How AI helps us turn big SEO data into tailored plans for our clients

How we use AI to read big SEO data exports, find what matters for each Webflow and Shopify client and act on it, with every figure checked.

We recently audited our own website, in particular the page titles and meta descriptions. We ran Screaming Frog, which crawled 611 URLs (every page, image, script and file it could find) and handed back the usual pile of spreadsheets.

Working through that would normally take a LOT of time. But now we can use AI to help, in our case, Claude.

Claude read the exports and we quickly found 13 pages still using an old title format with a pipe separator (we like a consistent format, even if that doesn't really impact SEO, it looks much better), and a footer link that pointed to a site we used to support that had recently shut down.

It then made the changes in Webflow: 19 page titles, including a few the crawl couldn't reach, and the footer link on every page. We checked each one and published the same day.

That's one of the uses where AI helps us in our SEO work. As a small agency we can act on data at a scale that would traditionally require a much bigger team, delivering a report, plan and action list tailored to individual clients.

What AI actually does in our SEO work

In short, it reads. A lot. SEO throws up masses of data: crawl exports, a year of Google Search Console figures, thousands of keyword suggestions, GA4 reports. Most of it lands in CSV files with tens of thousands of rows, and the useful part is usually a handful of lines.

Claude reads the whole lot, checks and cross references each source against the others and tells us what stands out.

From this we can decide what to do about it. A lot of the time it can be instructed to make the change too. For example, in Webflow or Shopify, through a direct connection, we could ask Claude to implement the 50 improved meta descriptions it helped find, once we'd reviewed, edited and approved them.

It doesn't replace knowing what good SEO looks like. It means we spend our time on judgement rather than spreadsheets. (I've done enough VLOOKUPs for one lifetime.)

The tools we use

Our setup isn't exotic. Most of it is tools we already used, connected to Claude through Model Context Protocol (MCP), a standard way for an AI to talk to other software, and APIs.

  • Claude, on Anthropic's Team plan
  • Claude Code, which I mostly use from the command line (CLI). Most of what's in this post can also be done with Claude Code in the Claude app, but a good working knowledge of the CLI and version control like GitHub is a huge help
  • SE Ranking for keyword research and rank tracking (amongst other stuff). Its MCP server comes with paid plans, so if you want to look into this (I highly recommend it) then check whether your plan includes it. Ahrefs and Semrush provide similar setups
  • Google Analytics' own open-source MCP server on GitHub, for GA4 data. It's free, but it isn't a quick setup (definitely needs you on the command line, and a Google Cloud console account with a project to set up OAuth. We use our Google Workspace Business account)
  • Screaming Frog for crawling, and exports from Google Search Console. Again, this requires the full paid version, but its recent MCP and API connectors make this software an absolute must
  • Direct connections to Webflow and Shopify

We don't keep every MCP or direct connection switched on. Claude Code lets us turn off the ones a job doesn't need, which saves tokens and keeps each session focused on the client in front of us.

Saving key findings in a Claude Artifact like a Claude Doc means work can start in one place and carry on in another. I might dig into a client's data in the command line in the morning, then later on, pick up the same working document in a Claude project in the app. Both see the same doc, so nothing gets lost between the two. (This post was written exactly that way.) Our project in GitHub knows all about this post too. Key points and progress are saved to versioned MD files, which helps us know which post to start working on next, in the knowledge of this one.

Webflow: turning visitors into enquiries

Most of our Webflow sites are for service businesses, ours included. The goal is enquiries, so the questions we ask of the data are about the pages people land on and whether they get in touch.

For those sites we use AI to:

  • check titles, descriptions and headings across the whole site, then, once we've reviewed and edited them, fix them in Webflow
  • find the searches each service page could realistically win, local ones included
  • see which pages lead to a contact form and which don't
  • check how the business is described in AI answers, such as Google's AI Overviews

That last one is newer. Being named in an AI answer is fast becoming part of being found, and it's built on the same foundations as good SEO. I've written about that in Good SEO is the foundation of AI SEO.

Shopify: working at scale

Shopify is a different job. A store's value is in its products and collections, and there can be thousands of each. The questions are about which ones get found, which ones sell and what's holding the rest back.

One of our clients has over 18,000 products, with more added every month. We pulled every product, collection, page and redirect from Shopify, crawled every product page and analysed that against a year's worth of data from Google Search Console. A lot of very large CSV files… In this particular case we found every product page came back live and indexable, which was good news. Of 1,331 existing 301 redirects, 118 needed attention: chains (a redirect pointing to another redirect) and leftovers from duplicated or removed products.

Nobody is going to read 18,000 product pages by hand. Even viewing a CSV with 20,000 rows and maybe 50 columns is a challenge. With AI we can read all of them, spot the patterns and plan to fix things in batches, while still looking at each problem on its own merits. In this particular instance we created a plan of actions as a Claude artifact, that we could also download (and if we'd have needed, share the Claude copy with the client).

GA4 adds the behaviour side: which products get looked at, which get added to the basket and where people drop out. For an ecommerce store, that's where SEO and sales meet. For a smaller example of AI on a Shopify store, see Using AI to write Shopify page titles and meta descriptions.

Making sense of thousands of rows

Keyword research is a good example. A tool like SE Ranking will happily give you thousands of related searches, and most of them don't matter to a particular business.

For one recent client, we started with SE Ranking's UK keyword data and ended with a shortlist of 33 keywords, each mapped to the page that should rank for it. Along the way, Claude grouped the searches by meaning and flagged where the volume came from the wrong audience (homeowners, in that case, not businesses). It also left out national terms that property portals and big agents already own.

Search Console works the same way. A year's worth of data spread across 6 or 7 CSVs is a lot, but the questions are simple: which pages are close to page one, which have dropped and which get seen but not clicked. AI answers those quickly, and it'll show its working if you ask. Importantly, it feeds your marketing strategy with useful information.

That's the part I value most. The output is shaped around our clients (their site, their market, what they actually sell), not a generic report with a logo.

Every figure checked before it reaches you

Speed is only useful if it's right, so we always do real-world comparisons: checking a sensible cross-section of what AI produces. In fact we've got our Claude instructions set to make sure we do just that.

A good example is when we first ran a rank check for the client just mentioned, the tracker put their brand name in second place. A live search the same day showed them first. Rank trackers take a snapshot from one place at one time, so a quick live check is part of the process before any figure goes in a report.

The same goes for every change. When Claude fixed our own page titles, each one was checked on the live site after publishing. The AI does the legwork. The judgement and the sign-off must stay with us.

Client data stays where it belongs

We work with our clients' data, and that often includes their own customers' data (Shopify sales, for example), so it's important we handle it carefully.

We connect Claude to the platforms we work in, such as Shopify, through authorised connections. Sensitive data stays inside those Claude sessions. We don't download exports to email around, and we don't store them in our code repositories. Those are set up to ignore such data files, with automated checks that stop the wrong things being saved.

We use Claude on Anthropic's Team plan, which comes under their Commercial Terms. Those terms say Anthropic may not train its models on what customers put in.

Why this matters now

This is tech, and tech moves fast. AI is moving faster than most of it at the minute, and we strongly believe, love it or loathe it, that agencies that don't keep up will find it hard to catch up. Their clients may well notice before they do.

We're still a small team, and AI hasn't changed that. What it has changed is how much we can see and do for each client in the time we have. For us, that's the point. Not more automated SEO, but better, more individual SEO, with people still making the calls.

How much of your website's data is sitting in spreadsheets nobody has had time to read, or make the most of?

Did Claude write this post? No. Well, it definitely helped. It gave me structure, corrected my speeling and held my hand. I have a lot to say, just need the help getting my words across.

Want your SEO data working harder for your business? Let's talk.