Original Investigation by Jason Packer
The original investigation into how ChatGPT scrapes Google Search and leaks user prompts into Google Search Console was conducted by Jason Packer. This important privacy research revealed how Google's tokenization system makes ChatGPT prompts visible to website owners.
Read the full investigation on Quantable:
Read Jason Packer's Original ArticleHow the Prompt Leak Works
When ChatGPT uses its web browsing tool, it doesn't search Google the way a human would — typing in a phrase and scanning results. Instead, it first translates your conversational prompt into one or more keyword-style queries, then fires those queries against the web.
Google logs every query it receives. Website owners can see those queries in Google Search Console, tagged to the pages that appeared in the results. What Packer's investigation found: some Search Console queries read less like conventional keyword searches and more like fragments of chatbot conversations — natural-language phrases that appear to originate from inside a ChatGPT session.
The “leak” isn't OpenAI sending your prompts to Google. It's the prompt-to-query translation step: a conversational prompt becomes a keyword query, and that keyword query enters the web's existing logging infrastructure.
From Aiso's panel of millions of anonymized AI conversations
One prompt becomes ten queries
A September 2024 entry shows the translation step in practice. An AI assistant, given the query “Make a list of the 10 most important AI papers in 2024, check the latest news & tell me about the latest rumors in AI”, first generated these ten Google search queries before attempting to answer:
- 1. “Top 10 AI papers of 2024”
- 2. “Most important AI research papers in 2024”
- 3. “Latest news in artificial intelligence”
- 4. “Recent developments in AI technology”
- 5. “Rumors and trends in the AI industry”
- 6. “Upcoming AI conferences and publications 2024”
- 7. “Breakthrough AI advancements in 2024”
- 8. “Current state of artificial intelligence research”
- 9. “Emerging AI technologies and innovations”
- 10. “Predictions for AI in 2024”
Each of those ten queries was sent to Google as a real search, logged by Google, and visible in the Search Console of any site that ranked for them. A researcher publishing AI paper roundups would have seen all ten in their Search Console — each looking like an organic keyword search, with no indication it was a fragment of one chatbot user's question.
What This Means for Marketers and SEOs
Search Console queries that read like natural-language questions rather than conventional keyword searches may be AI-generated — echoes of what real users are asking chatbots. This gives you a partial, indirect window into the prompt clusters driving traffic to your content.
The inverse is equally useful: if you want your content cited in AI answers, you need to cover the keyword clusters an AI will generate as web queries when a user asks a relevant question. That's why mapping AI fan-out queries matters — the queries ChatGPT generates are the content targets you need to rank for.
For a fuller picture of how to measure your brand's visibility in AI answers, see our complete guide. And if you want to track the actual prompts people use when asking about your category, Aiso's panel surfaces real, anonymized conversations — not just keyword-style signals filtered through Google's tokenization layer.
Related: Understanding Google Tokenization
If you're interested in understanding the technical mechanism behind how Google's tokenization system works, we've written a detailed explanation:
How Google's Tokenization Turns 'openai + index + chatgpt' Into Search Queries