Platform
AI visibility monitoring Competitors Brand AEO Score Content Studio Content library Campaigns Search & AI demand Site & page audits Citations & gaps Agent ControlResources
Learn
Knowledge center Research Data News & updatesDevelopers
Developer docs Developers & API OpenAPI spec llms.txt Free Tools Pricing Log inPlatform · Autonomy
Agent Control is the decision desk for the content loop. An agent works the steps nobody needs to watch — reading the gaps your monitoring already found, generating the brief and the article, then proposing a publish with its destination and its source query attached. The default is manual, so nothing runs on its own. Propose mode auto-drafts and stops. Autopilot is an opt-in you can revoke in one selection, and every proposed, approved, rejected or refused action is written to an audit trail. Pages that go stale are refreshed in place, on the same URL, through the same gate. And the decision no longer needs the dashboard: the notification email carries an Approve and publish button, and replying to that email asks for a change.
Real AI queries · no credit card · 3 free checks a day
Detection, the brief, the article and the proposed publish are the loop’s work. The publish is the step that stops: it lands in the queue with its destination, the query that prompted it and a link into the draft, and it sits there until you approve or reject it — with an optional reason recorded on a rejection.
A brand begins in Manual: nothing runs on its own and every step is one you start. Propose is the next rung — auto-drafts content, but you approve every publish. Autopilot is the third, and it is a choice you make rather than a level you drift into.
In Autopilot the publish no longer waits for a person, and three things still hold: a destination must be connected, the fact-check still applies, and every run is recorded. Switching back to Manual or Propose is one selection and applies on the next load.
For a key driving the loop, auto-publish unlocks only after its own record reaches 20 reviewed proposals with 95% approved. Below that line the key proposes. The number comes from the decisions you have already made, so it reflects experience rather than a setting someone typed.
A published page that stays uncited past its observation window is refreshed in place — the same URL, no duplicate article — and your live page only changes when the update is published. The refreshed draft is fact-checked again before it ships, and a page that is already cited, or still inside its window, is left alone.
Each proposal, approval, rejection and refusal is recorded with the action, who or what requested it, the destination, the exact prompt that was sent, the credits it spent and what came back. Technical details stay collapsed on the card until you ask for them, and the same records are readable per key through the public API.
The “your article is ready” email is no longer only a notification. It carries the draft, what the agent is unsure about, and an Approve and publish button. Approving runs the same account-scoped approval the queue uses — the same row, the same audit entry — so the only thing that changed is that you did not have to open the app to make the call.
The Reply-To on the email carries that thread, so replying “make the headline more bold” is the whole instruction. The agent applies it to that draft and emails back the revised article with a fresh Approve and publish button, so the loop can continue from the inbox. The reply is handed to the revision engine for that one article; it is never interpreted as a command that could choose a destination, approve itself or change a limit.
Above the button the email lists what the agent is unsure about, most serious first: claims it could not back up with your sources (quoted back to you), drafts written without your documents, unfinished placeholder text, claims it is only partly sure of, an article that has not been fact-checked yet, a short article, a draft without structured data. When there is nothing to raise it stays silent rather than promising the draft is clean.
The button opens a confirmation page — the publish happens on the confirm, with the destination and the flags shown again, so a preview or a forwarded link cannot publish anything. Each link is single-use, expires, is bound to one account and one draft, and is revoked when the draft moves on.
“publish” or “approved” in a reply takes the same path as the button, and “reject” or “no” declines the proposal and records it. “stop” switches the email channel off and confirms once; only you can turn it back on. Revisions are capped per day and per thread, so a long thread cannot run up a bill on its own.
Only in a mode you selected, and only for a key that earned it. New brands start in Manual. Propose drafts and waits for your approval. Autopilot is an explicit selection, publish still requires a destination, and the agent cannot waive the fact-check. You can move back to Manual at any point.
The detection, the brief and the article. On its scheduled pass the agent takes an opportunity forward on its own, then creates a pending approval for the publish — the queue in the screenshot is that hand-off. Nothing reaches your site until you approve it there.
If it stays uncited past its observation window it is treated as stale, and the agent regenerates the article in place with a refresh directive. Your live URL does not change until the update is published, the previous draft is kept as a revertable revision, and a page that is already cited is never refreshed as maintenance.
The same gate as Content Studio. Facts have to come from the numbered sources in the brief, the finished article is checked against them, and anything unsupported blocks the publish. For an agent the refusal is absolute — it cannot supply the override reason a person can.
Open any card: the action, who requested it, the destination, the source query and the technical payload on demand. The same records are available per key through the public API’s audit endpoint, including the exact prompt, the credits spent and the HTTP result.
No. Manual and Propose are permanent settings. Autopilot is a convenience for a workflow you already trust — not a milestone you are expected to reach.
Yes — the button opens a confirmation page and there is no sign-in. The publish happens when you confirm it, not when the link is opened, because mail scanners pre-fetch links and a link that published on open would publish the article by itself.
Plain instructions about that draft — headline, angle, tone, length, a section to add or cut — as you would say them to a colleague. “publish”, “reject” and “stop” are also understood as the decisions they state. A reply cannot pick a destination, approve itself, spend beyond the plan or change a setting.
A revision is the normal content-revision charge, the same as asking for one in Content Studio; if the account is out of credits the agent says so in its reply instead of failing silently. Revisions are also capped per day and per thread, after which it declines without charging anything.
