Black SEO Analyzer Now Runs Inside Claude and ChatGPT

Until November 10 my own homepage said an AI assistant “has never loaded your pages.” That was wrong, and I changed it. Claude and ChatGPT can both fetch a page and read it. Ask one to fix your SEO and it will open four or five pages and hand back advice that is correct and would fit almost any site.

That mistake is worth a minute, because it points at the real limit. The model can read. What it can’t do is read your whole site, check every page the same way, and count. That part is a crawler’s job, and now the two can work together.

What shipped

Black SEO Analyzer is now a connector for Claude and ChatGPT. You add it from the assistant’s directory, sign in with your email, and ask for what you want in plain words: “Audit example.com.” “Check this page.” “What changed since the last crawl?”

The crawl runs on our servers. bseoa fetches the pages, runs 300+ checks across 16 modules, and hands the assistant a list of issues. Each one has a severity, the URL it was found on and a hint for fixing it. The assistant can check a single page, crawl up to 2,000 pages, summarize the issues, pull the full list for any one of them, report on one URL, preview a title and description as a search result, and compare two crawls. The connector page has the setup steps and the reference lists every tool.

Nothing to install. Results stay with your account, so you can come back next week and ask what changed.

Why I built it as a connector

I use LLMs for more of my work every month. Somewhere along the way I stopped wanting a separate app for each job. I want the tools to show up where I’m already working.

That’s what MCP is for. The Model Context Protocol is a standard Anthropic published in November 2024, and OpenAI adopted it in 2025. A program uses it to tell a model what it can do: here are my tools, here is what each one takes, here is what comes back. The model decides when to call one. That second half is tool calling, and it’s the piece that turned chat assistants from things that talk into things that can do work.

The split suits both sides. A model is good at reading a pile of results, deciding what matters first, explaining it to a person and writing the fix. It is bad at being a crawler. It can’t hold 264 pages in its head, it won’t apply the same rule to page 3 and page 203, and if you ask twice you may get two different counts. A crawler is boring and exact. It does the same thing every time, which is the whole point of having one.

The speed part

Here is a real run. On November 9 I asked Claude to audit sethserver.com, my own site, through the hosted connector. The crawl covered 264 pages in 2 minutes and 18 seconds. Then Claude told me where to start:

• 171 meta descriptions are too long and will be cut off in results
• 26 pages share a title with another page
• 16 pages load a CDN script with no integrity check
• 1 internal link is broken, and it named the post

I sell an SEO tool, so make of that list what you will.

That 2:18 is slower than bseoa can go. The hosted crawler holds itself to two pages a second per site so it doesn’t lean on anybody’s server, and 264 pages at that pace comes out to about what I got. The checks aren’t the slow part. Being polite is. If you run the desktop app or the CLI on your own machine against your own server, you set the concurrency yourself, and the default is 20 requests at once.

The other half of the speed is the model. In Claude Code, Cursor and Codex, bseoa runs as a local MCP server next to your repository. The assistant reads the issue list, opens the template that produces those 171 descriptions, changes it, and crawls again. The compare tool then says what got fixed, what is new and what is still there.

The homepage fix at the top of this post went that way. I described the problem to Claude Code in one sentence. It rewrote the copy in English and Spanish, measured the figures it quoted, and opened a pull request. I read it, changed one sentence and merged it.

One caveat so I don’t repeat my homepage mistake in the other direction. In a Claude or ChatGPT chat window, the assistant can’t touch your site. It tells you what to change and you change it. The edit-and-recrawl loop needs an assistant that can reach your code.

Where this is going

For years a technical audit has been a project. Somebody runs the crawler, exports a spreadsheet, writes it up, and the fixes land whenever they land. By then the site has changed again.

When the crawler is a tool the model can call, an audit is a question. You ask it after a deploy. You ask it before a launch. You ask it about one page because something looks off. The team that finds a broken canonical the day it ships is ahead of the team that finds it in next quarter’s report, and they got there without working any harder.

I think most software ends up here. The programs a model can call will get used, and the ones that need a person to click through them will get used less. I would rather build for that now than port to it later, and if your work is keeping sites healthy I think the same is true for you. Learn to work this way while it’s still early and a little rough.

And it is a little rough. The hosted crawler doesn’t render JavaScript, so a single-page app still needs the desktop version. Models still get things wrong, which is the reason to hand them measured results in the first place.

Try it

The hosted connector is free for 14 days with no card, then $29 a month for 25,000 crawled pages every rolling 30 days. If you want something to practice on, demo.blackseoanalyzer.com is a made-up bakery site I keep broken on purpose. Add the connector to Claude or ChatGPT and ask it to audit the bakery.

-Sethers