Why "signature word" does not just mean your most-used word

The stat is a comparison, not a ranking

If the signature word were simply your most-frequent word, every single chat in the world would return the same answer: "the", "you", "a", "i". Function words dominate any body of text by a wide margin, and they carry no information about who you are in a conversation. A stat that returns "the" for everybody is not a stat.

So the question the calculation asks is different. Not "which word do you use most?" but "which word do you use much more than the other person does?" That reframing changes the output completely. The winner is a word with a large gap between two rates, and a word can win with a modest number of uses as long as the other person almost never reaches for it.

This is why your signature word often surprises you. You are not being shown your habit in isolation. You are being shown the part of your vocabulary that is distinctively yours inside this specific chat, which is a different thing and, in practice, a more revealing one. The same person can have a different signature word in three different chats.

Step one: rates, not counts

Raw counts are unusable the moment two people write different amounts, which is nearly always. If one person contributes 60,000 words to the chat and the other contributes 20,000, the heavier writer will out-count the other on almost every word by default, and every signature word would just go to whoever typed more.

So each word is converted into a rate: uses per 10,000 words written by that person. A word used 30 times by someone who wrote 60,000 words is a rate of 5 per 10,000. The same 30 uses from someone who wrote 20,000 words is a rate of 15 per 10,000, three times higher. Same count, very different habit.

Per 10,000 is chosen for readability rather than mathematical necessity. Individual words in a chat are rare events, so per-100 or per-1,000 figures come out as long decimals, while per-10,000 keeps most interesting words in the range of whole numbers you can hold in your head.

Step two: the ratio, and the +1 that stops it exploding

Once both sides are rates, the gap is expressed as a ratio: your rate divided by their rate. A ratio of 4 means you use the word four times as often, adjusted for how much each of you writes.

The problem is obvious. Very often the other person has used the word zero times, and dividing by zero is not a number. Worse, near-zero is almost as bad as zero: if their rate is 0.01, the ratio comes out in the hundreds, and one accidental use by them would collapse it. The stat would be violently unstable at exactly the point where it matters most.

The fix is +1 smoothing. Before the division, 1 is added to both sides. A zero-rate denominator becomes 1, so the result is always finite, and the difference between "never used it" and "used it once or twice" stops producing wild swings. The cost is that the ratio is pulled slightly toward 1, meaning slightly conservative. That is the correct trade: a signature word should be reported with confidence or not reported at all.

Step three: the 8-use floor

Smoothing tames the denominator but it does not fix the numerator. A word you typed exactly once, that the other person never typed, still produces a ratio above 1, and there are thousands of such words in any long chat. Typos qualify. So do proper nouns that appeared during a single afternoon, and words autocorrect mangled once and never again.

So a word must reach 8 total uses across the chat before it is eligible to compete at all. Below that, no ratio is even calculated. Eight is high enough to exclude one-off noise and low enough to keep genuinely distinctive vocabulary that appears in real conversation rather than in every message.

These two guardrails do different jobs and you need both. The floor decides which words are allowed to be considered. The smoothing decides how their scores behave once they are. Remove the floor and the leaderboard fills with typos. Remove the smoothing and it fills with words the other person happened to use once.

The arithmetic on a made-up chat

Take a two-person chat. A wrote 40,000 words, B wrote 20,000. Here are four candidate words for A, with invented counts chosen to show each rule doing its job.

Word "honestly": A used it 60 times, B 4 times. A's rate is 60 / 40,000 x 10,000 = 15 per 10,000. B's rate is 4 / 20,000 x 10,000 = 2 per 10,000. With smoothing the ratio is (15 + 1) / (2 + 1) = 16 / 3 = 5.33. Eligible (64 total uses) and a strong gap.

Word "the": A used it 1,800 times, B 900 times. A's rate is 450, B's rate is 450. Ratio is (450 + 1) / (450 + 1) = 1.00. Both people use it identically once you adjust for volume, so despite being the most-used word by a huge margin it scores essentially nothing. This is exactly why your most-frequent word almost never wins.

Word "anyway": A used it 12 times, B 0. A's rate is 3, B's rate is 0. Ratio is (3 + 1) / (0 + 1) = 4.00. Eligible at 12 uses. Note what smoothing did here: without it this would have been a division by zero, and the word would have had to be discarded or given an artificial infinite score that beat everything.

Word "pterodactyl": A used it 2 times, B 0. Rate 0.5 against 0, so the smoothed ratio would be 1.5. It never gets calculated, because 2 total uses is under the floor of 8. Without that rule, a single joke about dinosaurs could define your year.

Ranking the three eligible words: "honestly" at 5.33, "anyway" at 4.00, "the" at 1.00. A's signature word is "honestly", a word used 30 times less often than "the".

How to read your own result

A high ratio on a moderately common word is the most informative outcome. It means the word is a real part of how you write and the other person genuinely does not share it. That is the case where the stat is telling you something true about your voice.

A word you do not recognise as yours usually means it clustered in a short period. Eight uses is a low bar, and eight uses in one week about one topic passes it. If the result feels off, that is normally what happened, and narrowing the recap's time window will show you whether the word survives across a longer stretch.

In group chats the comparison is against the rest of the group combined rather than one other person, which makes the denominator much larger and much steadier. Signature words in big groups therefore tend to be more strongly distinctive when they do appear, because out-rating several people at once is harder than out-rating one.

Signature words are single tokens. Recurring multi-word phrases are handled separately as top expressions, and the two stats often disagree, since a phrase can be shared by both people while one particular word inside it is not.

One last thing worth knowing: the ratio is directional. Your signature word and theirs are computed from the same table read in opposite directions, which is why the pair of results is frequently more interesting than either one alone.