A MentionBeat agent is an assistant that does generative-engine-optimization work inside your workspace — measuring, reading the pages that beat you, drafting, queueing fixes. It reaches 40 tools built on your own data, 9 of which can write. Every run is priced before it starts and logged step by step, and it stops before anything that cannot be undone.
The reason to automate GEO is not that the thinking is hard — it is that the loop is repetitive and nobody finishes it. Agents exist to carry the repetitive half. A run starts in whichever of these four shapes fits the job, and none of them spends a credit before an estimate is on screen.
Ask for an outcome in your own words — “find the three prompts we lost this month and tell me why” — and the agent plans, calls what it needs and answers. Threads are kept, named and re-openable.
When a conversation passes its retention window it says so at the top of the transcript, rather than quietly serving you a beginning that is missing.
Eighteen named jobs ship with the product, written as GEO procedures rather than prompts: answer the prompt we lost, tear down the page that is beating us, correct a wrong answer an engine is giving about us, tune the prompt suite.
Edit any of them for your workspace, and revert to the shipped version whenever the edit stops earning its keep.
One instruction, run down every row. Paste a list of pages, prompts, competitors or product URLs, add a column that says what to work out for each, and the grid fills in.
Columns are priced individually, a single cell can be corrected by hand, and anything that depended on that cell knows it changed.
Named page shapes — comparison, alternatives, FAQ, category explainer — each carrying the evidence it needs before it can be written, and the case for using it at all.
Building the same page family from your approved brief, without an agent, stays free in Content Studio.
The uncomfortable question about any agent is what it does while you are not watching. MentionBeat answers it with three modes you pick per run, and the mode sits on the composer where the work starts rather than buried in a settings page.
| Mode | What it does | The caveat that comes with it |
|---|---|---|
| Plan | Plans only. Nothing is carried out. | Read-only tools still run, so the plan is grounded in real measurement rather than a guess. Every change is recorded as a proposal and executed by nobody. |
| Manual | Asks before every change. | It runs until it wants to change something, then stops with the exact call in front of you. Approving continues the same run — one trace, not two. |
| Auto | Runs to the end without asking. | Auto removes the human, not the rules. Every change still passes the policy gate, the per-run ceiling and the kill switch. |
Manual is the default. A tool that shipped Auto as the default would be optimising for the demo rather than for the Monday morning after it.
Forty tools, and the split matters more than the total: 31 of them only read. They pull your visibility scorecard, the prompts you lost, the answers verbatim, the rivals named instead of you, the domains engines cite, your product brief, your site audit and the open web.
Nine can write, and every one of them writes inside MentionBeat rather than out on your property — queue a fix, compose a draft, request a measurement run, propose a skill, track a source, call an endpoint you configured yourself.
The trace records the difference between called and proposed — not run, so a plan never reads like an action after the fact.
Agent runs consume the same credits as measurement, metered per tool call. Because an agent decides for itself how many tools a job needs, the honest estimate is a range: one call if the first answer is enough, up to the ceiling if it uses every step. Both ends are shown before you start.
A run with no tools costs nothing, and says so — the model is answering from your brief alone. Agents are on every plan, including Free. Credits are the only limit, and there is no automation tier to buy.
A chat window that forgets is a demo. Work an agent finishes becomes something you can find again, version, hand to somebody, and send out — deliberately, from one place.
Documents, batches, reports and saved views written by a run, each with its history. Correct one and the things downstream of it know. Pages built from your brief stay in Content Studio — nothing appears in both places.
One surface sends anything outward, and it dry-runs first. CMS destinations receive drafts — never a published page — so the last read before your site changes is a human’s.
A finding an agent turns up lands in the same fix queue as everything else, and closes on the same terms: only when an automated re-check confirms the page really changed.
Agents are included on every plan, including Free. Start in Plan mode and read what it would have done.