When does planting more documents in a review corpus stop being worth what it costs?

## Why this is a judgment and not a calculation The mechanics are settled by the time this question arrives: one passage owns one query neighbourhood, copies buy nothing, breadth costs distinct documents. What remains is a decision somebody has to own — how far to push, and what to claim from where they stopped. It has no single right answer because the inputs are a cost curve, a value distribution and a tolerance for exposure, and reasonable people weigh them differently. ## The cost curve rises, and it rises for a structural reason The order in which questions fall is not random. Thin neighbourhoods go first: a question the corpus barely answers can be taken with one aimed document. Once those are gone, what remains is the well-covered material — exactly the questions that many genuine buyers already answered in detail — and each of those needs a plant that clears a stack of established passages, often several documents per phrasing. So the count is not linear in the questions moved. The first five neighbourhoods might cost seven documents and the next five might cost forty. Anyone quoting an average is hiding the shape that matters. ## The value curve is uneven the other way Moved questions are not worth the same. A handful sit immediately before a purchase decision — fit, compatibility, safety, whether the cheaper option is equivalent — and those carry nearly all the payoff, because a false answer there changes what somebody does. The rest of the family is decoration. That asymmetry is what makes the stopping decision tractable at all: the right move is usually to stop far earlier than "we covered the family", once the decision-adjacent questions are held. ## Three costs that are not document counts 1. **Attribution accumulates and cannot be withdrawn.** Every post is tied to a listing and an account. A campaign that reaches thirty documents has left thirty attributable writes, and the pattern of a single controlled set of listings producing consistent, similar-purpose content is more visible than any single post. Later deletion does not undo the fact that the write happened. 2. **Held ground decays.** A live marketplace keeps producing genuine reviews and answers. A neighbourhood held today can be outranked next quarter by nothing more than ordinary buyer activity, so coverage is a subscription, not a purchase. Budget for upkeep or accept a shrinking result. 3. **The pipeline can move under you.** Chunking, encoder, or ranking changes shift every position at once, and none of that is observable from the write side. The dependency is on a configuration nobody involved controls. ## The alternative-use test The strongest form of the stopping question is not "is one more document worth it" but "is this construction still the cheapest route to this payoff". Corpus poisoning buys influence over answers to questions people ask. If what is actually wanted is influence over one specific decision, and there is a shorter path to it, the marginal document is the wrong purchase however cheap it is. A lead who cannot say when their favourite method stops being the right method will over-invest in it. ## What the stopping point is worth as a result In a sanctioned exercise the count is the deliverable's substance. "Eleven of forty enumerated buyer questions moved, for thirty-one distinct posts, of which the first six questions cost seven posts and the rest cost twenty-four" is a statement about how much protection genuine coverage actually provides — and it is falsifiable, because someone can re-run the same phrasings later. Be careful about what is being claimed on either side. A count is not an assurance that a cheaper route does not exist; someone better at aiming passages may move the same questions for fewer documents, and a corpus with different coverage would price differently. Equally, a small count of moved questions is not "the assistant is fine" if the moved ones were the decision-adjacent ones. Report the shape of the curve, not a single number, and let the person who owns the product weigh it. ## What a weak answer looks like It treats the budget as the only constraint and pushes until the money runs out; it quotes an average cost per question and hides the rising curve; it ignores that the writes are attributable and permanent; and it assumes ground taken stays taken in a corpus that people are still adding to every day.