Auditor agent: the X-ray, claude-seo, and the list with evidence
The first agent worth building is the one that only reads. It runs the X-ray, runs claude-seo, and returns a list of problems, each with the line that proves it. It fixes nothing: you decide what to change, and the 2026 yardstick filters out what's just a fad.
4Run every recommendation through the 2026 yardstick
Goal: run the read-only X-ray and claude-seo, and have the agent list problems with evidence, changing nothing.
01 · From the X-ray to the list you decide on
Tap each block.
Click each block in the diagram to see what it does.
The auditor only reads. Whoever writes to the site goes through a gate.
02 · How to build the auditor
Tap each step.
1. X-ray in a file
Run the method's auditoria.sh with the home page and the pages that matter, and save the output to a file. That's the evidence the agent will read.
The script only makes read requests (curl). It can run on any public site, including a competitor's.
2. claude-seo
claude-seo needs Python 3.10 or higher and Claude Code. The /seo audit, /seo technical, and /seo geo commands run without paying anything beyond your subscription.
The DataForSEO extension (search volume, SERP, backlinks) requires a paid account: the README mentions a minimum top-up of US$ 50 and US$ 1 for a test (README from 03/06/2026; check the site).
Install it after readingPrefer to clone and read install.sh before running it. A script piped straight into bash runs without you seeing what it does.
3. The request
Ask for a table: finding, evidence line copied from the file, severity, and proposed fix. And say what it can't do: change files, run commands that write, call APIs.
A finding without an evidence line is discarded, as in the check from lesson 3.
4. The yardstick
claude-seo's GEO module treats llms.txt as an emerging standard, and its schema reference on FAQ dates from 2023. Mark those items as [yardstick] and decide by what holds in 2026.
Blocking a training bot doesn't block a search bot: robots.txt is decided by purpose (course 1, lesson 6).
03 · The X-ray, claude-seo, and the request to the agent
git clone https://github.com/AgriciDaniel/claude-seo.git
cd claude-seo
less install.sh # read it before running
./install.sh
claude # opens Claude Code
/seo audit https://www.seusite.com
/seo technical https://www.seusite.com
/seo geo https://www.seusite.com
Request to the auditor agent
Read auditoria.txt (and the /seo audit report, if there is one).
Return a table with up to 10 findings: finding | evidence line copied from the file | severity (high/medium/low) | proposed fix.
Rules:
- A finding without an evidence line doesn't go in.
- Do not change any file, do not run any command that writes, do not call any API.
- Mark [yardstick] every recommendation about llms.txt, FAQ, or "optimizing for AI".
- Bots in robots.txt: separate search, user action, and training.
Redirects: http and non-www go to a single address (308). Nothing to do.
robots.txt: ClaudeBot is in the same group as the search bots, but it is a training bot. Decide by purpose.
courses-sitemap: 0 hreflang annotations. The agent points it out; you know the context (the listing pages are Portuguese-only and the translated version is a different site).
llms.txt 200: [yardstick] optional and cheap, not a citation lever.
04 · Practice: your first auditor
Use your own site. If you don't have one, use a public site you know well.
E1 · X-ray in a file
20 min
Run auditoria.sh on the home page and 3 more important pages, and save the output to auditoria.txt. If you have Claude Code, also run /seo technical on the home page.
Done when: auditoria.txt exists with the 4 pages, each with status, title, canonical and the internal link count.
E2 · The list with evidence
30 min
Give the auditor's request to the agent. Check each evidence line against the file and write your decision next to each finding: fix, ignore (with the reason) or [régua].
Done when: you have 5 findings, each with a line that really exists in auditoria.txt and your decision, and the project's git status shows the agent changed nothing.
05 · Test what you learned
The explanation appears after your answer.
Question 1 of 4
The auditor agent flagged a problem, but didn't copy any line from the output. What should you do?skill
Correct
A finding without evidence is an opinion. The request requires the line copied from the output; without it, the finding doesn't make the list.
The core runs with Claude Code and Python. Market data (exact volume, SERP, backlinks) comes from DataForSEO, which is paid. Lesson 7 shows when that pays off.
Not this one
The core runs with Claude Code and Python. Market data (exact volume, SERP, backlinks) comes from DataForSEO, which is paid. Lesson 7 shows when that pays off.Review: How to build the auditor
Question 3 of 4
In the real output, ClaudeBot appears in the group with OAI-SearchBot and Claude-SearchBot. What is the finding?
Correct
Search, user action and training are different purposes. Allowing or blocking training is a separate decision, and it doesn't affect search.
The GEO module suggests creating llms.txt to show up more in AI answers. How should you handle it?
Correct
An Ahrefs study (137 thousand domains, 06/15/2026) found that 97% of the files get no requests at all. You can have one, but not as a lever.
Not this one
An Ahrefs study (137 thousand domains, 06/15/2026) found that 97% of the files get no requests at all. You can have one, but not as a lever.Review: How to build the auditor
Quiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.
Choose an option. The explanation appears after your attempt.
06 · To take with you
Read-onlyThe auditor reads and points things out. Writing to the site goes through a gate.
EvidenceEach finding comes with the line that proves it; no line, no entry.
Free firstauditoria.sh and the claude-seo core. DataForSEO is optional and paid.
2026 yardstickllms.txt is optional; the FAQ rich result ended on 05/07/2026; robots by purpose.