Fifteen accounts out of 2,164 could not be classified by anybody, and that is the sentence in the whole study we are proudest of. It is seven-tenths of one percent. It is a rounding error you could bury in any footnote and no reader would ever ask after it. We put it on a shelf instead, in its own labelled box, and left it there.
Here is how the box came to exist. The raw material was 202 different platform labels, the free-text things accounts write about themselves, a list as unruly as any you have seen. We read those 202 down to thirty-one categories, which is already an act of interpretation nobody should pretend is neutral. Where a label was too thin to trust, we read the biography underneath it, the actual sentences a human being wrote about what they do. And at the end of all that reading, fifteen accounts still refused to sit in any of the thirty-one. Not because we ran out of patience. Because forcing them anywhere would have been a lie with a decimal point on it.
Think about what the alternative buys you. Coffee sorts itself the same way a good roaster sorts a harvest: most cherries fall cleanly into grades, and then there is a small pile at the edge of the table that is neither ripe nor discard, and the honest roaster does not sweep that pile into the nearest sack to make the tally come out round. The dishonest one does, and the tally looks magnificent, and the cup is quietly worse for it. A hundred percent classified is the roast that swept the table. It reads as competence. It is usually just a bucket labelled “coffee shop” that has three accounts in it that are not coffee shops.
This is the reason the piece exists, and it is aimed squarely at anyone who has ever bought a social listening report and believed the totals. The totals are the least interesting thing in the document. What you should be reading for is the residual: the category that admits the instrument has an edge, that some of the world arrived in a shape the schema did not anticipate. Ask the report what it did with the accounts it could not place. If the answer is nothing, if every single row landed somewhere clean, you are not holding a more rigorous study than ours. You are holding one that decided fifteen inconvenient truths were worth less than a tidy pie chart, and made fifteen small classifications up to get there. Multiply that instinct across a dataset of two thousand and you can no longer tell which of the confident numbers were also invented.
We would rather you trust the thirty-one categories precisely because you can see the fifteen we kept out of them. The residual is not the study’s weakness admitting itself; it is the study’s warranty. It is the load-bearing “we checked, and these did not fit” that makes every other number in the document worth the pixels it is printed on. Confidence without a visible margin of doubt is not confidence, it is salesmanship, and the two look identical on a chart. The fifteen are our margin of doubt, printed at full size, so you never have to take our certainty on faith. You can read the full method in How The World Talks About Coffee, and the first thing we would point you to is not the biggest bar on the chart. It is the smallest, the one that says fifteen, the one that cost us nothing to hide and everything to keep.
So the next report that crosses your desk, do the ten-second test before you believe a word of the executive summary. Find the “other”. Find the “unclassified”, the “unclear”, the honest little shrug in the corner of the taxonomy. If it is there, somebody did the reading. If it is missing, somebody did the rounding, and rounding is where reports go to become fiction with good typesetting. A clean hundred percent is not the sound of a job finished. It is the sound of fifteen questions that never got asked, filed under answers.
