Skip to main content
Who decides an answer tells you which pages the models leaned on. The win plan turns that into one thing to do — per model, with the evidence under it and the date we will find out whether it worked.

​
One action per model, never a list

A plan with five actions per model is a plan nobody works. So each model gets:
  • Up to five deciding sources, each with its share, its n, and whether you are on it already.
  • A rung: proven (you are cited among the pages that decide this answer), better (you are cited and a competitor sits above you), or new (nothing of yours is there).
  • One next action: fix a page, get listed, pitch a listing, answer a thread, or write something cite-able that does not exist yet — with the page or domain to work on and the citations it rests on.
  • A credits estimate for that action. It is an estimate; the meter on the run is what you are actually charged.
The models are never averaged. The move that wins a prompt on ChatGPT is not the move that wins the same prompt on Perplexity, and one blended plan would fit neither. Where a model was read on two surfaces — a consumer answer and an API answer — the plan takes the surface with the most answers behind it and says which one it took, rather than averaging two different products.

​
Where the rung comes from

The same rule the deciding-sources map follows: the static per-engine table until your own workspace has read at least 20 successful answers on that source, and your own history after that. Every rung says which of the two it came from.

​
What the plan will not say

  • No forecast. No predicted rank, no predicted lift, no percentage that has not been measured.
  • No movement language. Nothing “improved” or “climbed”: until a re-read comes back moved, there is no proven change, and the plan ends on that sentence in so many words. Once the workspace has a moved verdict, the footer may use movement wording.
  • No number without its n. The plan states the evidence rung it is entitled to — measured at 7 or more answers, published between 2 and 6, hypothesis at one.

​
When we read it again

Every plan ends with when the answer gets re-read. The date is the day the next booked re-read for that prompt is actually read, a few days after its +7, +14 or +30 day mark. If nothing has booked one yet, it says “after the first action lands” rather than naming a date nothing is going to keep.

​
Where to find it

  • On the prompt page, as the first section, How to win this answer. See Prompts.
  • The Daily win plan works from the same plan every morning.
  • In chat, whenever you ask how to win a prompt — the agent follows the same procedure and shows you the sections per model.
  • As the MCP tool win_plan. It takes a required prompt — a prompt id or the prompt’s exact text — and an optional days window. Ask for a prompt you are losing; deciding_sources says which those are.
  • Over REST it is GET /v1/win-plan, with the same payload as win_plan.
The plan is composed over the deciding-sources map rather than recomputed from the readings, so the map, the page and the tool can never disagree about which page decided an answer.