How do I run a GEO audit? A checklist with a command for every check

How to use this checklist

Set your address once, then run the checks in order:

SITE=https://yoursite.com

The order follows the gates of the Katman method: first bots must read the page, then each page needs one clear address, then the machine-readable files must match, and last there must be pages worth recommending, measured against real answers.

Gate 1: can AI bots get in and read the text?

1. robots.txt doesn’t block search bots.

curl -s $SITE/robots.txt | grep -iE '^(user-agent|disallow):'

Pass when: no Disallow: / sits under *, OAI-SearchBot, Claude-SearchBot or PerplexityBot. OpenAI says sites that opt out of OAI-SearchBot aren’t shown in ChatGPT search answers, and that GPTBot (training) is a separate choice (OpenAI, “Overview of OpenAI crawlers”, read 15 Oct 2026).

2. Each bot gets a 200 through your CDN.

for b in OAI-SearchBot/1.0 ChatGPT-User/1.0 Claude-SearchBot/1.0 PerplexityBot/1.0; do
  echo "$b: $(curl -s -o /dev/null -w '%{http_code}' -A "Mozilla/5.0 (compatible; $b)" $SITE/)"
done

Pass when: every line ends in 200. A 403 usually comes from a CDN or firewall rule, not your code: see Cloudflare blocking AI crawlers.

3. The text is in the HTML without JavaScript.

curl -s $SITE/ | sed 's/<[^>]*>/ /g' | wc -w
curl -s $SITE/ | grep -o '<h1[^>]*>[^<]*' | head -1

Pass when: hundreds of words and a visible H1. A crawler that doesn’t run JavaScript sees only this.

4. Missing pages return 404.

curl -s -o /dev/null -w '%{http_code}\n' $SITE/this-page-does-not-exist

Pass when: 404. A “not found” screen sent with 200 makes every unknown URL look like a real, empty page.

Gate 2: does every page have one clear address?

5. One host. http, www and any preview copy lead to the main address.

curl -sI http://yoursite.com/ | grep -iE '^(HTTP|location)'
curl -sI https://www.yoursite.com/ | grep -iE '^(HTTP|location)'

Pass when: both return a 301 to your main https address, and a Google search for your brand on your host’s preview domain finds nothing.

6. Canonicals point to the page itself.

curl -s $SITE/pricing | grep -io '<link[^>]*rel="canonical"[^>]*>'

Pass when: each page names its own absolute URL.

7. The sitemap lists only live, canonical pages.

curl -s $SITE/sitemap.xml | grep -o '<loc>[^<]*' | sed 's/<loc>//' |
  while read -r u; do echo "$(curl -s -o /dev/null -w '%{http_code}' "$u") $u"; done

Pass when: every line starts with 200, and no login, checkout or noindex page is listed.

8. Links are real links.

curl -s $SITE/ | grep -o '<a [^>]*href="/[^"]*"' | wc -l

Pass when: at least as many as the pages in your menu. A div with onClick isn’t a link to a bot.

Gate 3: do the machine files match the pages?

9. llms.txt, if you have one.

curl -s -o /dev/null -w '%{http_code}\n' $SITE/llms.txt
curl -s $SITE/llms.txt | head -5

Pass when: 200, the first line is an H1 (the one required part of the spec, per llmstxt.org, read 15 Oct 2026), and no price or claim contradicts your pages. Google says you don’t need new machine-readable files or special schema to appear in AI Overviews or AI Mode (Google, “AI features and your website”, read 15 Oct 2026), so this check is optional. Background: What is llms.txt?

10. JSON-LD is valid and says what the page says.

curl -s $SITE/ | python3 -c "import sys,re,json; b=re.findall(r'application/ld\+json\">(.*?)</script', sys.stdin.read(), re.S); [json.loads(x) for x in b]; print(len(b), 'valid blocks')"

Pass when: a number above 0, no errors in Google’s Rich Results Test, and every price and claim in it is visible on the page.

Gate 4: pages worth naming, and measurement

11. A page for each customer situation. Write 10 to 12 sentences a customer would type into ChatGPT, without your brand. For each, find the URL whose H1 says it and whose first three sentences answer it. Example, for an invented invoicing tool: “I need to invoice a German client in euros from a US company” deserves its own page, with that sentence as the H1. Pass when: every situation has a page. On our sample site, 59% of ChatGPT visitors landed on two such pages (source: the site’s own analytics, human sessions only, 21 to 27 Sep 2026; one site, our own product, not a forecast; case study).

12. Measurement. Ask those questions logged out with web search on, in ChatGPT, Claude, Gemini and Perplexity, and record who is named and which pages are cited. Pass when: you have a dated baseline and a date 14 days later to ask again. The full method: How do I check whether ChatGPT recommends my site?

Checks 2 and 3 can also be run from the browser with the AI crawler checker.

What Katman does with this checklist

Katman MCP runs inside your AI coding tool. katman_audit_site runs checks 1 to 10 on the live site, through your CDN; katman_scan_code finds the cause of a failure in your code; katman_prompt returns the fix prompt with its acceptance check; katman_page_brief writes briefs for check 11; katman_visibility_run handles check 12 on your own OpenRouter key, about $2.40 a run on the default panel. Katman Pro is $99 a month for up to 3 sites, with fix prompts and page briefs uncapped. Or have us do it: Growth writes 30 pages a month for you. See what Katman does.

What it can’t do: it doesn’t fix anything by itself (your coding tool does), doesn’t track rankings, counts full audits per month, and nobody can promise an assistant will recommend you after every check passes.

Written by the Katman team, Rnify Limited. OpenAI, Google and llmstxt.org pages read 15 October 2026.