The Invisible Instructions Guiding AI Search

When you ask an AI chatbot a question about current events, it doesn't just rely on its pre-trained memory—it actively searches the web to find the answer. But what exactly is the AI typing into its internal search bar while you wait for a response?

Recent data suggests that the invisible mechanics governing how AI browses the internet are undergoing a dramatic shift. A company called Promptwatch, which specializes in tracking how chatbots like ChatGPT, Claude, and Gemini respond to prompts, recently noticed a massive anomaly under the hood of ChatGPT's search tool.

For weeks, the AI rarely restricted its searches to specific websites. The use of domain-specific operators—the equivalent of a human typing "site:nytimes.com" into Google—hovered at a mere 0.5% of all tracked queries. Then, almost overnight in early August, that figure skyrocketed to between 16% and 17%.

This sudden spike wasn't a glitch. It coincided precisely with OpenAI's rollout of a new model update, dubbed "GPT-5.6 Sol." At the time, OpenAI vaguely announced that the update would make the chatbot "more reliable with facts and provide more focused answers." The data reveals how they actually achieved this: rather than letting the AI scan the wild, unfiltered web, the system's hidden instructions are increasingly forcing it to look for answers only within specific, pre-approved digital walls.

This newfound focus comes with casualties. A few weeks after the initial spike, analytics showed a sharp decline in ChatGPT using Reddit as a source for its answers. While the exact system prompts remain hidden by OpenAI, the behavioral shift is clear. The developers are quietly turning the dials on the AI's "information diet," seemingly prioritizing authoritative domains over crowdsourced, conversational chatter to reduce hallucinations and improve factual accuracy.

This invisible tug-of-war over what the AI reads has birthed an entirely new industry: Generative Engine Optimization (GEO). For the last two decades, the internet has been shaped by SEO—the art of tweaking websites to please traditional search engine algorithms. Now, GEO consultants are stepping in to solve a new puzzle. They are building tools and strategies to ensure their clients' websites aren't left out when AI models decide which domains are worth querying.

The web isn't shrinking, but the gatekeepers are changing. As AI chatbots become the primary way many people find information, understanding how these systems filter, select, and ignore sources is crucial. The battle for digital visibility is no longer just about getting humans to click a blue link; it's about convincing an artificial intelligence that your data is worth reading in the first place.

Key Points

• Analytics reveal ChatGPT recently increased its use of domain-restricted searches from 0.5% to over 16%.
• This shift aligns with updates aimed at making the AI more factually reliable and focused.
• The AI has also notably reduced its reliance on crowdsourced platforms like Reddit for answers.
• A new industry called GEO (Generative Engine Optimization) is emerging to optimize content specifically for AI chatbots.

Why It Matters

As AI chatbots replace traditional search engines for many users, the algorithms dictating what these bots read will shape public knowledge. The rise of GEO indicates a fundamental shift in how digital content must be structured to remain visible.

Sources:

• ChatGPT search now uses the site:operator at scale — Simon Willison's Weblog