Co-occurrence versus direct attribution

Two different kinds of relationship

The analysis tab shows relationships between entities, and it is worth understanding that it actually tracks two different kinds of relationship, not one. The first is co-occurrence: two entities simply showed up in the same response, with no claim from the AI that they are connected in any particular way. The second is direct attribution: the AI explicitly linked the two, for example saying one brand is known for a particular attribute, rather than the two just happening to appear near each other.

These two signals are not the same strength of evidence. Co-occurrence is a weaker, statistical hint. It tells you two things were talked about in the same breath somewhere in your data, which can be genuinely useful for spotting patterns worth investigating, but it does not tell you the AI thinks they are related. Direct attribution is a stronger signal, because it reflects an actual claim the AI made about the relationship rather than a coincidence of proximity.

A concrete example

If a particular attribute and a brand show up together in eight responses out of a hundred, that is co-occurrence: worth a look, but not proof the AI considers the two connected. If the AI instead writes that the brand is known for that attribute, that is direct attribution — a much stronger claim you can act on with more confidence.

Before you trust a number

Before you lean on a co-occurrence number in a report or a decision, it is worth clicking into the cell to read the responses behind it. This is especially important when the sample behind that number is small, since a handful of responses that happen to mention two things together can look like a meaningful pattern when it is really just a few data points. Reading the actual text tells you quickly whether there is a real relationship worth reporting or just two unrelated ideas that turned up in the same answer.

It's also worth checking that the entities involved are genuine in the first place — an unrelated entity that was never cleaned up can rack up co-occurrence purely by showing up often, which looks identical to a real pattern until you check it.