Why autonomous AI agents can't run your SEO (yet)

Autonomous AI agents can't run your SEO yet because the parts of SEO that actually move rankings are judgement calls, and unsupervised agents go off the rails on exactly those. They're brilliant at the busywork. They're unreliable at the deciding. And SEO is mostly deciding.

An 'autonomous agent', in case the term is new, is an AI that doesn't just answer a question - it takes actions on its own: researching, writing, publishing, linking, changing your site without a person approving each step. That autonomy is the selling point. It's also the problem.

The three decisions I won't hand over

When I break SEO down, most of the value sits in three judgement calls. These are the ones I watch agents get wrong, and they're the ones I won't let a machine make unsupervised.

1. What to build

Deciding what page or piece of content to create next is a strategic call. It depends on what you already rank for, what your competitors own, what your business actually makes money from, and where the realistic gaps are. An agent doesn't weigh those. It sees a keyword with volume and writes a page. Do that on repeat and you get a site bloated with pages nobody needed - which, counterintuitively, can make your rankings worse rather than better. More is not the goal. Right is the goal.

2. How to link

Internal linking and anchor text - the specific words inside a link - are among the strongest signals you control on your own site. Done deliberately, they tell Google which page matters for which term. Done by an agent optimising for nothing in particular, they smear that signal across everything. A tool that decides your anchor text for you and won't let you override it is waving a flag I can't ignore. I've made the full case for why giving up control of your anchor text and internal links is a red flag, because it's the single clearest tell of a tool that will cost you later.

3. What will cannibalise

Cannibalisation is when two of your own pages compete for the same keyword, so Google can't decide which to rank and under-ranks both. Spotting it needs a view of your whole site and a judgement about intent - do these two pages really serve different searches, or are they fighting over the same one? I don't trust a tool to make that call at mass scale without oversight. An agent publishing at volume has no reliable sense of what it published last week, let alone whether it's about to compete with it.

The agent is a fast, confident junior. Fast and confident is exactly what makes an unsupervised mistake expensive.

Why 'mostly right' isn't good enough

People push back here: surely the agent gets it right most of the time? Probably. But SEO compounds, and so do mistakes. A wrong internal link or a cannibalising page doesn't announce itself. It sits there quietly diluting your signals while everything looks busy and productive. By the time the effect shows in your rankings, you've got weeks of decisions to reverse-engineer. 'Mostly right, invisibly wrong' is the worst combination for something you're not watching.

The other issue is confidence. An AI will hand you a cannibalising page or a nonsense link with exactly the same certainty it hands you a good one. It doesn't flag its own doubt because it doesn't have any. That's fine when a human is reviewing the output. It's dangerous when nobody is.

The manual override step is the product

So I don't run SEO with autonomous agents. I run it human-in-the-loop, and the non-negotiable piece is a manual override step - a point in every process where a person can catch and fix the thing the agent got wrong before it goes live.

• Let the AI do the mechanical work: drafting from a brief, gathering data, formatting, handling volume.
• Insert a human checkpoint before anything publishes or changes the live site.
• The human owns the three decisions above - what to build, how to link, what might cannibalise - every time, no exceptions.
• Treat the AI's confidence as meaningless. Verify, then ship.

This is the whole philosophy of human-in-the-loop SEO: the AI does the lifting, the human does the judging, and there's always a step where a dumb or hallucinated decision gets caught. It's slower than 'switch it on and walk away'. It's also the difference between compounding progress and compounding cleanup.

Where I might be wrong

I hold this with humility. Agents are getting better quickly, and I don't know what I don't know. The moment one can reliably judge cannibalisation across a whole site, link with genuine intent, and tell me when it's unsure, I'll happily hand more over and say so publicly. For now, the honest answer is that the autonomy is the bug, not the feature. Keep a human on the wheel. Let the machine do the pedalling.