A massive audit written entirely by AI can look impressive. It's a lot less impressive if nobody reads it closely enough to know whether it's right.
TL;DR: Chris Green argues that AI should assist SEOs, not do their reviews for them. Let software check the facts, let AI make the results easier to read, and keep the judgment calls for yourself.
🧠 The Trouble Starts When AI Is Asked to Judge
Chris Green, who writes about search marketing, published a piece on October 6 about where AI fits in SEO work. His answer is that AI should assist the person doing the review, not take it over. He's building a Chrome extension around that idea, combining standard SEO checks, browser data, and a language model.
The trap he describes is easy to fall into. You hand raw data to a model and ask it to tell you what to think. You get a confident answer, but you skipped the thinking that should have produced it.
A giant AI-written audit is the extreme version of this. It sounds thorough, but if no one reads it closely, it's worth very little. The problem isn't the output, it's the thinking you stop doing.
🔍 Software Finds the Facts. AI Makes Them Readable.
Many technical SEO checks have a yes-or-no answer. Green points to a few:
Did a URL return a 404?
Does a canonical tag exist?
Does robots.txt allow crawling a path?
Did a link destination change between the server HTML and the rendered page?
Canonical tags tell Google which version of a page should be treated as the primary one. Server HTML is what a site sends first, and the rendered page is what the browser builds after the scripts run. Those two can disagree, which is why that last check matters.
None of these need a language model. Plain software answers them faster, gives the same answer every time, and uses fewer resources. Starting from those facts gives you firmer ground than starting from a model's best guess.
I covered a few of these same checks, including soft 404s and canonical mix-ups, in a recent issue on static sites.
Once the facts are right, a language model earns its keep. Raw JSON or a giant spreadsheet is unpleasant to read, and AI can turn it into something short and usable. Green lists summarizing changes, drafting Jira ticket descriptions, rephrasing findings for clients or coworkers, and flagging shifts in meaning that aren't obvious.
Picture the pile of data behind one page: URL, status code, canonical, robots rules, anchor text changes, and more. An AI can boil that down to a sentence saying the link changes after rendering, but the server HTML already has a working URL, both versions land in the same place, and the anchor text is the same. You read one line and decide whether it matters.
🧪 Green's Own Test Showed Where the Model Breaks Down
Green tested Gemini Nano, a model that runs on your own device, on the differences between raw and rendered HTML. It did well at describing wording changes and what they might cause, like anchor text going from descriptive to "Learn more" and losing context.
It got much less reliable when asked to combine several structured technical facts into a final SEO judgment. It got confused and made up reasons. So Green changed the job: the model now helps explain the findings and the evidence instead of ruling on them.
There's a bonus to keeping a person in the loop. When the model disagrees with you or gets something wrong, that's a chance to look closer. Green found those misses pointed to noisy evidence, fuzzy terminology, weak logic, or his own bias, all of which a fully automated setup would have hidden.
🛠️ Split Your Own Review Into Three Jobs
Green's closing advice is a simple split, and it works as a checklist for your own process.
Let Software Check the Facts
Run crawlers, scripts, or browser tools for the yes-or-no items first: status codes, canonicals, robots rules. Do this before you read anything an AI wrote about the page.
Use AI to Cut the Reading, Not the Thinking
Ask for a short summary, a ticket draft, or a plain-English version of findings you've already confirmed. Then check the evidence behind it, especially when the summary sounds sure of itself.
Keep the Judgment Call for Yourself
Deciding whether a finding matters, and what to recommend, is your job. When the AI disagrees with you, find out why before you dismiss either side.
Hit reply
Where does AI save you the most time in an SEO review right now, and where has it handed you a confident wrong answer?
One thing I'm thinking about is how often "AI does the work" gets treated as the finish line. Green's version is less exciting than an AI-generated audit, but it keeps you close to the problem. If you only ever read summaries, you slowly lose the feel for why a page is broken.
Small gains count too. A tool doesn't have to make a single decision to be worth using, as long as it makes your evidence clearer and your repetitive work more consistent.
I'd rather be the person who can explain the audit than the person who forwarded it.
Till next time,
