumber will tell you that there was no test found), necessitating someone to actually verify that the test as written actually tests what the new requirements are.