Every week I used to open Google Search Console and Plausible, our website analytics, and try to work out what had changed and why. Answering a single question that way could take an hour or so.
Now I ask Claude.
What asking looks like
I ask questions in plain English: "How is the site doing?" "Is the Cherwell page catching anything?" "How many visitors came from ChatGPT this month?" Claude pulls the live figures from Search Console and Plausible, explains them, and tells me what it thinks they mean.
A question that used to take me an hour or so now takes a few minutes. I also read them alongside our marketing material and plans, rather than on their own.
It works through a small connection we built using MCP (Model Context Protocol), the open standard that lets an AI assistant use other tools. The connection gives Claude a handful of ways to read our data: clicks and impressions from search, the queries behind a single page, a before-and-after comparison, visitor numbers, enquiries, and where visitors came from.
Read-only, on purpose
The connection can read our data and nothing else. Its access to Search Console is the restricted, read-only level, and the Plausible key can only read statistics. If Claude gets something wrong, the worst it can do is give me a wrong answer. That's why we checked its answers before trusting them.
Checking it before trusting it
Before anyone relied on its numbers, we checked them against a Search Console export we'd taken by hand in August. Every day we spot-checked matched, and the total over the period matched to the click.
We learned something about Search Console along the way. Its report by search query doesn't add up to your total clicks, because Google leaves out rare and personal searches. For us, the queries it listed accounted for well under half of the month's clicks. If your query report and your totals have never agreed, that's why.
The week-one bug
In its first week, it got a comparison wrong. I asked how our pages had done before and after a change to the site's navigation, and it told me one of our blog posts had dropped out of search results. It hadn't.
The comparison only looked at the top couple of hundred pages in each period. Most of our pages had no clicks at all, so they all tied, and which of them made the cut was effectively random. A page could vanish from one period without anything having changed.
It was fixed within days. The comparison now reads every page, and there are tests for that mistake. If you work with these reports yourself, the same trap applies: a top-200 list isn't the whole picture.
From answers to changes
Once a week we also go through the numbers together: pages and searches that moved, pages that show up in search but don't get clicked, and searches a page almost answers. Claude proposes changes, each with its evidence. I approve or turn down each one, and Claude Code makes the approved changes to the site.
Every change gets a dated entry in a log, and we change one thing on a page at a time. Two or three weeks later we compare before and after, so we can tell whether it helped.
What hasn't changed
I still make the decisions. The difference is that I no longer spend an hour finding the numbers before I can make one.
Trying to do something similar? Talk to us. If you run a service desk, the same idea works there too: see Claude working with Freshservice.