I gave Claude my real SEO for a week. The site is contextbolt.com, which we run, with our real Google Search Console connected and one rule: do whatever Claude says unless it’s obviously going to hurt. I’d read a dozen threads claiming Claude had replaced someone’s agency, and I wanted to see it on a site that has to earn its traffic.
Claude turned out to be a brilliant analyst and a shaky strategist. It did the reading, the math and the first draft of the thinking better than most help I’ve had. The moments it tried to decide what I should do were the moments I stepped in.
- Claude is a fast analyst, not a strategist. The week proved it does the reading and the math brilliantly and the deciding badly.
- What it got right: quick-win hunting in Search Console, intent clustering, SERP gap analysis, and difficulty triage across a long list, all in minutes.
- What it got wrong: pushing an unwinnable keyword, keyword stuffing on a loose prompt, and quoting estimates as if they were Google’s own numbers.
- The one thing that made it work was live data. Without an MCP connection to a live source like ContextBolt SEO, plus your Search Console, it guesses confidently.
What Claude got right
The analysis was excellent, and it’s the part the hype undersells.
- Quick wins I’d ignored for months: On day one I asked it to read our Search Console and tell me where I was leaving clicks on the table. It came back with pages sitting at the bottom of page one and queries with real impressions and almost no clicks, each a title rewrite or a couple of internal links from moving up. I knew that data was in there. I’d never sat down to dig it out. For the same job as a repeatable routine, see tracking keyword rankings with AI.
- Keyword clusters: I dumped a few hundred keyword ideas on it and asked for groups by search intent. What would have been an afternoon of dragging rows around a spreadsheet came back in under a minute, sorted, deduplicated, and labeled with the kind of page each cluster needed.
- The gap in the live SERP: I’d name a query, and Claude pulled the top ten results, summarized each one’s angle, and told me what every one of them missed. That’s a content brief written from the actual competition instead of a guess about it. I’ve paid for worse briefs.
- Difficulty triage: Forty headline ideas in, difficulty and volume out, scored against a small site, with a clear “skip this one” and a reason on each. Difficulty is mostly a count of how many sites link to the pages already ranking, which is how Ahrefs computes its own score, so it suits a loop. A keyword tool’s dashboard makes you check them one at a time. If you want to read those scores sensibly, I covered it in checking keyword difficulty without Ahrefs.
All four are reading and reasoning over data. Give Claude the numbers and a clear question, and it’s faster and more thorough than I am.
What Claude got wrong
The decisions are where the agency-killer threads go quiet.
- A keyword I had no business chasing: Mid-week it got excited about a term with great volume and built a plan around it. The difficulty was high, the top ten were all huge domains, and my site had no chance of cracking it this year. Claude knew all of that, because it had pulled the data itself, and it still recommended the play with total confidence. A good strategist reads those numbers and says “not yet.” I overrode it.
- Over-optimizing the moment I got lazy: I asked it to “optimize this page for [keyword]” with no constraints, which is a common way to prompt. It stuffed the keyword into the title, every heading and the first line of half the paragraphs. Google’s spam policies are explicit that keyword stuffing is the kind of thing that gets a page demoted. The lesson was that vague instructions get keyword soup. With real constraints, the rewrite was good.
- Estimates treated as gospel: More than once Claude quoted a volume or difficulty number as fact. Those are estimates, as in every keyword tool. Search Console was the exception, because that data is ours. Claude needed reminding which source was which.
- A myth about AI results: It suggested adding special markup to “rank in AI results.” Google’s documentation on its AI features says there’s no secret schema to bolt on. The fix for AI visibility is the same as always: be useful and easy to cite.
None of these is a dealbreaker once a human is in the loop. The danger is the framing that says you can hand the whole thing over and walk away.
The one thing that made it work
Strip everything else out and the experiment came down to live data. I gave Claude two sources. The first was live SEO estimates from ContextBolt SEO, delivered over the Model Context Protocol, the open standard Anthropic introduced for letting Claude call external tools mid-conversation. That gave it keyword volume, difficulty, SERP results and competitor data on demand. The second was my Google Search Console, so it read my real impressions, positions and clicks rather than a third party’s model of them.
Before the connection was working, Claude was useless in exactly the way skeptics warn about. I asked how hard a keyword was and got a confident paragraph with a made-up number in it. Its training data is months old and never held live SEO metrics, so anything it says about a current SERP is a guess in a nice voice.
Once it could fetch the numbers itself, it was a different tool. Everything in the “got right” list happened because Claude had real data to reason over, and everything thin happened when it didn’t. The wider shift is covered in how AI agents are changing SEO.
What a week with Claude changes
I didn’t fire my SEO process. I got a very fast junior analyst who never gets bored. The research that used to be the slow part, pulling the SERP, scoring a list, finding quick wins and drafting a brief, now takes minutes, and that freed me for deciding what to do next and judging whether the output is any good.
So the hours saved were real, but the shape of the work changed more. I went from doing the research and the deciding to checking the research and doing the deciding.
If you run your own, give Claude live data first, then point it at the reading-heavy jobs: quick-win hunting, intent clustering, SERP gaps, difficulty triage, first-draft briefs. Keep the decisions, read its strategy as a strong first draft, give it constraints so it doesn’t stuff keywords, and remember estimates are estimates. The full workflow, including the prompts that only work with live data, is in keyword research with Claude.
Running it with ContextBolt SEO
The MCP server I used for the data half is one we build. ContextBolt SEO is hosted: you start a 7-day free trial, get one MCP URL, and paste it into Claude Desktop, Claude Code, Cursor or Codex. Claude can then pull live keyword data, difficulty, SERP results and competitor info inside any conversation, with no DataForSEO account to register, no deposit and no credentials to manage. After the trial it’s $35 a month for 1,000 credits, which is about a quarter of an Ahrefs plan, and you can cancel any time. SEO tool pricing lists every major plan with its billing basis if you want to check that ratio. The free Google Search Console connection sits alongside it, so in one chat Claude reads both the estimates and your real numbers.
It also suits a week-long experiment. It remembers every lookup across sessions, so when I checked the same keyword on day five the answer led with what had changed since day one. And it keeps each finding, so by Friday the whole experiment showed up on my SEO Dashboard as a page I could scroll in one go.

Every lookup saves to the SEO Dashboard, so a week of research ends up in one place.
The audit tool shipped after this experiment. If I ran the week again, I’d say “audit my site,” take the fix-list worst first, tell Claude to “fix the top few,” and let it edit the pages and re-check in the same chat. That closes the loop the week left open, from finding the problem to fixing it.