AI Search — The Owners of the Facts Are Down 20–40%, the Answer Boxes Sit Near Their Highs

Google’s leaked 2024 ranking documents contain an attribute called contentEffort — a language-model estimate of how much human effort went into a page. This week the search trade rediscovered it and drew the obvious conclusion: own facts the machine cannot generate, and the machine has to cite you. The tape has been running the same experiment for a year, with real money, and it reached a different verdict. The companies that own the facts are 20–40% below their highs. The two companies that own the answer box are near theirs.

The leak, and what it does and does not say

The attribute is real. In the May 2024 Content Warehouse leak, contentEffort sits in a page-quality module and is described as an “LLM-based effort estimation for article pages”; a separate attribute scores original content. Google’s AI answers are retrieval-based — they assemble from indexed sources rather than generate freely — and Google holds a patent on ranking documents by how much new information they add to what a user has already seen.

Two things in the popular version are not in the documents. The words “originality” and “talent or skill” come from Google’s guidelines for human quality raters, not from fields next to contentEffort in the API. And nothing in the leak says the score gates AI Overviews; the documents predate that product’s rollout and describe attribute names, not weights or whether anything is live.

So the heuristic survives — if a prompt can rebuild your page in ten seconds, the machine has no reason to cite you — but as a heuristic, not a mechanism. The market’s version of the same test is more instructive, because it has prices attached.