What ChatGPT Can (and Cannot) See on Your Website: A Comprehensive Audit
Technical Research
What ChatGPT can see on your website. And what it can't.
We tested five controlled sites to see what ChatGPT could retrieve, miss, or misread. The short version: static HTML and explicit links win; dynamic, image-only, or vague content often disappears.
Ben Tannenbaum · January 10, 2025 · 15 min read
We built five controlled test websites to check what ChatGPT could retrieve from static pages, links, JavaScript-driven sections, images, and hidden text. The highest-value finding was blunt: content in static HTML with descriptive links was retrievable; content that depended on user interaction, vague link text, or images was much easier to miss.
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.
Why This Matters
If ChatGPT can't "see" certain content on your website, that content effectively becomes invisible to users who rely on AI assistants to find information. Understanding these limitations is crucial for ensuring your content remains discoverable in an AI-first world.
Critical Questions We Set Out to Answer
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?
Research Methodology
We conducted a methodical audit of ChatGPT's web retrieval capabilities across a series of controlled test websites with varying levels of complexity:
- Built five test websites with controlled content structures
- Each site contained both general content and specific "marker content" with unique identifiers
- Deployed all sites using Vercel connected to GitHub repositories
- Tested ChatGPT's ability to retrieve specific information using "Browse with Bing" feature
- Documented results to understand what ChatGPT could and couldn't access
🧪 Test Sites Overview
Site 1: Text-Only Website
Purpose: Baseline test for simple content retrieval
Description: Basic static HTML with plain text content to establish retrieval baseline
Site 2: Blog Article Website
Purpose: Test link following with generic titles
Description: Main page with links to blog articles using non-descriptive titles
Site 3: Blog Title Hint Website
Purpose: Test link following with descriptive titles
Description: Blog articles with explicit, keyword-rich titles in link text
Site 4: Blog Content Hint Website
Purpose: Test partial title matching
Description: Articles with titles containing partial query matches
Site 5: JavaScript-Hidden Content
Purpose: Test JavaScript execution capabilities
Description: Content hidden in tabs, accordions, and dynamically loaded sections
Site 6: Image-Based Content (OCR)
Purpose: Test text extraction from images
Description: Critical information embedded only in image files
Site 7: White-on-White Text
Purpose: Test HTML parsing vs visual rendering
Description: Text hidden with CSS styling but present in HTML
Key Findings
Static Content Accessibility
High Impact
ChatGPT reliably retrieves static HTML content on main pages
Action Required: Ensure critical information is in static HTML on primary pages
Link Following Behavior
Critical Impact
Only follows links with explicit, relevant titles - ignores generic titles
Action Required: Use descriptive, keyword-rich anchor text for internal links
JavaScript Limitations
High Impact
Cannot execute JavaScript for dynamic content loading
Action Required: Make critical content available without user interactions
Search Engine Workaround for JavaScript Content
Medium Impact
OpenAI uses Bing and Google searches as a workaround to access JavaScript-rendered pages that have been indexed by search engines
Action Required: Ensure JavaScript pages are indexed by Bing and Google; move critical content (especially FAQ answers) above the JavaScript fold
CSS vs HTML Processing
Medium Impact
Processes HTML content regardless of CSS styling/visibility
Action Required: Text in HTML is accessible even when visually hidden
Image Text Extraction
High Impact
Cannot extract text from images, may hallucinate responses
Action Required: Provide text alternatives for image-based content
6 Key Actions to Optimize Your Website for AI Retrieval
1. Use Explicit, Descriptive Link Text
Critical
Ensure links to important content contain explicit keywords that clearly indicate what information can be found on the linked page
Implementation:
- Include target keywords in anchor text
- Avoid generic 'Click here' or 'Read more'
- Make link text descriptive and specific
2. Optimize Page Titles
High
Explicit page titles significantly improve content discovery, while vague titles result in content being missed entirely
Implementation:
- Use specific, keyword-rich page titles
- Include query-relevant terms in titles
- Avoid vague or generic titles
3. Keep Critical Content in Static HTML
High
ChatGPT can access hidden content in tabs and accordions, but content loaded dynamically after user interactions is invisible
Implementation:
- Place important info in static HTML
- Avoid user-interaction dependencies
- Ensure content is crawlable
4. Place Important Information on Main Pages
Medium
Content on main pages is reliably accessed, while secondary pages are only discovered under specific conditions
Implementation:
- Feature key content prominently
- Include summaries on main pages
- Use clear navigation structure
5. Avoid Image-Only Content for Critical Information
High
ChatGPT cannot reliably extract text from images, while it processes HTML content regardless of styling
Implementation:
- Provide text alternatives for images
- Use alt text descriptively
- Include text versions of image content
6. Ensure Search Engine Indexing
Medium
Even JavaScript-dependent content can be discoverable if it's properly indexed by search engines
Implementation:
- Focus on search engine crawlability
- Use server-side rendering when possible
- Implement proper SEO practices
Conclusion: Preparing for an AI-First World
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.
Final Thought
The rise of AI assistants represents a fundamental shift in how users discover and consume web content. Just as websites had to adapt to mobile devices and search engine algorithms in the past, they must now adapt to AI retrieval systems to remain visible and relevant in the evolving digital landscape.