Outcome Deals
Setup, Run, and Update
Three actions drive your outcome. Each one starts a chat with the agent. You run Setup once, Run many times, and Update when the evidence calls for it.
Where to find them
Autopilot starts Runs and Updates for you, once a day, after Setup. To start one yourself, open your outcome. Run next action starts a Run. The arrow beside it opens all three. You can also type / in any chat's message box and pick outcome-setup, outcome-run, or outcome-update.
At a glance
| Action | When | The agent | You | Changes your site |
|---|---|---|---|---|
| Setup | Once, first | Reads your site and Search Console history, asks questions, proposes a plan. | Answer, connect your code or name your platform, approve. | No |
| Run | Once a day on autopilot, all through the window | Makes one change to your site, with its prediction written down first. | Approve it or put it live, then tell the agent. | Yes, through you |
| Update | On autopilot when nothing is queued to try, or when you start it | Reviews the evidence and revises the plan. | Read it. Approve changes to the plan. | No |
Setup
Run it first, once. It's a conversation, not a form. The agent reads your site and your Search Console history before it asks you anything, so it only asks what it can't find out itself.
What it asks
- Your business. What the site offers, who it's for, and which parts of it are in scope.
- What to protect. The number is set by your deal: Google Search clicks to your site. So this is about what must not happen on the way there. Pages not to touch, claims not to make, words your brand doesn't use.
- Where it can act. Where your site's pages are edited, what you've tried before, and what you think might work. This is where you connect GitHub or Bitbucket, or tell it your platform. Connect your website.
Specific beats complete. “Don't touch /pricing, legal reviews it” helps more than a long brand document.
What it proposes
Before it saves anything, the agent shows you its plan. It lists the kinds of change it wants to start with (its levers) and a first set of predictions (its hypotheses). Each prediction names what will change and what result would prove it wrong. Push back on anything that looks off. Nothing is saved until you approve.
After Setup, the Levers and Hypotheses tabs fill in. Your site is untouched.
Run
Your main action, and the one Run next action starts. Each run makes one change.
What happens in a run
- 1The agent records any readings that are due on earlier changes.
- 2It picks one change: one page, one lever, one prediction. It ranks every candidate by the clicks it's expected to add in the time your window has left.
- 3It writes down its prediction before making the change, including what would prove it wrong.
- 4It writes the change. If your code is connected, it opens a pull request. If not, it gives you a handoff: exactly what to change and where.
- 5You put it live, then tell the agent in the same chat, with the page's address.
- 6The agent checks that Google has seen the change, then measures it on the searches that page was changed for.
Tell the agent when it's live
A change you publish without saying so can't be checked, measured, or counted. A short message is enough: “Live: https://example.com/blog/choosing-a-crm”. Why it matters.
Two kinds of change
Some runs go for quick wins on pages that already show up in Google: a sharper title and description, a clearer page, better links to it. These can pay off inside your window. Other runs build new pages for searches your site doesn't show up for yet. New pages usually take months to earn clicks, so they're what keeps working after your deal ends. The agent alternates between the two.
If your site has little search history yet, most runs build new pages.
When a run makes no change
Sometimes a run ends without a change. It may send you to Update because every open prediction is waiting on results. It may send you to Setup because something it needs, like access to your code, is missing. Do what it points to, then run again.
How often to run
On autopilot the agent runs once a day. Report each change live soon after you put it up, so the next run plans with it in view. It won't make a second change to a page Google hasn't seen yet, because two changes on one page can't be told apart.
Update
Update is the agent stepping back. It changes nothing on your site. It:
- Judges each prediction against what happened: supported, refuted, or inconclusive. “Not enough evidence yet” is a valid answer.
- Works out why, when clicks stay flat after changes have had time to show.
- Retires what didn't work, adds new predictions, and rebalances new ideas against proven ones.
Autopilot runs it when nothing is left to try, and in the last stretch of the window every run is an Update. You can start it yourself once a few changes have had time to show results. Running it every day doesn't help: results take time to arrive. Update can't change the number your deal is measured on.
A rhythm that works
| When | What to do |
|---|---|
| Day one | Setup. Autopilot starts the first Run soon after you approve the plan. |
| Every few days | Read the run emails. Approve or apply each change and report it live. |
| Once results come in | Autopilot runs Update. Read what it found and approve anything new. |
| The last days | Autopilot stops making new changes and measures what moved. |