While rewriting our pages this summer we pulled 37,621 real queries out of Google's autocomplete across the seventeen languages our site speaks — seeded with what people might want, expanded wave by wave into what they actually type. We were looking for keywords. What we found was a portrait of the category, and the sharpest split in it sits exactly on «ai boyfriend vs ai girlfriend».
That phrase itself is typed. The assumption behind it — that these are two flavours of one product — is what the data quietly took apart.
How we counted, and what the numbers can claim
Autocomplete is built from real query volume, so a phrase that appears is a phrase people type; how often, it does not say. Our prominence signal is convergence: how many distinct phrases lead to the same suggestion. A phrase that fifteen different seeds all reach is central to a market; one that surfaces once is a leaf.
Everything below is that kind of evidence — directional, honest about its limits, and still far better than the language intuition we used before it.
India: two different products wearing one name
The starkest result in all seventeen corpora. Around «ai girlfriend», Indian queries are one third games and APK downloads and nearly another third photo editing — generating a girlfriend picture to post — with barely one query in twenty about conversation. Around «ai boyfriend», conversation is the single biggest slice, led by the plainly human phrase «ai boyfriend to talk».
Same country, same category label, opposite intents: the girlfriend is searched as an image, the boyfriend as a listener.
Iran: the buyers are women
Persian was the one market whose cleanest commercial demand — download, app, Telegram, free, the whole purchase-shaped tail — clustered around the AI boyfriend, not the girlfriend. The girlfriend term there is tangled in dating-sim games and forum threads about actual long-distance relationships; the boyfriend term is a product being shopped for.
We rewrote our Persian pages accordingly, and it is worth saying out loud: in at least one market of seventeen, the default assumption about who uses AI companions is simply backwards.
The languages that refuse the question
Turkish sevgili, Vietnamese người yêu, Kazakh сүйікті — beloved, partner, dear one — are genuinely genderless, so their markets never split the category in two. One word carries both searches, and the boyfriend-or-girlfriend framing arrives only with English.
Vietnamese adds a joke at the category's expense: «ai» is the Vietnamese word for «who», so the literal string «bạn gái ai» reads as «whose girlfriend» — and a third of the queries around it turn out to be celebrity gossip about footballers' partners. Language shapes what a market can even ask for.
What nobody searched for
The absences taught us as much as the volumes. Persistent memory — the feature we consider our strongest — has almost no direct search demand in eleven of seventeen languages; people meet it as a complaint («it forgets everything»), not as a request. And in Japan the queries around AI companions lead with words for danger and risk: that market is not comparing products, it is deciding whether the category is trustworthy at all.
!A feature list answers questions people are not asking. The searches say: tell me what this is, tell me if it is safe, speak my language, and let me make my own — in roughly that order.
What did we change because of this? Our seventeen title tags now each lead with what that market actually types — and the product answer to «boyfriend vs girlfriend» turned out to be the honest one all along: here, gender is one word in a sentence you write. Describe who you want to meet. The category's big binary is, mechanically, a text field.