Agentic SEO
Hand the regular SEO work to an AI agent: what to paste, how much to let it do, and routines to give it.
An AI agent such as Claude can do most of what you do in Copilot SEO: add your sites, read the numbers, work through the Action plan, record changes and report back. You decide how much it does on its own.
What it still needs you for
- Creating your account.
- Connecting Google. Google asks you to approve access in your browser.
- Paying, if you need more than the free plan gives.
- Letting the agent in. Approve it when it asks to sign in, or create a token for it. See Connect an AI assistant.
The agent tells you when it needs one of these, and gives you the link.
Start with one prompt
Copy the starter prompt from Use Copilot SEO with your AI agent or the Connect AI assistant page into your agent. It reads our guide for agents, then:
- checks what is set up, and sends you links for anything only you can do
- asks which sites, your brand names, what you want from it, whether it can edit your site, how hands-off to be and how often to check in
- adds your sites, waits for the first import and sets your brand terms
- shows you the top of your Action plan
- suggests routines that fit your answers
How hands-off?
| Level | The agent | It needs |
|---|---|---|
| Report only | Reads your data and reports. Changes nothing. | A read-only token is enough |
| Suggest and ask | Proposes each change and waits for your yes. This is where it starts. | Write access |
| Hands-off | Makes the changes you agreed on, records them and reports afterwards | Write access, and a way to edit your site |
Copilot SEO never changes your site. Hands-off means the agent edits it, for example through your code or your CMS, and then records the change here.
Routines
| Routine | How often | Good for | What the agent does |
|---|---|---|---|
| Rewrite the five weakest titles | Monthly | More clicks | Finds pages with lots of impressions and a weak CTR for their position, drafts new titles and descriptions, records them once they are live, and next month reports what moved |
| Weekly Action plan digest | Weekly | Business owners | The few changes worth making this week, anything that moved sharply, and reviews that came back |
| Annotate every release | After each release | Developers | Adds an annotation with the version and what changed, so you can read a rise or fall against it |
| Monthly report across sites | Monthly | Agencies and anyone with several sites | Every site side by side against the period before, what moved and why, and the next actions, written to send on |
| Decay watch | Every two weeks | Blogs, guides and other older content | Pages losing clicks, the searches they lost, what lines up in the annotations, and what to refresh |
| Experiment check-in | Every two weeks | Bigger changes, such as a new template | How each running experiment is doing, without calling a result too early |
| Onboard a new site | Once for each site | New sites and clients | Adds the site, sets up URL groups for its main sections and shows its first actions |
The prompt for each routine is on Use Copilot SEO with your AI agent, ready to copy. An agent connected through the MCP server also sees each one as a ready-made prompt, along with one called get_started.
Decay watch needs a plan that keeps at least 56 days of history. See Plans, trials and billing.
When do routines run?
When your agent runs them. We do not run routines for you, and there is nothing to set up in Copilot SEO.
Ask your agent to schedule a routine in whatever it runs in. Many agent apps can run a task on a schedule. Developers can run one from cron or a CI job with the REST API. Or paste the prompt whenever you want it.
Copilot SEO keeps working without an agent: it imports new data every morning, refreshes the Action plan, reviews your changes and sends the weekly report.
What to expect
- The numbers run about two days behind, as everywhere else in Copilot SEO.
- A change's review arrives about 30 days after it is marked implemented. The outcome is a direction, not proof. See Working an action.
- One change per action. If the agent changes three things on a page at once, the review cannot say which one worked.
- Agents can misread data. If a report surprises you, check the figures in the app before acting on it.