SEO/GEO Bench live

Tracked terms sitting in the top ten, counted weekly across the six client domains. Nothing moves before week four: the first pages are still outside the index, and a page does not count until it has been indexed for twenty-one days.

Six fresh domains per run, matched on a DR 0–4 authority band and on niche, each with the same sixty tracked terms held constant across every arm for the full sixteen weeks. A task is one delivered unit of work — a technical fix, a brief, an article, a landing page, or an answer block. Positions are pulled weekly from the client's own Search Console, and a page does not count towards anything until it has been indexed for twenty-one days.

Citation share is the GEO half. Each week a fixed prompt set is put to four answer engines and the cited sources are read; the share is the fraction of those prompts where the client's domain is among them. It moves on different work than rankings do, which is why a system can lead one column and trail the other.

The adaptive arm rewrites its own term priorities, internal linking rules, and brief structure from what the last four weeks returned. The control arm is the same agency with those three frozen at their week-one settings, and it is what the multiple in the headline compares against. Cost is everything billed to deliver a task — model tokens, research seats, crawl budget, and publishing — so a per-task floor survives even on the cheapest model.

We define and run this benchmark ourselves. Other harnesses ran with stock tooling and our prompting, so treat the ordering as directional rather than as a settled ranking of the harnesses.