Step-by-Step: How to Uncover the Exact Queries GPT-5 Sends ( query fan-out )

When ChatGPT searches the web, it doesn’t send your prompt as-is.
Behind the scenes, GPT-5 performs a process known as a query fan-out — it rewrites your original question into one or more search queries designed to fetch the most relevant, factual, and high-authority web results.

These rewritten queries are not guesses; they’re the actual terms the model sends to the web index it accesses.
For example, when you type:

“What are the best tools for improving Core Web Vitals?”

GPT-5 may silently send:

1. “best tools to improve core web vitals”

2. “core web vitals optimization software 2025”

3. “page speed insight alternatives”

These are the precise search queries that shape the model’s information retrieval — and ultimately influence the output you see in ChatGPT’s answer box.

You can access them through the metadata.search_model_queries field embedded in ChatGPT’s internal network response.
This guide will show you, step-by-step, how to uncover those exact queries, interpret them, and use them strategically for AI-driven SEO.

What Are search_model_queries?

Every time GPT-5 triggers a web search, it generates an internal metadata structure to record what it asked the search system for.

Inside that metadata, there’s a field called search_model_queries, which contains the final, cleaned queries GPT-5 used.

Example JSON snippet:

"metadata": {
  "search_model_queries": {
    "type": "search_model_queries",
    "queries": [
      "best seo tools for small businesses 2025",
      "seo software comparison for startups",
      "affordable keyword research tools"
    ]
  }
}

Each item in the queries array represents one independent search fan-out that GPT-5 ran to collect data before generating your answer.

Understanding these queries allows you to see how the model:

• Interprets your intent

• Simplifies or restructures your phrasing

• Chooses topical angles to retrieve from the web

Step-by-Step: How to Uncover the Exact Queries GPT-5 Sends

This method uses built-in browser developer tools — no external software or extensions are required.

Step 1: Trigger a Web Search in ChatGPT

1. Open ChatGPT (GPT-5) in your browser.

2. Type a prompt that clearly requires real-time information, such as:

   “What are the top AI content writing tools in 2025?”

3. Wait until the “Searching the web...” message disappears and the model finishes responding.

You’ve now created a conversation that includes at least one network call to the OpenAI search subsystem.

Step 2: Identify the Chat ID

1. Look at your browser’s address bar.
   You’ll see a URL in this format:

    https://chat.openai.com/c/68d13850-abc123xyz
   

2. Copy the alphanumeric code after /c/ — this is your conversation ID.
   Example:

    68d13850-abc123xyz
   

You’ll use this ID to locate the relevant network response file in the next steps.

Step 3: Open Developer Tools

1. Right-click anywhere on the ChatGPT page and choose Inspect (or press Ctrl + Shift + I on Windows, Cmd + Option + I on Mac).

2. Click the Network tab at the top of the DevTools panel.

3. Keep it open for the next step.

Step 4: Reload the Page and Filter Requests

1. Refresh the ChatGPT page (Ctrl + R).

2. In the search bar within the Network panel, paste your chat ID (e.g., 68d13850).

3. This filters all network requests related to that specific conversation thread.

You’ll now see a list of JSON or XHR (XMLHttpRequest) calls that represent background communications with OpenAI’s servers.

Step 5: Find the Relevant JSON Response

1. In the filtered list, locate a request whose “Type” column says “fetch” or “xhr”.

2. Click on it, then open the Response tab on the right-hand side.

3. Scroll or expand until you see data formatted in JSON.

   

Step 6: Search for the Term search_model_queries

1. Inside the Response tab, press Ctrl + F (or Cmd + F on Mac).

2. Type search_model_queries into the search field.

If your prompt triggered a web search, you’ll see something like this:

"metadata": {
  "search_model_queries": {
    "type": "search_model_queries",
    "queries": [
      "top ai content writing tools 2025",
      "best ai copywriting software",
      "ai writing tool comparison 2025"
    ]
  }
}

These are the exact search queries GPT-5 issued to its internal retrieval system or connected search engine.

Step 7: Record and Analyze Your Results

Copy these queries and store them in a spreadsheet or text document.
Over time, as you collect more samples from different prompts, you’ll begin to notice patterns such as:

• The average number of queries per prompt (often 2–3)

• The linguistic simplification (e.g., removing stop words or question phrasing)

• The semantic intent grouping (e.g., “comparison,” “top,” “best,” “alternatives”)

This process reveals how GPT-5 interprets different prompt types — from transactional to informational — and which phrasing it prefers for retrieval.

Why This Insight Matters for SEO/GEO

Being able to see GPT-5’s real queries gives SEOs a direct view into the model’s retrieval mindset — effectively, how AI interprets search intent.
This unlocks several strategic benefits:

1. Understanding AI Search Intent

GPT-5 reformulates prompts into concise, keyword-optimized forms.
By observing this, you learn how LLMs think like search engines — stripping fluff, focusing on entities, and using power terms like “best,” “compare,” or “guide.”

2. AI Visibility Optimization

If you align your on-page content, headings, and FAQs with the kinds of phrases GPT-5 uses, your chances of being cited or retrieved rise substantially.

Example:
If GPT-5 queries “best seo audit tools 2025,” make sure your page explicitly uses that heading or variation.

3. Topic Cluster Planning

Fan-out queries reveal how GPT-5 divides a concept into related searches.
This helps you structure pillar pages and subtopics that reflect the same clustering logic the model uses internally.

4. Authority Signaling

GPT-5 likely favors pages with high relevance and structured clarity.
Pages that match the exact query syntax (with schema, clear H2s, and precise keyword targeting) tend to be more discoverable in AI-driven search.

Practical Application for SEOs

Here’s how you can use this discovery systematically:

---

Tips for Better Analysis

• Prompt diversity: Test different prompt types (questions, comparisons, instructions).

• Categorize intent: Separate informational, navigational, and transactional queries.

• Observe structure: GPT-5 favors concise, lowercase, keyword-dense phrasing.

• Track frequency: See which modifiers (“best,” “top,” “guide,” “2025”) dominate your topic.

• Benchmark competitors: Compare AI-generated queries against SERP results for overlap.

Limitations and Ethical Use

• Not all responses include search_model_queries — only prompts that trigger real-time retrieval.

• This data may change as OpenAI updates internal APIs.

• Avoid automation or scraping beyond personal analysis; stay compliant with platform terms.

• Queries ≠ ranking signals — they reveal AI retrieval intent, not final ranking factors.

Conclusion

The metadata.search_model_queries field is a direct lens into GPT-5’s information retrieval layer — the point where AI meets search.
By uncovering and analyzing these fan-out queries, SEOs can understand:

• How ChatGPT interprets user intent

• Which phrasing patterns the model prefers

• How to align content for AI-driven discovery

In the age of Generative Search, where large language models are becoming the new intermediaries between content and users, this understanding isn’t optional — it’s foundational.

Knowing how GPT-5 rewrites, expands, and fans out queries gives you the power to optimize content not just for Google’s crawler, but for the AI systems that now decide what information gets surfaced.