Anchor entropy: a better lens than ratio buckets

Anchor entropy: a better lens than ratio buckets

We slice profiles into buckets — exact-match, partial, branded, naked, generic — then argue over percentages. A more information-theoretic frame is to measure the entropy of the anchor distribution: how unpredictable is the next anchor, given the profile?

The intuition: natural profiles are high-entropy. Real people phrase links a hundred idiosyncratic ways, so no single anchor dominates and the distribution has a long, varied tail. Manufactured profiles are low-entropy — a handful of templated anchors repeated, producing a spiky, predictable distribution that compresses well.

On one hand, this captures something the bucket method misses: two profiles can have identical exact-match percentages but wildly different repetition structure, and the repetitive one looks engineered. On the other, entropy is sensitive to sample size and to how you tokenize anchors, so the metric isn't turnkey.

A 2023 analysis comparing penalized and healthy profiles found healthy ones carried substantially higher anchor diversity — many low-frequency unique anchors — while penalized profiles concentrated mass on few strings. Suggestive, not conclusive, given the usual confounds.

Limitation: I'm not aware of Google explicitly describing an entropy feature; this is an inferred model, not a documented one.

Practical read: count your unique anchors and their frequency spread, not just your category percentages.

Open question: does Google's classifier read distributional shape (entropy, concentration) rather than the human-legible category ratios we obsess over?