GEO agent
How the GEO agent probes the engines that answer questions, what it holds against your pages, and how its fixes reach you as pull requests.
What it does
The GEO agent audits how your product reads to the engines that answer questions. Every two weeks it asks your own model a fixed set of prompts and records whether you or a competitor came up, then names the page that should have earned the citation and says why it did not.
Nothing it finds is a proposal. A site fix is a recommendation with a pull request behind it, so the audit writes the recommendation and a separate fix run writes the code.
When it runs
Every 2 weeks. It is on as soon as the product has a Site surface, since auditing publishes nothing. One reading of what the engines answer means little on its own, so each audit waits for two weeks of it. The cadence comes from the agent and is not editable, and it has no weekly cap. After each audit, up to five fix runs start on the most urgent open recommendations that have no pull request yet.
What it needs
The GitHub connector, and a Site surface. The reason is plain: a page is fixed by a pull request on the repository that serves it, so without the repository connected the agent has nothing to leave you and does not run.
What it reads
The foundation doc, the growth plan and the product memory first. Then the site as an answer engine reads it.
- The site snapshot, rebuilt that day: robots.txt, the sitemaps, the pages with their titles, meta, headings, canonicals, structured data and links, plus the files crawlers look for at the root such as
llms.txt. - The week's visibility probe: up to fifteen prompts, derived from the foundation doc and editable by you, asked of your own model on your own provider key. When the probe cannot run, the audit runs on the site alone and says so.
- The recommendations already open, so it weighs them instead of writing them twice.
- The market-index reads of a Site surface, for what happens outside your own traffic.
What it leaves you
No proposals at all. It leaves recommendations: one per page and problem, rated critical or warning, kept from week to week. The fix runs that follow open one draft pull request each, and you approve a recommendation by merging that pull request or dismiss it for good. A later probe or crawl is what resolves one.
The skills it runs
It enters through ai-seo, and may activate schema and competitors.
Editing its instructions
The agent's page in the app carries the instructions it is given at the start of every run, and you rewrite them there; the probe prompts are edited from the same place. The cadence is not editable. The switch reads on, off or paused, where paused means Kwern stopped it for a reason and will resume it when that reason clears.
Competitors agent
How the Competitors agent watches the pages your rivals publish, what it counts as a substantive move, and the digest it writes each week.
Google Ads agent
What the Google Ads agent reads from your account, the five kinds of change it proposes, the monthly budget it paces them against, and how campaigns start from a chat.