Human-in-the-loop: where AI helps SEO and where it wrecks it

The honest answer to "can AI do SEO?" is: yes, the boring half of it, and only if a human is checking the work. That is what "human-in-the-loop" means - the AI does the mechanical lifting and a person makes every judgement call before anything goes live. Get that split right and AI makes you faster. Get it wrong and it quietly wrecks the site while looking productive.

I run my own work this way, so this is not a theory. I use AI for the absolute minimum amount of thinking possible, then I decide. Below is the map I actually use for what belongs where.

The split that matters: mechanical work vs judgement calls

SEO is two very different jobs wearing the same jacket. One job is mechanical - repetitive, rule-based, high-volume. The other is judgement - weighing trade-offs where the "right" answer depends on context only a human holds. AI is genuinely good at the first and confidently wrong at the second.

A useful way to think about it: AI is like a very fast, very literal junior who never gets tired and never pushes back. Fantastic for grunt work. Dangerous the moment the task needs someone to say "hang on, that does not make sense for this business."

AI does the mechanical lifting. A human makes the call. The moment you flip that, you are outsourcing your judgement to something that has none.

- Tom McSherry

Where AI genuinely helps

These are the tasks where I hand work to AI without much hesitation, because they are mechanical and easy to check:

• Clustering a big keyword list into topic groups so I can see the shape of a market fast.
• Drafting first-pass title tags and meta descriptions across dozens of pages, which I then edit.
• Summarising a competitor's page structure or a pile of Search Console data into something readable.
• Turning my rough notes and a real outline into a first draft I can rework.
• Spotting the obvious technical patterns - missing headings, thin pages, duplicate titles - across a whole site at once.

Notice the pattern: in every case the AI produces a draft or a list, and a human decides what is right. It compresses hours of copy-paste into minutes. That is the win, and it is a real one.

Where AI wrecks things

The damage happens when AI is trusted with the judgement half of the job - usually because someone bought a tool that promised to remove the human entirely.

Commercial judgement

AI will happily tell you a keyword with 5,000 monthly searches is more valuable than one with 90. It has no idea that the 90-search term is "emergency plumber Wellington" from someone with a burst pipe, and the 5,000-search term is students writing essays. Search volume is a weak signal for commercial value, and AI leans on it hard. That call has to be human.

Judging intent and cannibalisation

Ask AI to expand a keyword list and it will cheerfully hand you fifteen variations of the same intent, then suggest a separate page for each. Now you have pages competing with each other and a diluted site. More pages can mean worse rankings, not better, and AI does not feel that tension the way an experienced person does.

Facts, links and anything with consequences

AI hallucinates - it states things that are simply not true, with total confidence. It invents statistics, misremembers how a Google feature works, and drops internal links with weird robotic anchor text that reads as a red flag. Publish that unchecked across a hundred pages and you have scaled a mistake, not a strategy.

The manual override step is the whole product

The thing that separates useful AI SEO from dangerous AI SEO is not the model. It is whether there is a step where a human reads the output, catches the dumb thing, and overrides it before it ships. That override step is not overhead - it is the actual product. If you would like the behind-the-scenes version, I have written up how I actually use AI in my SEO process.

This is also why I am wary of set-and-forget AI SEO tools that promise to run your whole strategy while you sleep. They are selling the removal of the exact step that keeps the work sane. When the AI does something daft at 2am - and it will - there is nobody in the loop to stop it.

A simple rule you can use

Before you let AI touch a task, ask one question: if this is wrong, does a human catch it before it goes live? If yes, hand it over and enjoy the speed. If no, keep it in human hands, or add a review step until the answer becomes yes.

That is the entire philosophy. AI for the mechanical grind, a person for every judgement call, and a hard override step in between. I hold this view strongly, and I would be happy to be proven wrong - but every time I have watched someone remove the human to save time, the site has paid for it later.

For what this philosophy looks like applied end to end across a website's content rather than stated as a principle, see using AI for your website content without wrecking it.