The same plumber counted four times - and why we did not ship the fix
Measured across 72 answers: 454 extracted names for 424 real businesses. The obvious fix absorbed 61 duplicates - and made 72 real shops disappear. We did not ship it.
In this article
Yesterday we wrote that our extractor could not read Japanese, Korean, Chinese or Arabic. Fixed, it reads them. Today's question follows: is what it reads correct?
The 72 answers collected yesterday - six markets, two trades, three assistants - were enough to find out, without paying for a single new call. We read them again.
One plumber, four lines
Inside a single competitor table in Tokyo, the firm Kurashian appears under four different labels: its name alone, its name followed by a colon and a sentence of description, its name in brackets, and its name with a ranking mention attached. Four lines, one business.
In Dubai, Urban Company appears three times - once per phrasing of its former name. In Lyon, Ring Twice, Yoojo, ThreeBestRated and StarOfService each appear twice: once alone, once followed by a dash and the word Lyon.
A table showing the same rival on three lines is not only miscounting: it dilutes that rival's share of voice and falsifies the ranking.
The tally, name by name across the six markets: 454 extracted strings for 424 real businesses. Thirty duplicates. Japanese holds half of them, but French has some too - and that detail is what made the cause interesting.
Two causes, one of which hits everybody
The first is an alphabet story, like yesterday. Our splitter cuts a name at its first separator: « Brand - great reviews » becomes « Brand ». Its list of separators is written in ASCII characters. But Japanese and Chinese write colons and brackets full-width, with no space around them: nothing matches, so the name keeps its whole sentence.
The second is not exotic at all. The split only applied to names found in a bulleted list. Names the assistant puts in bold mid-paragraph never went through it. That is why « Ring Twice » and « Ring Twice - Lyon » count as two in Lyon: French was affected all along, and nobody had measured it.
The fix, written and measured
One separation rule, shared by every extraction path, extended with full-width punctuation. About thirty lines. Re-run over the 72 answers: 454 strings become 424. Kurashian is one business again, so is Urban Company.
Then we counted what came in and what went out, answer by answer - and that is where it went wrong.
61 duplicates absorbed.
That was the point.
72 names lost
- among them Maison Deschamps, a real Lyon bakery, extracted before and no longer.
85 new entries
, including « check the qualifications » and « the word-of-mouth reflex ». Advice, not shops.
Why a correct fix breaks something else
The mechanism is worth understanding, because it is invisible in the code. A name longer than forty characters is rejected: an old guard, written to keep sentences out. In Japanese, an assistant's question typically runs forty-five characters, so it fell. Split at its bracket, it is down to twenty-one: it passes.
The length gate was therefore filtering noise by accident in the scripts that have no spaces. Fixing the split removed that unintended protection. And since a competitor table only shows eight names, every fragment of advice let in pushes a real one out.
We did not ship. The fix is written, measured, and filed in the repository with these numbers. It goes out when advice detection works in all four scripts: that is the real job, and it needs those rules written language by language.
What this changes for you
Count the rows of a competitor table before believing it.
If one name is there twice under two forms, the ranking is false and the shares of voice are diluted.
A tool can be wrong in both directions at once.
Ours inflated the number of rivals and let advice through: fixing one exposed the other.
Ask what is being counted, not only how many.
« Eight competitors » means nothing until you know whether all eight are businesses.
Be suspicious of a figure that never moves.
The cap of eight was reached in nearly all our measurements: that was a ceiling, not a result.
The limits
72 answers, six markets, two trades, a single morning of collection: this is a sounding. Sorting « business » from « advice » was done by hand, and some judgement enters it - a name like « Professional Setup » could be a company. The thirty duplicates need no judgement: they are strings beginning with the same name.
We are publishing a fix we did not ship because that is the useful information. A measurement tool that only told you about its successes would measure badly, and report it worse.