ChatGPT search is now using Google’s site: operator to filter results by domain. Promptwatch, a monitoring service that tracks how chatbots respond to specific prompts, caught this in action and published their findings this week.

If that sounds familiar, it should. The site: operator has been a staple of Google search for decades. Type site:github.com rust web framework and you get results only from GitHub. It’s basic but effective.

ChatGPT is apparently now doing this automatically when it thinks focusing on specific domains will improve results. Which is fine. Makes sense, even.

The problem isn’t the feature. It’s what’s already growing around it.

The GEO Industrial Complex

Promptwatch isn’t just tracking chatbot responses for fun. They’re part of what they’re calling “GEO”, short for Generative Engine Optimization. It’s SEO for chatbots. The same playbook, different search box.

The pitch is predictable: companies need tools and consulting to increase their “presence in replies to prompts” inside ChatGPT, Claude, and Gemini. Promptwatch uses automation to track how often your site shows up in chatbot answers, publishes aggregate reports, and presumably sells you advice on how to improve your numbers.

We’ve seen this movie before. In the late 90s and early 2000s, SEO went from “write good content with relevant keywords” to a multi-billion dollar industry of consultants, tools, and increasingly elaborate tricks to game Google’s algorithm. Some of it was legitimate. A lot of it was snake oil.

Now we’re doing it again, except the algorithm is a black box language model instead of a black box search ranking system. At least with Google you could sometimes figure out what changed. With LLMs, you’re optimizing for a moving target that might give different answers to the same prompt on consecutive tries.

What Actually Works

Here’s the thing: if ChatGPT is using the site: operator, it’s looking for domains it already knows are relevant. You don’t optimize your way onto that list. You get there by actually being a useful source of information on your topic.

The fundamentals haven’t changed. Clear, accurate content. Good information architecture. Structured data where it makes sense. If you were doing SEO right, you’re probably fine for GEO too.

What definitely won’t work: paying a consultant to help you “optimize for prompt presence” when nobody, including OpenAI, can tell you exactly how the model decides which sites to query.

Other Tools News

Ramp, the corporate spend management company, launched its own AI model router this week. It’s called Router, because naming is hard.

Router lets you switch between different language models via an API. It’s the same idea as services from OpenRouter, Martian, and a half-dozen other companies in this space. Send a request, Router picks a model or lets you specify one, you get a response.

Ramp says Router is aimed at companies that want to experiment with multiple models without managing a bunch of different API integrations. Fair enough. The model routing space is getting crowded, but there’s clearly demand.

Bun 1.4 also shipped this week, the first stable release since the project’s Rust rewrite earlier this year. The release notes downplay the rewrite and focus on compatibility improvements: 1,517 new tests from the Node.js test suite, 2,900 bug fixes.

The interesting new feature is Bun.WebView, which lets you spawn a headless browser instance and interact with it via a JSON API. Think Puppeteer or Playwright, but built into the runtime. Simon Willison has already started experimenting with it as a shot-scraper alternative for automated web scraping and screenshots.

If you’re already using Bun, worth checking out. If you’re happy with Node or Deno, there’s no urgent reason to switch.

The Bottom Line

GEO is real in the sense that people will definitely pay money for it. Whether it actually works is another question.

If you’re a publisher worried about chatbot visibility, focus on being correct and comprehensive. Structure your content well. Make it easy to parse. That’s not exciting advice, but it’s free and it’ll probably work better than whatever a GEO consultant wants to sell you.

And if someone tries to pitch you on “prompt optimization” services, ask them to show their work. Specifically, ask them to explain how they’re controlling for the fundamental randomness of LLM outputs. If they can’t give you a straight answer, save your money.

The internet doesn’t need another optimization grift. We’re still cleaning up from the last one.