CrazySEO README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the CrazySEO listing page.
CrazySEO
Find out whether ChatGPT, Gemini and Perplexity recommend your site — and why they don't.
An open-source GEO / AEO auditor and MCP server. No account. No card. Your keys.
Your customers have stopped googling. They ask an assistant "best CRM for small law firms?" and buy whatever it names. If it doesn't name you, you don't exist — and no analytics tool will tell you, because there is no click to measure.
This audits it directly: it asks the questions your buyers ask, records the answers word for word, and shows you which sources the models cite instead of you.
Why another one
Most tools in this space ask the model "what do you think of Acme?" — which tells you nothing, because naming the brand in the prompt guarantees it shows up in the answer.
This asks blind. Your brand never appears in the question, so the answer you get back is the answer a real buyer gets. That is the whole difference between a number you can act on and a number that flatters you.
Three more things a prompt wrapper can't do:
- Verbatim receipts. Every result carries the exact prompt and the model's raw answer, so you can paste it into ChatGPT yourself and reproduce it.
- Honest coverage. If an engine fails or rate-limits, the report says so. It never quietly drops an engine and calls the remaining average your score. If a site can't be reached, it refuses to score it rather than inventing a number.
- A real readiness audit. 11 checks against your actual site — not a prompt.
Quick start — no API keys needed
The readiness audit runs against any site with zero configuration and zero cost:
Nothing to install, no signup, no key. Or run it from source:
(That's real output — stripe.com, unmodified.)
11 scored checks, all first-party fetches: HTTPS · AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended…) unblocked in robots.txt · schema.org markup · title + meta · readable without JavaScript · canonical · Open Graph · single-H1 structure · image alt-text coverage · sitemap referenced · broken internal links (sampled).
Plus llms.txt, shown but deliberately unscored: Google's own AI-optimization guidance says it has no effect in Google Search, so it can't fairly cost anyone points — but some agents do read it, so it's still worth seeing.
Exits non-zero when checks fail, so you can gate a deploy on it:
Machine-readable output for pipelines:
The visibility audit — bring your own keys
To measure what the engines actually say, add whichever keys you have. Any one is enough; more engines means more coverage.
It reads your site, infers what you sell, generates the questions your buyers would actually ask, and asks all of them — blind — across every engine you configured.
(Real output, trimmed for length. Note the last line: when an engine fails, the report says so instead of quietly averaging the rest and calling it your score.)
Runs on your own keys — a full audit is typically a few cents.
Use it from your agent
It ships as an MCP server, so Claude, ChatGPT or any MCP-capable agent can run it directly:
It's also packaged as a Claude Code plugin, which wires up the MCP server and the skill below in one step:
Then just ask:
"Is my site visible in AI search? What's cited instead of me?"
| tool | what it does | keys |
|---|---|---|
readiness_check | 11-point AI-readiness audit | none |
audit_plan | site facts + buyer questions to ask, for a keyless audit | none |
audit_score | scores what the agent found: visibility, leaderboard, citation gap | none |
audit_visibility | blind multi-engine visibility + verbatim answers | BYO |
citation_gap | sources cited on the questions you lose | BYO |
engine_status | which engines this install can query | none |
You don't need an API key to get the interesting part. audit_plan and
audit_score let the assistant you're already talking to be the engine: it
fetches your site, writes the buyer questions, answers them with its own web
search, and hands the results back for scoring. Ask "is my site visible in AI
search?" and you get a real citation gap with no signup, no key, no billing.
The tool still does the arithmetic, and still refuses a question that names your
brand — that guard is what makes the number worth having. It measures visibility
in that assistant, which the report says plainly. For the cross-engine
picture, use audit with your own keys.
There's also a SKILL.md so agents know how to interpret the
numbers — that readiness isn't visibility, that 0% is a normal starting point,
and that partial engine coverage must be reported as partial.
Use it as a library
How this stays free
Everything here is free forever, self-hosted, AGPL — full capability, no feature locks, no telemetry, no account, no upsell in the output.
There is no paid tier today, and this README won't pretend otherwise. If one ever appears it will be for the single thing self-hosting genuinely can't give you — scheduled re-runs, stored history and alerts, the infrastructure to watch a number move over months without running anything yourself. The auditing itself stays here, complete, under AGPL.
Which is also the answer to "what stops this going closed later": the licence. Anyone offering it as a hosted service has to open their version too, including whoever wrote it.
Contributing
Issues and PRs welcome. The checks are deliberately conservative: every one has to be something an assistant or its crawler actually depends on, and verifiable by hand. If a check can't be justified that way, it doesn't belong here.
CONTRIBUTING.md covers the setup, the bar a new check has to clear, and what an engine adapter needs to get right.
License
AGPL-3.0. Use it, fork it, run it commercially, embed it in your own tooling. If you offer it to others as a hosted service, that service must be open too.
Built because "am I in the answer?" deserves a real measurement, not a vibe.