I've been doing technical SEO for over 15 years, and this is the biggest change to how I work that I've seen. The AI doesn't just suggest what to do anymore. It pulls the data, writes the script, builds the report and hands me something I can ship.

The trick is giving it real data. On its own, an AI knows SEO but can't see a single live SERP or search volume. Connect it to DataForSEO and a crawl of your site, and it can do most of what you'd log into Semrush for, at a fraction of the cost.

What does this look like in practice?

Here are two jobs I ran on October 15, 2026, with the exact prompt, what came back, how long it took and what the API charged.

What is DataForSEO?

It's the raw SEO data, sold by the call, with no subscription.

DataForSEO sells the same kinds of data the big SEO platforms show you (SERPs, search volumes, rankings, backlinks) through an API. You pay as you go: a $50 minimum deposit, no monthly fee, and the balance never expires. There's no dashboard to learn. Claude is the dashboard.

Can Claude + DataForSEO really replace Semrush?

For most day-to-day jobs, yes, and the data for each one costs cents.

Here's the same work done both ways. The data costs are what DataForSEO charges at its published rates; Claude's work runs inside your subscription.

What you give up is Semrush's polished interface, its clickstream traffic estimates, its deep history and set-and-forget projects. What you gain is an assistant that doesn't just show you the data but writes the report, the fix or the page from it.

Which AI should you use?

Most of the big ones can do this. I recommend Claude.

ChatGPT's Codex, Gemini and others can all call an API and work through a crawl export. I use Claude, and the rest of this article focuses only on it. Three reasons:

16 jobs to hand Claude this week

What can it actually do for me?

These come from my own client and site work and from what other SEOs and the AI labs are publishing. Each card has a prompt you can paste in and the data it needs. Where I've priced the job, the dark pill shows what Claude's tokens cost at API rates, from my test of 8 SEO jobs on every Claude model.

How much time will it actually save?

Research and reporting that took most of a day now takes minutes.

The savings come from the pulling, sorting and formatting, the part of SEO that used to fill whole afternoons. In Keyword.com's 2026 survey, 89% of SEOs using AI save at least 4 hours a week, and a third save more than 10. How much you get back depends on how much of your week is that kind of work.

Which site crawler do you need?

Any real crawler will do. Claude alone only works for tiny sites.

Screaming Frog is free up to 500 URLs and $279 a year after that. Sitebulb starts at $18 a month. Either one catches what an AI reading page by page will miss: redirect chains, orphan pages, duplicate content and canonical conflicts across the whole site.

Claude can crawl a small site itself, but I wouldn't go past about 10 pages, and it won't be as thorough as a real crawler. Treat it as a spot check.

How do I connect Claude to DataForSEO?

Two ways. Both take about ten minutes.

1. The MCP server. Add DataForSEO's hosted endpoint, https://mcp.dataforseo.com/v3/mcp, to Claude and authorize once in the browser. A local option runs with npx dataforseo-mcp-server@latest.
2. A skill file. This is what I use. My API login sits in a .env file on my Mac, and a short skill tells Claude where it lives, which endpoints to use, to check my balance first and to report the cost of every job.

Then give it a real job. This is the prompt that turns a crawl and a few cents of API calls into a Semrush-style audit:

Here's my Sitebulb export (internal_all.csv). Pull ranked keywords and
search volume for every indexable URL from DataForSEO, then build one report:
technical issues ranked by traffic at risk, pages in positions 8 to 20 worth
pushing, and the 10 fixes to ship first. Report the API cost.

Five rules that keep it accurate and cheap

1. Ask for the deliverable, not advice. "Build a keyword gap CSV and a one-page summary" beats "how should I find keyword gaps?"
2. Keep keys out of the chat. Credentials live in a local file that Claude loads per call and never prints.
3. Make it show its work. Raw numbers in a table, sources linked, API spend reported. If it can't cite it, it doesn't ship.
4. Trim what you feed it. Cutting a Screaming Frog export from about 70 columns to 15 saves 52% of the tokens. More in my Claude cost breakdown.
5. Save what worked. Every good run goes into the skill file, so next month the whole job is one line.