I use AI every single day. It’s woven into almost everything I do — research, content, technical audits, feed work, campaign analysis. I’m not here to tell you AI is overhyped or that you should be cautious about it. Quite the opposite.

But I am going to be straight with you about something that’s been doing the rounds lately — this idea that you can set up an AI agent, point it at your SEO or your Merchant Centre feed, and walk away. That it’ll just… handle it.

It won’t. And I’ve got the battle scars to prove it.

The pitch sounds brilliant. The reality is messier.

The promise is “self-driving SEO” — a bot that crawls your site, spots the opportunities, makes the changes, and improves your rankings. All for a fraction of what you’d pay a specialist.

On paper? Compelling. In practice? Agents misinterpret context, apply changes blindly, and create errors at scale. Businesses end up with broken pages, generic content, and falling visibility — spending more time cleaning up the mess than they saved on the tool.

And that’s just SEO. When you bring Google Ads and Merchant Centre into the picture, the stakes get considerably higher.

Let me tell you what actually happened with my Merchant Centre feed

I was working on supplementary feed spreadsheets — trying to automate product data updates to get products better positioned for AI Overviews and shopping visibility. The logic was sound. The execution, when I let the agent run it without tight oversight, was a disaster.

Wrong data in wrong fields. Product attributes mapped incorrectly. Titles pulling from description columns. Availability statuses that made no sense. The kind of errors that don’t just sit quietly — they trigger disapprovals, tank visibility, and in a worst case, flag your account.

Google Merchant Centre has zero tolerance for certain errors. Price mismatches, invalid GTINs, conflicting brand data — these aren’t warnings you can sit on. They cause immediate product disapproval, and if they pile up, you’re looking at account-level suspension and an appeals process that takes time you simply don’t have.

Here’s what changed everything: when I worked with the AI — me driving the strategy, checking the field mapping, validating the outputs before they went anywhere near a live feed — it worked brilliantly. Fast, accurate, genuinely useful. That’s the version of this that actually delivers results.

That’s not a compromise. That’s the model.

Why agents fail when left alone

It’s not that the technology is bad. It’s that it’s being deployed without the human layer it actually needs to function properly.

A Gartner survey found that only 15% of IT leaders are even considering fully autonomous AI agents — and the ones scaling AI successfully are not pursuing full autonomy. They’re building systems where agents handle the heavy, repeatable work and humans guide the nuanced decisions.

Without that human layer:

  • Agents don’t understand your business goals. They chase volume, not conversions. They’ll optimise for traffic signals without understanding what actually matters to your clients or your revenue.
  • They make technical decisions without understanding tradeoffs. Should you consolidate pages? Restructure navigation? Rewrite a product title? Agents oversimplify these calls — and one misapplied canonical tag can break dozens of pages. One schema error can wipe out rich snippets across an entire domain.
  • Speed becomes a liability. Agents can make thousands of changes in days. That’s the selling point. It’s also the risk. When something goes wrong at that pace, you’re not dealing with one bad title — you’re dealing with a catalogue of problems that all happened while you weren’t looking.
  • They can’t read brand voice. They produce generic copy. They hallucinate. They’ll create content that’s factually wrong or completely off-brand, and it goes live unless someone catches it.

And in Google Ads, the damage can be immediate

The same principle applies to paid. AI recommendations inside Google Ads are not neutral — they are always in Google’s interest. Not always yours.

Some of the most damaging things I see in account audits:

What went wrongWhat it actually caused
Auto-apply settings left onGoogle changed bid strategy without permission
Broad match + Max biddingBudget spent chasing volume, not conversions
Broken conversion trackingEvery optimisation decision built on bad data
PMax cannibalising SearchLower CTR, lower conversion rates, higher spend
Old negative keyword lists never reviewedBlocking your own best keywords from showing

These aren’t edge cases. They’re patterns I see regularly — and they’re exactly the kind of thing an AI agent will either miss entirely or actively make worse if it’s been given too much autonomy over account settings.

AIO is even less ready for automation

AI Optimisation — getting your content cited in AI Overviews, showing up in LLM responses, being the source that gets pulled — is still an emerging discipline. The signals are different from traditional SEO. The rules are still being written.

Automating something we don’t fully understand ourselves yet is a fast route to wasted effort and potentially harmful content patterns. The businesses I see doing this well are doing it carefully, with a human making the calls on what gets published and why.

So what does AI actually do well?

✅ Works well with human oversight❌ Not ready to run alone
Technical audits & crawl analysisStrategic planning & prioritisation
Keyword clustering & gap analysisBrand voice & editorial judgement
Feed data structuring (with validation)Merchant Centre field mapping unsupervised
Metadata drafts for reviewGoogle Ads bid strategy decisions
Schema markup suggestionsConversion tracking setup
Competitor research & reportingAccount-level settings changes
Content briefs & outlinesLink building outreach

The businesses winning with AI right now are not the ones who handed over the keys. They’re the ones using it to go faster — with a human still in the driving seat.

Even Anthropic agrees — and they just proved it

As I was finishing this post, Anthropic released two major model updates: Claude Opus 4.6 (5 February 2026) and Claude Sonnet 4.6 (17 February 2026). Along with them came a new product called Claude Cowork — where Claude works autonomously alongside you on tasks in the background.

Here’s what struck me. This is Anthropic — the company building some of the most capable AI in the world right now. And their own framing for their most powerful agentic feature is not “let it run.” It’s co-work. Working with you. Not instead of you.

Opus 4.6 can run financial analyses, do research, build spreadsheets and presentations — but inside a framework explicitly designed around human collaboration. Even their agent teams feature in Claude Code is built so that humans direct the work, not abdicate it.

That’s not a coincidence. That’s the people who built these models telling you, plainly, how they’re meant to be used.

The most capable AI tools available right now are designed to amplify a skilled human — not replace one. If the people building the technology aren’t betting on full autonomy, you probably shouldn’t be either.

What this actually looks like in practice

When my Merchant Centre feed work went wrong, it wasn’t the AI’s fault. It was mine — for expecting it to manage something that needed my expertise to validate. When I stepped back in, we fixed it together in a fraction of the time it would have taken me alone.

That’s the real value of AI in SEO and AIO right now. Not replacement. Not automation for its own sake. A genuinely capable collaborator that needs direction, context, and someone who knows what “good” looks like.

“The future of AI is not about replacing humans. It is about empowering them.” — Rajeev Butani, CEO of MediaMint

I’d go further: the future of SEO and AIO is human expertise, moved faster by AI. That’s what I build for my clients. And it’s why I’m still very much in the room.

Sources