We publish one figure more than any other. A sixth of it rests on titles we could not check.

The number we publish most often is 7.7% — the share of the world’s Netflix top-10 places held by Korean titles, 37,750 of 492,900 across 93 markets and five years. It is on our front page, in our articles, and in the sheets we offer to companies.

This week we finally asked how much of it we have actually checked.

The rule, and the hole in it

A title enters our Korean count because its name matches a Korean work in Wikidata. That rule can fail in exactly one direction: another country made something with the same name.

The guard against it is Netflix’s own label. Netflix classes its global charts as English and Non-English, so a title that shows up on the English global chart is not the Korean work of that name, and we drop it. That guard works — it is how Suits and The Perfect Couple left our catalogue in an earlier correction.

But the label exists only on the global charts. A title that never reached a global top 10 has nothing to check against. We keep it, because dropping titles for lacking a label we cannot obtain would quietly delete real Korean shows.

Keeping is not the same as confirming. Here is the size of the difference:

Why we had never published this

We had. Sort of. Two of our pages have said for weeks that 197 titles carry no language label, and that sentence is true.

It is also the wrong unit. “197 titles” out of nine hundred sounds like a rounding error. 16.1% of the places behind our headline figure does not. The same fact, counted the way the number is actually used, is roughly six times more alarming — and it is the version that matters if you are deciding whether to rely on the figure.

We counted titles because titles were what the collector happened to have. That is not a reason.

A second check, which disagrees about which titles are the problem

The language label asks whether a title is not English. It never asks which work this is. For that we hold a Wikidata item number, matched by number rather than by name — the fix we made in August after discovering that name matching is case-sensitive and was silently dropping Korean films filed as LAND and DETOUR.

By that measure the picture is much better: 99.0% of the places sit on titles with an item number. Only 379 places — 1.0% entered on a name match alone.

Neither check is a superset of the other. A title can carry a Wikidata number and still have no language label; the reverse happens too. So we publish both rather than the flattering one.

What we did not do

We did not remove the 6,078 unchecked places to make the number cleaner. It would have taken one line, the headline would have moved from 7.7% to something slightly lower, and it would have been smaller without being truer — those titles are not known to be wrong, only unverified.

We also did not stop publishing 7.7%. A figure with a stated confidence is more useful than no figure, and considerably more useful than a figure whose confidence is not stated.

How this came up

It came from another desk. A colleague auditing a different dataset had been counting the rows their tool silently dropped, found none, and then went looking in the opposite direction — for rows it silently let in. They found 369 withdrawn workplaces being counted as active.

We had done the same audit that morning and stopped at the same halfway point: we had counted what our collectors throw away, confirmed it was nothing, and never asked what they wave through.

A pipeline has two silences and most people only ever check one.

What this means for the figure

7.7% stands. What changes is the sentence around it: of the places it counts, 82.5% are confirmed not to be English-language works of the same name, 1.4% are known to be ambiguous, and 16.1% are unverified in that specific way. On the other axis, 99.0% are identified by a database number rather than by their name.

If you use our figure, that is the disclosure you should have had from the start. It is on the world share page now, in the same table as the number itself, which is where it should have been.