Chat statistics, and what they are actually worth

Counting messages, and two ways to get it wrong

A WhatsApp export is not a list of messages, it is a list of lines. One message can take five of them, line breaks and blank lines included. Counting lines therefore overstates volume every time, and the error is not uniform: it grows with people who write long, which distorts exactly the comparison you are trying to make.

The mirror mistake is throwing away media. A <Media omitted> line carries no text, but it is a real turn in the conversation. Across a corpus of 196,830 messages, media accounts for 27%: ignoring it erases a quarter of the conversation, and specifically the quarter belonging to whoever answers with voice notes instead of text.

The usable rule: a media message counts for rhythm, initiation and balance, never for text statistics. Average message length, vocabulary range, number of questions: those are computed on text alone, otherwise a voice note counts as a zero word message.

Who texts the most, and why that is not enough

Raw volume takes two lines of code and three seconds to read. The problem is that it blends three very different behaviours:

• Volume: how many messages, all types together. Somebody who writes in ten short bursts what another says in one message looks twice as invested.
• Initiation: who opens the conversation after a silence. This is the number that answers "am I the only one reaching out", and it has nothing to do with volume.
• Follow-up: who writes twice in a row without getting a reply. That one is the hardest to read, and the most telling.

All three can point in opposite directions in the same conversation. Somebody who writes fewer messages but opens eight conversations out of ten is not less present, they are present differently.

Reply time, and its three definitions

It is the most quoted measure and the worst defined. Three choices change it completely, and nobody ever says which one they made.

A reply time quoted without those three details is not comparable to another, even on the same conversation.

Rankings need a floor

In a group, everything turns into a podium, and that is where counters get ridiculous. In one real conversation from the corpus, a participant wrote 1 message out of 14,653. With no participation floor, they show up in percentage rankings, sometimes at the top: their single message was a joke, so they become "funniest in the group" at 100%.

A work thread with 54 participants, all of them raw phone numbers, over three months, defeats every ranking at once. A counter that holds up on that one holds up anywhere.

The number that says the most: the slope

A total is a photograph, a slope is a story. In the corpus, one thread goes from 18,645 messages in one year to 794 the next. No total would have told you that, and nobody in that conversation saw it coming: the decline spreads over months, and each week looks like the one before.

That is why splitting by year or by half year beats one global figure. Seven years of history gives you something to work with; three months only shows a state.

What no counter can tell you

An export holds no deleted messages, no read receipts, and nothing that was said out loud. A silence in the file could be an argument, a house move, a week of childcare or a broken phone. The file says there was a silence and since when, not why.

Any reading that jumps from observation to cause invents half of what it claims. A dated number beats an explanation: "since March, 6 messages out of 10 are under five words" can be checked, "he lost interest" cannot.

If you want to see what these measures look like on your own conversation, start with the export. And if you are wondering where that file goes, that is a good question.