What ChatGPT Can (and Cannot) See on Your Website: A Comprehensive Audit

As AI tools like ChatGPT become increasingly integrated into how people find and consume information online, understanding what these systems can and cannot "see" on your website is crucial for content creators, marketers, and developers.

Introduction

With the rise of AI-powered search and retrieval systems like ChatGPT's Search and Bing-powered search features, the way users discover and interact with web content is fundamentally changing. Instead of browsing through search results and visiting websites directly, users increasingly ask AI assistants to find and summarize information for them.

This shift raises critical questions for website owners and content creators:

• Can AI systems like ChatGPT effectively "see" and retrieve all the content on my website?
• What types of content structures or technical implementations might prevent AI systems from accessing my content?
• How can I optimize my website to ensure it's fully accessible to AI retrieval systems?

To answer these questions, we conducted a methodical audit of ChatGPT's web retrieval capabilities across a series of controlled test websites with varying levels of complexity.

Methodology

We created a series of increasingly complex test websites, each designed to test specific aspects of ChatGPT's retrieval capabilities:

1. We built five test websites with controlled content structures
2. Each site contained both general content and specific "marker content" with unique identifiers
3. We deployed all sites using Vercel connected to GitHub repositories
4. We tested ChatGPT's ability to retrieve specific information from each site using the "Browse with Bing" feature
5. We documented the results to understand what ChatGPT could and couldn't access

Test Sites Overview

We created the following test sites, each with a specific purpose:

Overview

Initial tests with Site 3 (https://ai-retrieval-test-site-3.vercel.app/) revealed important findings about ChatGPT's ability to access JavaScript-dependent content.

Test Questions and Results

Key Findings

Implementation Implications

These findings have significant implications for developers:

• If you want content to be accessible to ChatGPT, it must be included in the initial HTML
• JavaScript-based progressive enhancement may hide content from AI systems
• Dynamic loading patterns need careful consideration if AI accessibility is important

Recommendations for Developers

Site 4: Image-Based Content (OCR Test)

Test: Mythica Capital and Founding Date

Question: "Can you visit https://ai-retrieval-test-site-4.vercel.app/ and tell me what is the capital of Mythica and when it was founded?"

Analysis: This experiment revealed significant limitations in ChatGPT's OCR capabilities. Despite the image containing clearly legible text about the fictional kingdom of Mythica, ChatGPT was unable to extract this information. Instead, it generated a plausible but entirely fabricated response. This demonstrates that ChatGPT cannot reliably "read" text embedded in images, even when that text is clearly visible and formatted in a standard font. Notably, ChatGPT delivered its incorrect response with apparent confidence, without indicating any uncertainty about the information.

Site 5: White-on-White Text Test

Test: Hidden Treasure Location

Question: "Can you visit https://ai-retrieval-test-site-5.vercel.app/ and tell me where the secret treasure of Eldoria is hidden?"

Analysis: This experiment revealed that ChatGPT processes the raw HTML content of webpages rather than a visual representation. Despite the text being styled as white-on-white (completely invisible to human visitors without selecting the text or inspecting the code), ChatGPT was able to perfectly extract this information. This demonstrates that CSS styling which affects visual presentation (like text color) has no impact on ChatGPT's ability to access and retrieve content. This finding stands in stark contrast to our OCR test, highlighting that ChatGPT can "see" invisible text in HTML but cannot "see" visible text in images.

Summary of Results

Actionable Insights for Marketers and Developers

Conclusion

Our comprehensive audit of ChatGPT's web retrieval capabilities has revealed both strengths and limitations in how AI systems access and retrieve web content. While ChatGPT can successfully retrieve static content and execute JavaScript to access hidden content in tabs and accordions, it struggles with following links without explicit relevance signals and cannot access content that requires user interactions to load dynamically.

These findings have significant implications for website owners, content creators, and marketers who want to ensure their content remains discoverable in an AI-first world. By implementing the actionable insights from our research, you can optimize your website for AI retrieval and ensure that your valuable content doesn't become invisible to users who rely on AI assistants to find information.

As AI retrieval systems continue to evolve, understanding these capabilities and limitations will become increasingly important for effective digital content strategy. By staying informed about how AI systems interact with web content, you can adapt your approach to ensure your content remains accessible and discoverable, regardless of how users choose to find it.