What an AI SEO agent is
A chat assistant answers one prompt. An agent is given a goal and tools, and works through several steps on its own: read Search Console, crawl the site, find pages losing clicks, draft a fix, publish it, check the result.
In SEO, the tools are what make it an agent: access to your search data, your pages and, crucially, the ability to change them. That last permission is where the value and the risk both come from.
What agents do well
- Connecting data. Lining up rankings, Search Console clicks, crawl results and analytics for every page, every week, without anyone exporting spreadsheets.
- Noticing patterns. Pages with lots of impressions and few clicks, pages slipping from page one, pages competing for the same search.
- Drafting at volume. Dozens of title rewrites, alt texts, internal links or schema fixes in minutes, each with a before and after.
- Keeping the cadence. SEO rewards steady weekly work, and an agent doesn't get busy with other priorities.
Where AI SEO agents go wrong
The ways agents go wrong have a common shape: they're allowed to act at scale and nobody is reviewing.
- Content at volume. An agent that can publish can publish a lot. More pages without more value is what Google's spam policies call scaled content abuse.
- Confident mistakes. Google's guidance on generative AI notes that generative models don't retrieve facts but predict a likely sequence of words, so their output can contain inaccuracies.
- Unreviewed technical changes. A wrong noindex tag, canonical or robots.txt rule can remove pages from Google. An agent that publishes tags and structured data without review will eventually ship one.
- Risk to the whole site. The spam policies apply to sites: those that violate them may rank lower or not appear in results at all, so a large batch of low-value pages is a risk beyond the pages themselves.
Four levels of agent autonomy
It helps to be explicit about how much an agent is allowed to do. Most of the safety comes from choosing the right level for each kind of change.
| Level | What the agent does | Where it fits |
|---|---|---|
| 0. Suggest | Lists issues and recommendations | Research tools. Safe, but the work still waits for you. |
| 1. Draft for approval | Prepares the finished change; a person approves | The right default for content, outreach and anything unusual. |
| 2. Ship low-risk changes | Publishes reversible, mechanical fixes on its own | Titles, alt text, internal links, valid schema, sitemaps. |
| 3. Ship everything | Publishes all changes and content unattended | Where the risks above come from. |
- 0SuggestSafe
- 1Draft for approvalRight default
- 2Ship low-risk fixesNamed types only
- 3Ship everythingWhere it goes wrong
A sensible setup runs at level 1 for most work and level 2 for a short, named list of change types, and never at level 3 for content or anything sent in your name.
Guardrails to insist on
Before you give any agent write access to your site, check for these:
- Approval by default. Nothing goes live until a person says yes, unless you've opted specific change types into automatic publishing.
- Scoped automation. You choose which kinds of change can ship on their own, and risky ones (robots.txt, noindex, canonicals, redirects) are excluded.
- A record of every change, with what it was before, so anything can be reviewed and undone.
- Articles arrive as drafts. An agent can write the draft; a person decides whether it's worth publishing.
- Nothing sent in your name without you. Outreach emails wait for a person, every time.
- Sources for claims. Drafts should show where facts came from so checking them is quick.
- One accountable owner. One person is responsible for what ships, and every change can be traced to their approval or to a rule they set.
Build your own agent, or buy one?
You can build a capable SEO agent with a workflow tool or a coding assistant, an SEO data API and your CMS's API. It's flexible and cheap to start. The cost is maintenance: APIs change, edge cases break silently, and the guardrails above are yours to build.
Buying makes sense when you want the guardrails, the publishing integrations and the weekly cadence handled for you. Either way, apply the same checklist.
How Sojig's agent is limited, by design
We make Sojig, so here's how its autonomy works, as built. You choose a mode in approvals and autopilot. Manual is the default and publishes nothing on its own. Auto publishes only low-risk fixes. Full auto publishes every fix it can, but articles and new pages only ever become CMS drafts. Some fixes always wait for a person in every mode, and outreach emails always wait because they go out in your name.
Strategy happens in conversation with Ady, the assistant inside the product, and nothing Ady suggests changes until you click Apply.
Illustrative example