The mistake that makes everything look promising
Any system that scrapes what people are saying about an artist will find the same post quoted, reposted, aggregated and re-aggregated. Counted naively, forty appearances of one opinion read exactly like forty people holding it.
That single error is enough to make a flat artist look like they are breaking. It is also invisible in the output — the chart goes up either way.
Counting occurrences instead
Each piece of chatter is filed once, with a fingerprint, and carries a count of how many times it has actually been seen. Forty people saying something and one person saying it forty times are stored as different facts and are not summed into the same number.
The same rule applies to evidence generally: a claim that reappears somewhere else is recognised as the same claim rather than filed again as a second, independent confirmation.
What this still cannot tell you
Whether the people are real. De-duplication catches repetition, not fabrication, and no amount of counting distinguishes a genuine small audience from a purchased one with any confidence.
What it does do is stop the tool from lying to you by accident, which is a lower bar than truth and a much more common failure.
Where this number comes from.
Audience chatter is stored once per distinct item with an occurrence count, so 40 mentions of one thing and 40 separate people saying it are recorded as different facts. Evidence carries the same fingerprint rule, so a repeated claim cannot be filed as a second confirmation.
A uniqueness rule on the stored record rather than a check performed at read time — a duplicate cannot be written, so it cannot be counted later.