Case study · AI product

A free tool that audits a 25-page site in under a minute, for Google and for AI search

I built it alone. Paste a web address and you get three scores, a to-do list ranked by impact, and code you can paste in to fix each issue. On a real run it checked 25 pages of djangoproject.com in 57 seconds and scored the site 76 out of 100. Data costs: €0.

  • Role: sole developer
  • Stack: Python, FastAPI, Playwright, Claude Code agents, Docker, GitHub Actions
  • Checks: Google and Bing, answer engines, and 14 AI crawlers including ChatGPT, Claude, Perplexity and Gemini
  • Code: github.com/elkushi/elkoushy_seo_geo_aeo
  • Where it runs: on my own machine and in Docker. I don't host it publicly, so every screenshot here is from a real run.
The report for djangoproject.com: an overall score of 76 out of 100, separate scores for SEO, AEO and GEO, and a short summary of what to fix first.
A real audit of djangoproject.com, run on 26 October 2026: 25 pages checked in 57 seconds.
What this proves: I can turn a vague marketing question, "is this site visible?", into a tool that answers it the same way every time and hands over the fix.

The problem

People don't only search Google any more. They ask ChatGPT, Perplexity or Google's AI answers, and those tools quote some websites and ignore others. Most audit tools still check only for Google, and the good ones are paid.

I wanted one free report that covers all three:

  • SEO: can Google and Bing find, read and rank the pages?
  • AEO (answer engines): is the content structured so a search engine can lift a direct answer from it?
  • GEO (AI search): can AI tools reach the site, understand who is behind it, and cite it?

What you get

< 1 minfor a 25-page audit
76/100the score it gave djangoproject.com: 25 pages in 57 seconds
3 scoresSEO, AEO and GEO, plus an overall score
14AI crawlers checked: ChatGPT, Claude, Perplexity, Gemini and others
0 €free data sources only; the AI step is optional
  • A plan in three stages: Week 1 quick wins, then 30 days, then 90 days. The biggest, cheapest fixes come first.
  • Ready-made code for each fix: structured data (JSON-LD), robots.txt rules and an llms.txt draft.
  • The report as a web page, a PDF and a Markdown checklist.
The audit page while it runs: crawling, SEO, AEO and GEO checks get a tick one by one.
The audit runs live. Each step ticks off as it finishes.
A high-priority issue, no Organization schema on the homepage, with ready-made JSON-LD code and a Copy button.
Every issue says why it matters and gives code to paste in.

How it works

  1. A Python crawler reads the homepage, robots.txt, the sitemap and up to 25 pages.
  2. Rule-based checks score each area. Every issue gets a priority and a place in the plan.
  3. The app requests the homepage as each of the 14 main AI crawlers, to see whether the server lets it in. A site can allow a bot in robots.txt and still block it at the firewall.
  4. A template turns the findings into the report, and headless Chromium prints the PDF.

The same pipeline runs from the command line. There, five Claude Code agents can take the findings further: an SEO auditor, an AEO analyst, a GEO analyst, a competitor researcher and a report writer.

AI crawler access table: 14 bots such as GPTBot, ClaudeBot and PerplexityBot, what each one is used for, whether robots.txt allows it, and the live response code.
Which AI bots can reach the site: the robots.txt rule and the live answer from the server.

Built to be used by strangers

It is written as if strangers will use it, because a tool that takes any address from anyone needs guard rails:

  • Sign-in by emailed link. There are no passwords to store or leak.
  • Each user has a monthly quota, and sign-in requests are rate-limited.
  • The crawler refuses private and internal addresses, and checks again after every redirect. A user can't make the server fetch its own network.
  • One site can't be audited over and over in a day, so the crawler stays polite.
  • The container runs as a normal user with every Linux capability dropped. Keys live in .env, outside git.
  • Automated tests cover parsing, robots rules, the address guard, sign-in, quotas and cross-user isolation. GitHub Actions runs them on every push, against both SQLite and Postgres.

The optional AI test

One more step asks a language model the questions a buyer would ask, and checks whether it names or cites the site. It runs on a free, rate-limited tier, so it's off by default and switched on with one setting.

← All work