On Sunday evening, I gave an AI agent a job on my own website.
Not “write me ten SEO tips”. Not “tell me how to improve my marketing”. A proper assignment: read my real Search Console data, examine the website, identify something commercially useful, and prepare the changes.
One important condition: don't publish anything.
I've spent years auditing Google Ads accounts. I don't give a bidding recommendation a free pass because it looks clever. So I wasn't about to give an AI agent one either.
The question wasn't whether AI could produce the work. It was whether the work could stand up to inspection.One Sunday. One live business website. No production deployment.
First, it had to look at the evidence
I used ChatGPT Work to investigate Compulsion, with access to my connected Google Search Console data and the public website.
Search Console wasn't there to provide a trophy figure. It was one source of evidence. I wanted to know whether the website connected its existing proof of work to a clear commercial offer, not whether an agent could turn an early dataset into an impressive-looking growth story.
The agent compared the search queries and landing pages with the site's structure. Existing case studies were being discovered, but there was no dedicated page explaining an independent Google Ads audit or the PPC support I offer estate agents. That was a gap in the site's commercial journey worth investigating — not proof of demand or a forecast of enquiries.
It also reviewed the site's headings, canonical links, indexing controls and internal links. The technical evidence didn't justify rebuilding a functioning website.
Instead, it proposed two focused pages: an independent Google Ads audit and PPC for estate agents, with contextual links from relevant case studies and the account-governance article.
That's the sort of recommendation I can evaluate: a specific change, a reason grounded in the actual site, and a way to test whether it helps after publication.
Then the agent found the actual website files
Here's where it stopped feeling like a glorified writing assistant.
The agent located the existing website source and compared it with what was actually live. That mattered: I wanted it working from real files, not guessing at the design from a browser window.
Working from those files, it built the proposed pages inside the existing design system: the same typography, dark palette, navigation, contact routes and supporting scripts.
It produced a full staging archive, metadata, structured-data drafts, a change log and what it called QA evidence. Its automated link checks reported no broken local routes.
That's useful work. It reduces the gap between discovering an issue and having something concrete to review.
But there was still a difference between the agent saying the checks passed and me having evidence that they did.
The screenshots told a different story
The agent said it had checked desktop and mobile layouts. So I opened the screenshots.
One image labelled “desktop” was captured at a mobile-sized viewport. Another labelled for the estate-agency page showed the Google Ads audit page instead. A screenshot meant to demonstrate the contact route showed the homepage hero, not the form.
The pages might have rendered correctly. The code might have been fine. But those screenshots didn't establish either claim.
- Wrong viewport: “desktop” label, mobile-sized capture.
- Wrong page: estate-agency label, audit-page content.
- Wrong evidence: contact-route claim, homepage hero image.
None of these proves that the pages were broken. All three mean the submitted screenshots were inadequate proof.
I sent the findings straight back into Work and asked it to verify the actual URL, H1 heading and measured viewport dimensions for every replacement capture.
It acknowledged the discrepancies, withdrew the affected images as evidence and began correcting the handover. At the time of this draft, that correction was still awaiting final independent sign-off.
No dramatic AI meltdown. Just the most ordinary, important thing in professional work: someone checked the evidence and refused to approve what it didn't support.
Why this matters beyond web development
This is familiar territory in paid search.
A Google Ads report can show rising conversions without proving that more profitable customers came through the door. A higher ROAS can reflect a change in reporting, product mix or attribution. A campaign can improve after a bidding change without that change being the cause.
An AI agent can now write the code, run the test and prepare the evidence. That's powerful. It also creates another place where unsupported confidence can slip through.
Which is why I keep coming back to the same distinction: observed result, claimed cause and verified proof are three different things.
It's also why my earlier encounter with an AI that invented a childhood struck such a nerve. The problem wasn't just what it said. It was how naturally it said something it couldn't substantiate.
My rules for using an AI agent
- Start with the actual source. Real account data, real files, dates and scope. Not invented benchmarks or assumptions dressed up as findings.
- Make every important claim inspectable. A screenshot must show the right page at the stated viewport. A test result must relate to the code being approved.
- Separate preparation from permission. Finding a problem doesn't authorise an agent to change production, access another client's data or alter a budget.
- Keep a way back. Know what changed, preserve the original, and have a controlled deployment and rollback route.
- Measure afterwards. Shipping new pages is an intervention. Whether they improve the business remains a question for subsequent evidence.
And the website?
Still live, still under my control, and still running its existing deployment. The two new service pages remained in staging.
Crucially, that staging build carried a noindex instruction across the site. Sensible for a private test. Very much not something to upload over the production website by accident.
The final decision on service scope, case-study claims, pricing and publication stayed with me. As it should.
The bit I came away with
I didn't learn that AI can write website copy. We knew that already.
I learned that an agent could find the real source files, work with connected evidence, prepare meaningful changes and respond to an independent technical challenge — all within a single workflow.
It also showed that a polished handover can still contain bad evidence.
AI doing the work is a step forward. AI proving the work is the standard worth aiming for.
Not bad for a Sunday night.
Russell Anson runs Compulsion, an independent Google Ads and paid-search consultancy. This article documents a real AI-assisted staging exercise on Compulsion's own site. It is not a claim of SEO growth, increased enquiries or client performance.
Related: My AI invented a childhood. What about your ad budget? · Who still has the keys to your Google Ads account? · Selected work and case studies.