How llms.txt made ChatGPT my #1 traffic source (a free IP API, 8 weeks in) Tags: webdev, seo, ai, cloudflare

I run HackMyIP, a free IP/privacy toolkit: IP lookup, VPN/proxy and DNS-leak detection, email-breach check, WHOIS, a CIDR calculator, and a free no-key API. It's about 8 weeks old. I'm writing this up because one decision moved the needle more than any backlink or keyword I chased: I made the site readable by AI assistants, not just by Google.

Here's the part that surprised me. Looking at the last 28 days of analytics:

• chatgpt.com referrals: 157 sessions (107 of them engaged)
• Google search: 44 sessions

That's roughly 3.5x more real traffic from ChatGPT than from Google. For a small, young site with thin domain authority, that ratio is not what I expected, and it's almost entirely because of a few small files most SEO guides never mention.

What I actually shipped

Three AEO ("Answer Engine Optimization") surfaces, all static, all boring to build:

1. /llms.txt — a markdown index of the site written for an LLM: a one-line summary, the questions the site answers, and a categorized list of every tool with a one-line description and URL. Think of it as a sitemap that reads like documentation instead of XML.
2. /llms-full.txt — the expanded version with more detail per tool, for assistants that will read a longer file.
3. An OpenAPI 3.1 spec + an ai-plugin.json manifest under /.well-known/, so an assistant (or a plugin runtime) can discover the API and the exact endpoints programmatically.

The API itself is the thing those files point at: 10 endpoints, no key, no signup, JSON responses, CORS enabled (Access-Control-Allow-Origin: *). For example, GET /api/ip returns your IP plus geolocation, ASN/ISP, and a privacy classification:

{
  "success": true,
  "data": {
    "ip": "…",
    "location": { "city": "…", "region": "…", "country": "…", "timezone": "…" },
    "network": { "asn": 3462, "isp": "…", "tls_version": "TLSv1.3" },
    "privacy": { "type": "residential", "score": 90, "grade": "A", "is_vpn": false, "is_datacenter": false }
  }
}

Why this works (my best read of it)

When someone asks ChatGPT "what's a free IP lookup API with no key?" or "how do I test if my VPN is leaking DNS?", the assistant needs a citable, machine-readable source. A normal marketing page buries the answer in layout and JS. An llms.txt hands the assistant exactly what it needs: here are the tools, here's what each does, here are the URLs. The OpenAPI spec does the same for the API. You're not gaming a ranking. You're lowering the cost for an assistant to recommend you accurately.

Two things I'd stress if you try this:

• Be honest in the file. Every claim in my llms.txt is something the site actually does. If an assistant cites you and you don't deliver, that's worse than not being cited. (My API really is keyless; my leak tests really run client-side.)
• It compounds with usefulness, not instead of it. llms.txt didn't create the tools. It made the existing useful tools discoverable to a new class of referrer. If the underlying thing isn't genuinely good, an index of it won't help.

What I'd do the same again

• Write llms.txt by hand, structured by user-intent ("questions this site answers") not by your internal nav.
• Ship a real OpenAPI spec if you have any API at all. It's the difference between "a site with an API" and "an API an assistant can call."
• Keep it current. A stale index that lists tools you removed is a trust leak.

I'm not claiming llms.txt is magic. 8 weeks of data on one small site is a single data point, and "direct" traffic is noisy. But the chatgpt-vs-google gap has been consistent enough that I now treat AI-discoverability as a first-class channel, not an afterthought. If you've got a developer tool or an API and you've only optimized for Google, you're probably leaving your fastest-growing referrer on the table.

Files are live if you want to see the format: llms.txt · the API.