Everyone has the same headline. I gave Claude my SEO and it replaced my agency. I read a dozen of those threads, got suspicious, and decided to actually try it. Not a slick demo on a throwaway site. My real site, my real Google Search Console, a full week, and one rule: I would do whatever Claude told me to do unless it was obviously going to hurt.
The site is contextbolt.com, which I run on my own. I connected Claude to live SEO data through an MCP server and plugged in my Search Console, so it could see real keyword numbers and my actual clicks and positions instead of guessing. Then I let it lead. Keyword research, audits, title rewrites, the lot. I took notes the whole time.
Here is the short version before the long one. Claude is a brilliant analyst and a shaky strategist. The week it spent reading, reasoning, and drafting was some of the best SEO help I have ever had. The moments it tried to decide what I should do were the moments I had to step in. The trick to a useful Claude SEO experiment is knowing which half is which.
- 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, over-optimizing into keyword stuffing, and trusting estimates like they were Google’s own numbers.
- The one thing that made it work was live data. Without an MCP connection and my Search Console, it just guesses confidently.
- The verdict: keep Claude on the analysis, keep the decisions for yourself, and you save real hours.
How I set up the experiment
The setup took about twenty minutes, and most of that was me reading my own notes.
I gave Claude two data sources. The first was live SEO estimates through 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 could read my real impressions, positions, and clicks rather than a third party’s model of them.
That second source matters more than people think. Estimates are guesses, even good ones. Search Console is the one place that knows the truth about your site. Putting both in front of Claude is what separates real SEO work from a confident chat. The full map of why is in my Claude SEO guide; this post is what happened when I stopped reading about it and ran it.
Then I set the rule. For one week, Claude drove. I asked it what to work on, and I did the work. I only overrode it when a call was clearly going to cost me. Those overrides turned out to be the most interesting part.
What Claude got right
This is the part the hype actually undersells. The analysis was genuinely excellent.
It found quick wins I had been ignoring for months. Day one, I asked it to read my Search Console and tell me where I was leaving clicks on the table. It came back with a list of pages sitting at the bottom of page one, queries with real impressions and almost no clicks, each one a title rewrite or a couple of internal links away from moving up. I knew that data was in there. I never sat down to dig it out. Claude did it in one prompt.
It clustered keywords the way I would, only faster. 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.
It read the live SERP and handed me the gap. This was the standout. I would 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 gap is a content brief written from the actual competition instead of a guess about it. I have paid for worse briefs.
It triaged a long list of keywords by difficulty in the time it takes to make tea. Forty headline ideas in, difficulty and volume out, scored against a small site, with a clear “skip this one and here’s why” 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 is a job built for a loop. A dashboard makes you check those one at a time. Claude ran the loop. If you want the honest way to read those difficulty scores, I wrote that up in checking keyword difficulty without Ahrefs.
The common thread is that all four are reading and reasoning over data. Give Claude the numbers and a clear question, and it is faster and more thorough than I am. By Wednesday I had done more useful research than I usually manage in a month.
What Claude got wrong
Then there were the decisions. This is where the agency-killer threads go quiet, so I will not.
It pushed me toward a keyword I had no business chasing. Mid-week it got excited about a term with great volume and built a whole plan around it. The difficulty score 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 the same numbers and says “not yet.” Claude read them and said “let’s go.” I overrode it.
It over-optimized the second I got lazy with a prompt. I asked it to “optimize this page for [keyword]” without any constraints, which is exactly how most people use it. It dutifully stuffed the keyword into the title, every heading, and the first line of half the paragraphs. The page read like a robot wrote it for another robot. Google’s spam policies are explicit that keyword stuffing is the kind of thing that gets a page demoted, not promoted. The lesson was not “Claude is bad at this.” It was “vague instructions get you keyword soup.” When I gave it real constraints, the rewrite was good.
It treated estimates like gospel. More than once Claude quoted a volume or difficulty number as if it were fact. Those numbers are third-party estimates. They are directionally useful, not Google’s actual ledger. My Search Console was the exception, because that data really is mine. Claude needed reminding which source was which, and left alone it would have built plans on numbers that were only roughly right.
It chased an AI-search tactic that does not work. It suggested adding special markup to “rank in AI results.” Google’s own documentation on its AI features is clear that there is no secret schema you bolt on to get picked up. The fix for AI visibility is the same boring stuff as always: be genuinely useful and easy to cite. Claude had absorbed a popular myth and repeated it with a straight face.
None of these are dealbreakers. Every one of them is fine the moment 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 one variable. Live data.
The first hour, before I had wired up the MCP connection properly, Claude was useless in the specific way the 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 stale and never contained live SEO metrics in the first place, so anything it says about a current SERP is a guess in a nice voice.
The minute it could fetch the numbers itself, it turned into a different tool. The same model that hallucinated difficulty scores now looked them up. The same model that guessed at the SERP now read it. Everything in the “what it got right” section above only happened because Claude had real data to reason over. Everything thin and hand-wavy happened when it did not.
That is the whole game, and it is why I keep saying the data layer is the product, not the model. This is also the deeper shift behind how AI agents are changing SEO. The agent is getting better at doing the work; the bottleneck is whether it can see the real world while it does it.
What a week with Claude actually changes
The honest outcome was not “I fired my SEO process.” It was “my SEO process got a very fast junior analyst who never gets bored.”
The research that used to be the slow, annoying part now takes minutes. Pulling the SERP, scoring a list, finding quick wins, writing a first-draft brief. That is most of the time I spend on SEO, and Claude ate through it. What it freed up was the part I am actually good at, which is deciding what is worth doing and judging whether the output is any good.
So the time saving was real and large. But the shape of the work changed more than the total hours. I went from doing the research and the deciding to checking the research and doing the deciding. That is a better trade for me. It might not be for everyone, and it is definitely not the autonomous robot the headlines promise.
Should you run your own Claude SEO experiment?
Yes, with one expectation set correctly. You are hiring an analyst, not a head of growth.
Give Claude live data, because without it the whole thing collapses into confident guessing. Then point it at the reading-heavy jobs first. Quick-win hunting in your Search Console, intent clustering, SERP gap analysis, difficulty triage, first-draft briefs. That is where it earns its keep on day one. The full workflow, including the prompts that only work with live data, is in my keyword research with Claude walkthrough.
Keep the decisions. Read its strategy advice as a strong first draft and then apply the judgment it does not have. Watch for the over-optimization reflex and give it constraints. And remember that estimates are estimates. Do that, and a week with Claude is one of the higher-leverage things you can do for a small site.
Running the experiment with ContextBolt SEO
Full disclosure, since you are reading this on the ContextBolt blog. The MCP server I used for the data half is one I build.
ContextBolt SEO is a hosted SEO MCP server made for exactly this. You subscribe, you get one MCP URL, and you paste it into Claude Desktop, Claude Code, Cursor, or Codex. From then on Claude can 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. It is $35 a month for 1,000 lookups at launch pricing, which is about a quarter of an Ahrefs plan.
Two things made it fit a week-long experiment specifically. It remembers every lookup across sessions, so when I checked the same keyword again on day five the answer led with what had changed since day one, at no extra credit. And it saves each finding to a ./seo-findings/ folder as markdown, so by Friday I had a written trail of the whole experiment sitting in my project. The free Google Search Console connection sits right alongside it, so in one chat Claude reads both the estimates and my real numbers.
If you would rather not pay for data, you still can run the experiment. Read the SERP by hand, lean on free checkers, and use your Search Console, which is the only paid-grade data that costs nothing. The method is the same either way.
The real result of my Claude SEO experiment was not a replaced agency. It was a clear division of labor. Claude does the reading, the math, and the first draft of the thinking, fast and without complaint. I do the deciding. Wire up the data, keep the judgment, and a week with Claude is genuinely worth running.