Google's own AI judged us against Google Translate. Google lost 108 to 1.
118 blind rounds across 11 country variants of Spanish, Portuguese, French, Arabic and Chinese, refereed by Google's Gemini 4 Argon. Every sentence, translation and verdict is published.
November 1, 2026AIRSkip Inc.Get the data
108
won by AIRTranslate
9
ties
1
won by Google
92% of 118 blind roundsJudged by Gemini 4 ArgonEvery round read in both orders
On the map, Spanish is one language. On the street, it is at least twenty. A Mexican moviegoer buys palomitas, an Argentine buys pochoclos, a Colombian buys crispetas, and all three are buying popcorn. Most translation apps hand every one of them the same sentence.
AIRTranslate was built on the opposite idea: a translation should sound like the place you are standing in. So we measured that idea, with a referee who has every reason to favor the other side.
01The result
108 rounds to AIRTranslate. 1 to Google.
Each square below is one round: one sentence, asked for one country, answered by both systems and judged blind by Gemini 4 Argon. AIRTranslate was preferred in 108 of 118. Google Translate was preferred once. The other 9 were ties.
Tap any square to read that round
- AIRTranslate preferred
- Tie
- Google Translate preferred
Spanish (Mexico)
12/12
Spanish (Argentina)
12/12
Spanish (Colombia)
11/12
Spanish (Spain)
12/12
Portuguese (Brazil)
8/10
Portuguese (Portugal)
7/10
French (France)
9/10
French (Quebec)
8/10
Arabic (Egypt)
10/10
Arabic (Saudi Arabia)
10/10
Chinese (Taiwan)
9/10
Spanish (Argentina) · sentence 1 of 12
“Let's get popcorn and a soda before the movie starts.”
Google Translate
Vamos a comprar palomitas y un refresco antes de que empiece la película.
AIRTranslateJudge's pick
Vamos a comprar pochoclos y una gaseosa antes de que empiece la peli.
JudgeGemini 4 Argon
AIRTranslate uses the specific Rioplatense vocabulary ('pochoclos', 'gaseosa', and the informal 'peli') that is standard in casual Argentine speech, whereas Google Translate uses terms from other regions.
Round 13 of 118
The judge saw the two translations as A and B. The names in its reasons are filled in by us.
The pattern inside the grid is what makes it credible. Five of the nine ties are Portuguese, where Google has a dedicated version for both Brazil and Portugal, and four are sentences where the two systems wrote exactly the same words. The sweeps happen where one Google answer has to serve a whole continent.
Google offers exactly one Spanish and one Arabic, for every country that speaks them.
Across our four Spanish and two Arabic variants, AIRTranslate was preferred 67 times out of 68, with one tie.
02The example
One sentence, four variants
The same English, asked for four countries. Google Translate returns one sentence for all of them. AIRTranslate changes exactly the words a local would change.
The English
“Let's get popcorn and a soda before the movie starts.”
Google Translate
Vamos a comprar palomitas y un refresco antes de que empiece la película.
The same sentence, every country
AIRTranslate
Mexico
Vamos por unas palomitas y un refresco antes de que empiece la película.
Argentina
Vamos a comprar pochoclos y una gaseosa antes de que empiece la peli.
Colombia
Vamos por crispetas y una gaseosa antes de que empiece la película.
Spain
Vamos a por palomitas y un refresco antes de que empiece la peli.
03Beyond vocabulary
It runs deeper than nouns
Verbs change shape in Argentina. Shopping lists change words in Cairo. Even the subway has a different name in Taipei. The words in orange give Google's version away; the words in green are what a local says.
Spanish (Argentina)
“Hurry up, we're going to be late!”
- ¡Date prisa, vamos a llegar tarde!
- AIRTranslate
- ¡Dale, apurate, que vamos a llegar tarde!
Argentina speaks in voseo: apurate, never the Spain-and-textbook date prisa.
Arabic (Egypt)
“Give me a kilo of tomatoes and some bread.”
- أعطني كيلو من الطماطم وبعض الخبز.
- AIRTranslate
- هاتلي كيلو طماطم وشوية عيش.
In a Cairo market bread is عيش and you ask with هاتلي. Google writes the formal Arabic of a textbook.
Chinese (Taiwan)
“I'll take a taxi to the subway station.”
- 我將搭乘計程車去地鐵站。
- AIRTranslate
- 我會搭計程車去捷運站。
Taipei rides the 捷運. Google prints the mainland 地鐵 even when Taiwan is requested by name.
French (Quebec)
“We went shopping on Saturday morning.”
- On est allé faire l'épicerie samedi matin.
- AIRTranslate
- On a magasiné samedi matin.
A variant Google officially supports, and it still says faire l'épicerie, grocery shopping. Quebec says magasiner.
French (Quebec)
Google won this round“We rented a cute little cottage by the lake.”
- On a loué un beau petit chalet au bord du lac.
- AIRTranslate
- On a loué un petit chalet ben cute au bord du lac.
The judge found our ben cute too slangy for writing and preferred Google's plain beau petit chalet. A fair call, and we publish it with the rest.
04Availability
Most apps cannot even be asked
Before quality comes the question of whether you can ask at all. Each dot below is one country version of a language that a translator offers. Apple Translate has no country versions of Spanish, Arabic, French or Portuguese. It cannot be asked for Mexican Spanish, so there is no quality test to run: the request does not exist there.
474
AIRTranslate
country versions
Spanish for 23 countries, Arabic for 23, French for 32, English for 34, and more.
8
DeepL
country versions
US and UK English, Brazil and Portugal Portuguese, two Chinese scripts, Quebec French and one Latin American Spanish. One Arabic.
4
Google Translate
country versions
Quebec French, Portugal Portuguese and two Chinese scripts. One Spanish, one Arabic, for every country.
4
Apple Translate
country versions
US and UK English and two Chinese scripts. No country versions of Spanish, Arabic, French or Portuguese.
05Method
How the test works
- 1118
52 sentences, 11 countries
Everyday English, written so that a natural rendering differs by country in vocabulary, idiom, register or script. Sibling variants share the same sentences: 118 rounds in all.
- 2Oct 17
Both systems as they ship, same day
AIRTranslate and Google Translate as anyone can use them, on the same day. Where Google has no version for a country, it got its closest one, which is what a Google user asking for that country gets.
- 3A / B
A blind referee from Google
Gemini 4 Argon, at its default settings, saw the two answers as A and B and was asked which a native of that country would find natural. If it leans toward anyone, it is not us.
- 4113/118
Twice, with the order swapped
A win counts only when both orders agree; anything else is a tie. 113 of 118 verdicts were the same both ways.
06The data
Check every call yourself
Every sentence, every translation from both systems and every verdict with its reason, so anyone can re-run the study or dispute a call.