How Do You Audit Your Own Website With AI?

Why we are giving away our audit prompt

Every new client starts with an audit. For years that meant a person with a crawler, a spreadsheet, and two days. Then it meant a person with a crawler and an AI model, and half a day. Now the AI model with web access does most of the fetching, and the person's job is to decide what matters and in what order.

The prompt that drives that process is not the valuable part. The valuable part is knowing which of the forty things it flags will actually move the phone, and doing the work. So we are publishing the prompt. If you run it yourself and fix three things, good. If you run it, see the list, and decide you would rather someone else did it, you know where we are.

The other reason: most "AI SEO audit" prompts floating around produce confident nonsense. They report Core Web Vitals scores the model never measured and indexing counts it never saw. We wanted one that is built to be honest about what it cannot see, and the only way to prove that is to show it.

What the three passes do

Each pass returns the same four blocks: a pass/warn/fail scorecard with a one-sentence "what it costs you" for every failure, the stop-everything issues, a numbered fix list ordered by impact, and a list of things the model could not verify with the exact place for you to look.

How to run it

1. Open Claude, ChatGPT, or Gemini. Use a model with web browsing turned on. Without browsing, the model will make things up, and the prompt cannot stop it.
2. Go to the contractor website audit prompt on GitHub and open prompts/1-foundation.md.
3. Copy the whole file. Replace {{SITE_URL}} with your homepage and {{TRADE}} with your trade.
4. Paste it and run it. Read the fix list. Do the "owner must check" items yourself; they take ten minutes in Search Console and PageSpeed Insights.
5. Repeat with pass two, then pass three. Pass three also needs your business name and primary city.

Then run pass two on the top competitor in your map pack. The gap between their page table and yours is your work list for the quarter.

How the prompt keeps the AI honest

This is the part we spent the most time on, and it is why we think this prompt is worth running when most are not.

• It can only report what it fetched. Every finding must cite a URL. A problem on a page the model did not fetch is not allowed in the output.
• It has to name what it cannot see. Search Console, Analytics, Google Business Profile insights, and real Core Web Vitals scores are behind logins or tools the model does not have. The prompt forces an "owner must check" block for those, with the exact report name. If the model returns a confident number for one of them, it ignored the instruction and you should throw the answer out.
• Page content is data, not instructions. The prompt tells the model to ignore anything on the audited site that reads like an instruction. That matters more than it sounds; a competitor's site is an untrusted input.
• It writes for the owner, not the SEO. Every failure comes with one sentence on what it is, one on what it costs, one on the fix. No jargon the office manager cannot act on.

Where the prompt came from

We did not start from a blank page. In early 2026 we installed Corey Haines' open-source marketingskills collection, which includes an SEO audit skill written for SaaS marketers, and ran it for a few months. Then we rebuilt it for our clients: reorganized it into the three passes, rewrote every check in plain English for contractors, added the local, review, brand, competitor, and AI-search checks that the original did not cover, and added the "owner must check" discipline. The version we run internally also pulls live Search Console, Analytics, and Ahrefs data, scores fifteen weighted categories, and produces a 30/60/90 day plan. That part is not in the repo. The three prompts are, under an MIT license. Thanks to Corey for the starting point.

What we found when we ran it

On October 8, 2026 we ran passes one and two, as a script, against ten real roofing company websites in Orange County. All ten were reachable and none was thin, which surprised us. But eight of ten had images with no alt text, five had no business schema at all, three had title tags Google truncates, and one Anaheim roofer's site tells Google it is in Temecula. The full results are in day five, what broke on ten Orange County roofing websites.

If you want context on what a professional audit should deliver before you run your own, we wrote that up in what an SEO audit should actually deliver, and the trade-specific version for roofers in what roofer SEO should actually improve.

The rest of the series

• Day 2: The six foundation checks that come before everything else
• Day 3: Why your service pages don't rank
• Day 4: The local SEO audit that leaves your website
• Day 5: We audited ten Orange County roofing sites. What broke.

If you would rather we ran it

The full audit is the first step in the Foundation Engine, and the findings drive the work in our contractor SEO and AI for contractors services after that. If you would like the fifteen-category version run on your site with real data, request a free audit. You get the scorecard whether or not you hire us.