What AI assistants recommend for dating — and get wrong
More people now ask an AI assistant "what's the best dating app?" than read a review site. So we did something simple: we asked one, on the record, twenty different ways — and logged every answer. Then we did it again two weeks later.
Full disclosure up front, because it matters: we build an AI dating app (AI Exodus). We started this to see what assistants say about our own category. What we found is bigger than us, and mostly unflattering to the assistants. Here's the data.
How we ran it
We wrote 20 questions a real person might ask about modern dating — from broad ("what's the best AI dating app right now?") to specific ("is there an app where an AI clone of me talks to other people's clones?"). We asked each one in two modes:
- From memory — the assistant answering from its training, no internet.
- With live web search — the same question, but allowed to look things up.
We recorded which apps got named in every answer, then tallied them. Same questions, run on 2026-07-10 and again on 2026-07-24. The results barely moved — which is itself a finding.
Finding 1: It names the same three apps, over and over
Across 20 questions — many of them explicitly about AI and clone features that mainstream apps don't have — the assistant kept returning to the incumbents:
| App named | Answers (of 20) | Two weeks earlier |
|---|---|---|
| Tinder | 8 | 10 |
| Hinge | 8 | 10 |
| Bumble | 7 | 9 |
| OkCupid | 4 | 7 |
| Iris | 4 | 3 |
| Replika | 2 | 3 |
| Purpose-built AI-twin apps (combined) | 3, web-search only | ~0 |
Ask "what dating apps use AI twins or digital clones for matchmaking?" and the honest answer is a short list of small, new products. The assistant's answer was Tinder and Hinge — apps that do not do that. It reached for the biggest names it knew, not the ones that fit the question.
An AI assistant is an incumbency machine. It recommends what the internet has written the most about — which is almost never the same as what best answers your question.
Finding 2: It recommends a companion app for a dating question
Replika showed up repeatedly as a "dating" recommendation. Replika is an AI companion app — you talk to an AI, not to other people. It's a fine product; it is not a dating app. Recommending it for dating is a category error, and the assistant made it confidently, without flagging the difference.
This is the quiet danger. The assistant isn't just incomplete — it's sometimes wrong in a way that sounds authoritative. A newcomer to the space would walk away thinking Replika and Hinge are the same kind of thing.
Finding 3: New and niche apps are invisible without a live search
In memory mode, purpose-built AI-twin dating apps essentially never appeared. They only surfaced at all when the assistant was allowed to search the live web — and even then, only once or twice each, buried under the incumbents.
The lesson for anyone building or choosing: an AI's memory lags the market by a full training cycle. If a product launched after the model's cutoff, it does not exist to that model until (a) the model searches live, and (b) the open web has enough written about the product for the search to surface it. No web presence, no AI recommendation. It's the new SEO.
Finding 4: Ask "is it legit?" and it gets confused
The most striking failures came from identity questions. Ask an assistant whether a smaller app is legitimate and it frequently conflates the name with something unrelated — a crypto token, a survey site, a different company that happens to share a word. Brand collisions the assistant can't disambiguate become confident-sounding warnings or mix-ups.
For any product with a common word in its name, this is a real reputational risk: the assistant may answer a question about you with facts about someone else entirely.
What this means
If you use AI to pick tools or apps, take its recommendations as a popularity snapshot, not an answer. It over-indexes on brand volume, lags the market, and occasionally mislabels categories. For anything new, niche, or fast-moving, it will quietly steer you to the biggest incumbent.
If you're building something, the takeaway is blunter: the assistant will not recommend you until the open web describes you clearly and often. Web presence isn't a marketing nicety anymore — it's the input that decides whether an AI can see you exist.
Honest limitations
- This is one assistant and a 20-question sample — directional, not definitive. We're publishing the method so anyone can rerun it.
- We build in this category, so we had a reason to look. We've reported what the data said, including the parts where our own category is invisible.
- "Named in an answer" is a coarse metric. It captures visibility, not quality of recommendation.
Curious what an AI twin actually is?
AI Exodus is the AI dating app this study came out of. 10 minutes, free, delete anytime.
See how it works