We read 637,000 messages of stranger chat. Here is what actually works.

9 min read · updated 1 September 2026

Original research on 9,854 real anonymous conversations: how they open, why most die within ten messages, and the one thing that separates the ones that last.

Almost everything written about talking to strangers online is guesswork. Advice articles recycle each other, and the studies that do exist are mostly about dating apps — a completely different situation, where both people have already seen a profile and chosen to match.

So we counted it instead. What follows comes from 9,854 real anonymous chat conversations containing 637,832 messages, published as an open dataset. We measured shape rather than reading conversations: message lengths, turn-taking, question rates, and how far each conversation got before it stopped.

On the data. These are logs people chose to save and share, so memorable conversations are over-represented and the "hi → disconnect" mass is under-counted. Trust the ordering of the findings more than any single percentage. We also throw out one number entirely: the corpus contains zero non-ASCII bytes in 30 million characters, so its 0% emoji rate measures the publisher's filter, not human behaviour.

1. A message is four words long

The median message in stranger chat is four words. Not four sentences — four words. Nearly a fifth of everything sent is a single word: yeah, lol, same, nah. Only 5% of messages run past twenty words.

This is the single most violated rule online. People arrive from email and messaging apps where a considered paragraph signals effort, and here it signals the opposite: that you have not noticed what kind of conversation this is.

2. Mirroring is real, but it is not proportional

People do match each other's length — the correlation between a message and its reply is r = 0.48, which is strong for anything involving human behaviour. But the relationship is not a straight line, and this is where most advice gets it wrong.

They wroteTypical reply
1 word6 words
2–3 words7 words
4–6 words8 words
7–10 words9 words
11–15 words11 words
16–25 words16 words
26–40 words30 words

The curve is almost flat at the bottom. Someone who writes ten words does not get twice what someone who writes five gets — both get roughly one short line back. Replies only start scaling once the other person is genuinely writing at length.

3. Questions are how conversations open, not how they continue

Only 14% of messages contain a question mark. Not one in two — roughly one in seven. Five or six replies out of seven are somebody just saying a thing.

But the rate is not flat. It peaks around the third to sixth message, while two strangers work out whether there is anything here, and then falls away. And when we split conversations by outcome, the difference is clear: in the first eight messages, conversations that reached forty messages asked questions 20.5% of the time, while ones that died before ten asked 14.6%.

So ask freely at the start. It is the late questions that read as an interview — or as someone with nothing of their own to offer.

4. The strongest predictor is balance, and it is visible immediately

This is the finding we did not expect. We measured balance — the quieter person's share of the turns, where 100% is perfect back-and-forth and 50% means one person is doing twice the talking.

First 8 messagesDied (<10)Lasted (40+)
Balance67%100%
Questions14.6%20.5%
Message length2 words2 words
Multi-message bursts20.6%20.6%

Every conversation that got anywhere was in near-perfect alternation from the very first messages. Every one that died had somebody talking past somebody else.

Note what is identical in that table. Message length and burst rate are the same in both columns. If you compare whole conversations rather than their openings, the ones that lasted look like they use longer messages and more double-texting — but that difference vanishes once you control for warm-up. Those are consequences of a conversation going well, not causes of it. Copying them does nothing.

5. Sexual openers perform worst of every category

We classified every opening message and measured how far each kind of opener got.

OpenerMedian lengthReached 10+
Greeting + question38 messages77%
Question35 messages83%
Greeting only32 messages79%
Sexual opener21 messages66%

Opening sexually costs roughly a third of the conversation compared with saying hello and asking something. Whatever you are hoping the conversation becomes, leading with it is the least effective way to get there.

6. What an opener actually looks like

The median opening message is three characters and one word. Roughly half start lowercase and only 3.8% end with a full stop. The most common openers, in order: hi, hey, m, hello, f.

Dating-app research recommends a 40–90 character opener with a specific question. In stranger chat, only 6% of openers fall in that band. It is advice from a different room.

7. Almost nobody says goodbye

86.7% of conversations simply stop. Only 13.3% end with anything resembling a farewell. The median conversation runs 32 messages and then one person is gone.

This is worth knowing mostly so you do not take it personally. Someone vanishing mid-sentence is the norm, not a verdict on you.

So what do you actually do

Why we did this. voidani has AI companions for when nobody is around, and we needed to know what a real conversation looks like rather than what a language model assumes one looks like. Nearly every number above changed something in the product — the reply-length rules, the question rate, when it offers a game. The research is published because it is more useful shared than kept.
Try voidani