Sentiment Is the Wrong Unit

Sentiment scoring survives because it produces a number, and numbers are comforting. It is 68 percent positive this month, up from 64. Something has apparently improved. Nothing has been learned.

A sentiment score averages across a corpus that nobody reads in aggregate. But perception is not formed by averages. It is formed by whichever three or four items happen to be retrieved at the moment of the query.

Ten thousand neutral mentions and one damaging, highly citable, structurally privileged document do not average out to mildly positive. They average out to the document, because the document is what gets retrieved and the ten thousand do not.

Consider two items about the same company. One is a mildly critical opinion piece in a low-authority outlet. The other is a neutral, factual, entirely non-hostile line in a structured public record stating an outdated legal status.

Sentiment scoring flags the first and ignores the second. The second is the one that will be cited by AI systems, surfaced in due diligence, and repeated as fact for years, precisely because it carries no tone at all. Neutrality is not safety. Neutral facts propagate better than hostile opinions, which is exactly why a wrong neutral fact is the most dangerous object in your pool.

Retrieval frequency: how often does this item appear in the answers your audiences actually receive. Citation depth: how many other sources derive from it. Structural weight: is it parseable, and does it live somewhere systems treat as ground truth. Correction cost: what would it take to change it, and is that even possible.

None of these are as easy to chart as a sentiment line. All of them tell you what to do on Monday, which the sentiment line has never once done.