How It Gets Fixed
In the previous post I showed what agents find on someone else's site. Now — what to do with it next. How that turns into fixes.
Fixing your own project is easier than fixing someone else's. Because you have access to your own.
Google Search Console as the source of truth
The first thing I do on my own project is connect Claude Code to Google Search Console.
It's done through the GSC API. You set up an OAuth token once — and the agent pulls everything from there itself: which queries are already finding you, at what positions, how many clicks, which pages are slipping, where the indexing errors are.
This isn't what the agent sees from the outside (like in the previous post). This is what Google itself thinks about your site, from the inside. Big difference.
Then the agent merges this data with what it found from the outside — the technical side, SEO audit, structure — and produces a fix plan.
How fixes land on the site
Next you need access to the server. Not your local machine — the server where the site actually runs.
You SSH in, start Claude Code there. And start pasting commands from the plan. The agent on the server sees the project's code and makes changes directly to it.
It's not one command. Not even ten. It's dozens of runs. The agent isn't perfect. It makes mistakes. Sometimes it breaks what was working. You go back, check, give it another iteration.
The pasting can be automated — build a pipeline where the agent drives itself. But that's a separate story, not for today.
There are things you can't do from the terminal. In GSC you have to log in, verify domain ownership, manually submit the sitemap, check reports. That happens in a browser.
Claude for Chrome
For the browser part there's a solution. Claude has a Chrome extension — Claude for Chrome. It's in beta now, access by waitlist.
It works like this: you go to Google Search Console, open the extension, give it a command. Claude sees the tab where you're logged in and walks through the forms itself. Submits pages for indexing. Sets up the sitemap. Reads error reports. Everything you'd do by hand — but you don't.
The same logic works on any complex form. Meta Ads, Google Ads, registering a business with the tax office, setting up email campaigns. It removes the manual work where it used to take hours.
⏳ The reality on timelines
This isn't a one-off.
You have to keep tightening it — run the same loop once a week. Checked GSC, looked at what slipped, handed it to the agent, got the fixes, applied them.
On my main project (I won't name it) I broke into top-3 / top-5 over two months on a number of queries where sites with years of history are sitting.
And that's only the start. Real SEO payoff comes on a 3-6-9 month horizon. Google doesn't rank fast. It looks at user behavior and at how stable your changes are over time.
When this doesn't work
There's a limitation worth mentioning.
One of the sites I looked at was on Tilda. Tilda is a website builder. There's no server you can plug an agent into. No files you can edit. Which means — no surface for the agent to act on.
In cases like that Claude Code produces recommendations — what to change in title, description, what schema to add through HTML blocks, what to fix in the structure. From there the owner goes into the admin panel and does it by hand.
In theory you can automate that too — through Claude for Chrome, walking the Tilda interface. But too much hassle for the result. This isn't where automation pays off.
Bottom line
The audit is half the work. The rest is a loop:
- GSC via API → fix plan
- Claude Code on the server → fixes in code
- Claude for Chrome → anything that requires a browser
- Repeat in a week
Site on Tilda? Recommendations only. Site on your own server? Full loop.
You'll need patience measured in months. But it's the only approach that actually works. Everything else is either an agency for money, or nothing at all.