Learning source, trust source, and culture source name three distinct roles a body of content can play for an 📝Large Language Model (LLM): teaching it how people talk, giving it something to cite, and showing it what people currently believe.
A learning source shapes the model's register — vocabulary, phrasing, the texture of how a subject gets discussed. Its value accrues to model capability and is consumed at training time. A trust source is what a model will cite and stand behind: attributable, verifiable, and institutionally legible. Its value accrues to the cited entity and is consumed at retrieval time. The two are routinely conflated because a single domain can serve both, but they are purchased differently, measured differently, and can move in opposite directions at the same time. 📝Ryan Edwards characterizes the distinction as the one AI providers have been drawing for years when they describe social platforms as excellent training material and poor citation material.
Culture source is the third role and the one least served by the first two labels. It is what tells a reader — machine or human — what a population currently believes, feels, and does: sentiment, emergent consensus, and the secondary behaviors that precede primary ones. Insurers and brand strategists have long paid for exactly this signal, which is why a longitudinal archive of unprompted human conversation has value entirely independent of whether a single sentence of it is ever cited.
The property that separates culture from the other two is decay. Learning value and trust value both survive age; an authoritative source from 2019 remains authoritative, and a register once learned does not unlearn. Culture value expires quickly and is destroyed outright by manipulation, because a signal describing what people genuinely think stops describing anything once it is authored to be measured. This is the mechanism underneath 📝Marketers Ruin Everything: coordinated marketing degrades the culture layer specifically while leaving the learning and trust layers comparatively intact. It also implies a structural tension for any platform that tries to sell culture signal as a product — the more the signal is worth, the more that platform is obligated to detect and discount inauthentic participation in it.
The three layers are routinely collapsed into one during a visibility scare — 📝Reddit's ChatGPT Citation Collapse (August 2026) produced a wave of commentary that read a single retrieval metric as a verdict on all three at once.
I find this taxonomy does more work than any citation metric I track. Most arguments about whether a platform still "matters" for AI are two people describing different layers and assuming they disagree.