Corrective proposals

When the numbers stop moving, the report emits typed, evidence-backed proposals — each naming the pages, the reason, the human action and the machine instruction.

What it is

Six proposal types, each emitted only when the data calls for it: strike (refresh striking-distance pages), cadence (publishing fell short), stuck (published keywords still outside the top 100 after four weeks), links (short of the next stage's referring-domain gate), health (open crawl errors) and season (in-season terms with no page yet). Each proposal carries a key, a title, why (with the actual keywords and pages named), a human action, and a forClaude field written as a machine-executable instruction.

Also called: what should I fix · SEO recommendations · action items · proposals

See it
Corrective proposals
Prepared for
Ronan Petsch
1140 Ridge Line Rd · Fair Play
Shell package$33,325
Concrete & site$9,675
Doors & windows$6,450
Trim & finish$4,300
Total
$53,750
Signature
A proposal card: title, the why line with real keywords in it, the action, and Approve / Dismiss. Sample data — no customer information appears here.
How it works
  1. 1Each proposal type has its own trigger condition computed from the report's own data — no proposal fires without evidence.
  2. 2The why field quotes the real keywords, positions and pages so the recommendation can be argued with.
  3. 3forClaude is a directly executable instruction, which is what lets the implementer cron and the terminal feed act on the same row.
  4. 4Proposals are stored as JSONB inside the report, so the report and its actions can never drift apart.
  5. 5The autopilot's implementer applies only the keys in autonomy.autoApplyProposals; the rest stay pending for a human.
Why we built it

A verdict without a next step is a status update. Each proposal states the stake in the builder's own terms — the season proposal reads "Demand is peaking NOW on: … Miss the window and the next one is a year out"; the cadence one reads "Rankings compound off fresh, keyword-matched pages; a silent week stalls the whole ladder." Writing the machine instruction into the same row as the human one is what made autonomous implementation possible later without a second format.

The problem
  • Reports named problems without naming actions.
  • Recommendations were generic and could not be checked against evidence.
  • The same recommendation had to be re-explained to whoever did the work.
Sound familiar?
What you get
Every recommendation names the pages and positions it is based on
The action is written once and read by both a person and the implementer
No proposal fires unless the data actually triggers it

See it on your own jobs

Twenty minutes, your numbers, no slide deck. We’ll build one of your real buildings in front of you and send you the estimate link at the end — yours to keep either way.

or keep browsing features →