How does the GEO MarketMap agent decide what to do?
Before every reply the agent reads the project's current state from the database: what was measured, what was produced, what you decided. From that it proposes 1 move by name, with the reason, and does it when you agree: the missing page, the draft, the checked figure, the brief for a developer. It asks a question only when you alone can answer it. Nothing is ever published to your site on your behalf.
What does the agent read before it answers?
- The project card: what your business is, read from your site's pages and re-read only when you press for it.
- Your project notes, which outrank everything measured. If the card is wrong, the notes are where you correct it.
- The tracked questions, the assistants' latest answers, the rivals you kept and the niche you agreed.
- The site check's findings, the strategy for the month and the queue of pages it implies.
- The record of every decision: what was agreed, when, and what was set aside.
It reads these at the moment of use, never from memory of an earlier conversation, so a change in Settings reaches the agent on its next turn by construction. How the stores fit together is in knowledge and rivals.
Why does it propose 1 move rather than a list?
It proposes one thing, by name, from that state. Not a menu. Not a question. The next move, and why that one.
A list of 5 options hands the decision back to you with no more information than you had. 1 move with its reason can be agreed, corrected or refused in a sentence, and each of those is a decision the product records. Two further moves sit under the first as chips, so the alternative is 1 press away without being the default.
What will the agent never do?
- Publish to your site, your social accounts or anywhere else. It writes the piece and tells you where the file is.
- Promise a ranking or a position in an answer.
- Show a price or a balance in the conversation. Every control says its cost before it is pressed, once.
- Name the suppliers or models behind a measurement.
- Give a percentage without the count it was measured on.
- Answer with generic advice. If the answer is a page, it writes the page; if it is a developer's job, it writes the brief.
Every reply is read by a guard before you see it, and a reply that breaks one of these rules is rewritten before it lands. The rules are the same ones the public feature page describes.
What is a turn, and what does it cost?
A turn is 1 message from you and 1 reply, including any tool the agent ran inside it: reading the plan, writing a draft, filing a brief. Long work, such as a measurement that takes 10 minutes, does not run inside a turn; it runs as a job the tab can watch, and the conversation gets 1 line when it lands. A turn has a small fixed cost, said on the pricing page; the work it starts is charged at its own rate, at the press.
- Can I ask the agent anything, like a general assistant? +
- It answers about your project and its search and AI visibility, and it will steer a general question back to that. It is not a general chat assistant, and during the setup it does not answer at all.
- Does the agent remember earlier conversations? +
- It reads the record, not the transcript. Every agreement and every produced document is written into the project, so a new conversation starts from the same facts without re-reading old messages.
- In which language does the agent reply? +
- In the language of your interface setting when you press buttons, and in the language you type when you write. The pieces it produces follow the project's market language, set in Project settings.
Last updated 15 September 2026Markdown version