ChatGPT ?hints Parameter: See What ChatGPT Actually Searches For

ChatGPT doesn't just answer from memory. About a third of the time, it searches the web - and the queries it generates internally look nothing like what the user typed. An undocumented URL parameter called ?hints=search lets you see exactly what ChatGPT searches for and which sources it cites.

The insights from that data are more valuable than the parameter itself. ChatGPT's fan-out queries reveal a hidden layer of commercial, comparison-heavy search intent that doesn't exist in Google keyword databases. Companies using this data to find content gaps are seeing AI visibility jumps from single digits to double digits in under three months.

What Is the ?hints Parameter and Why Does It Matter?

The ?hints=search URL parameter triggers ChatGPT's web search mode directly from a URL. Append it to chatgpt.com with a query and ChatGPT will search the web, show its sources, and cite them in the response.

The URL format:

https://chatgpt.com/?hints=search&q=best crm for startups 2026

This works for any ChatGPT Plus user as of April 2026. OpenAI has never officially documented it - the parameter emerged from community discovery by Jerome Salomon in October 2025 and has been validated by commercial AI visibility platforms since. It's a UI hint that opens browsing mode, not a backend API override, but it reliably triggers search.

The real value isn't the parameter itself. It's what it exposes.

When ChatGPT searches, it doesn't send your exact prompt to the web. It rewrites your question into multiple internal sub-queries called "fan-out queries" (QFOs), retrieves candidate pages, and selects which ones to cite. The ?hints parameter lets you see all of this - the internal queries, the sources considered, the sources actually cited - through ChatGPT's conversation JSON export. That data reveals how ChatGPT discovers and selects content, which is the foundation for any serious AI visibility strategy.

The critical caveat: The parameter forces a 100% search rate. ChatGPT's natural search rate is about 31-34.5% of prompts (Nectiv baseline, Semrush Feb 2026 clickstream data). So you're seeing the ceiling of ChatGPT's search behavior, not average behavior. For competitive analysis this is useful - you're mapping what ChatGPT would find if it searched for your topic. For market share estimates, you need to account for the baseline.

Other useful parameters you can combine:

For more on how AI engines select sources generally, see How AI Answer Engines Actually Select Content.

What Do ChatGPT's Internal Search Queries Look Like?

ChatGPT's fan-out queries are longer, more specific, and more commercially oriented than what users type - and they don't match Google keyword patterns.

Nectiv's study of 8,500+ prompts (the first large-scale analysis of ChatGPT's internal search behavior) found that when ChatGPT searches, it generates an average of 2.17 sub-queries per prompt. Those queries average 5.48 words - roughly 60% longer than Google's average 3.4-word query. 77% of fan-out queries exceeded 5 words.

The queries follow a consistent pattern. ChatGPT strips filler, promotes entities (brand names, product categories), and aggressively injects commercial modifiers. A user prompt like "What's a good project management tool for remote teams?" becomes QFOs like:

• "best project management software remote teams 2026"
• "project management tools remote collaboration features comparison"
• "top rated project management apps small business reviews"

Peec.ai's analysis of 20 million QFOs (published April 2026) found that QFO word count roughly doubled between October 2025 and January 2026, while the number of QFOs per prompt stayed flat. ChatGPT is getting more precise with each query, not running more of them.

Semrush's clickstream analysis of tens of millions of ChatGPT sessions reveals the scale of this divergence from Google: 65-85% of ChatGPT prompts cannot be matched to any keyword in Semrush's 27-billion-term search database. About 70% represent "entirely new types of search intent" not seen in classical search logs - tasks, scenarios, and multi-step comparisons rather than keyword lookups.

Search rates vary dramatically by category. Nectiv's vertical breakdown:

If your industry triggers high search rates, optimizing for ChatGPT's web search layer matters significantly. If search rates are low, you need cross-web brand presence to influence training data instead.

How Do You Extract This Data?

Three methods, ranging from zero-code to enterprise-scale.

Method 1: Browser + JSON export (free, manual)

1. Open https://chatgpt.com/?hints=search&q=[your query] in your browser
2. ChatGPT searches and responds with citations
3. Open Chrome DevTools (F12) > Network tab, find the conversation JSON
4. Look for metadata.search_model_queries - this array contains every internal query ChatGPT ran
5. Look for sources_footnote (main citations), supporting_websites (considered but not cited), and safe_urls

Alternatively, use ChatGPT's built-in export (Settings > Data Controls > Export Data) to get conversation JSON in bulk.

Method 2: Bookmarklets and extensions (free, semi-automated)

JC Chouinard's ChatGPT Query Fanout Analyzer is a bookmarklet that fetches the active conversation's JSON and renders a table of all internal queries, cited URLs, and entities. One click, no code needed.

Chrome extensions like ChatGPT Internal Query Extractor and ChatGPT Sources & Citations Exporter automate the same extraction and export to JSON or CSV.

Method 3: API or commercial platforms (paid, scalable)

OpenAI's official Responses API supports a web_search_preview tool that returns structured citation data (URLs, titles, character indices) - fully documented and stable, unlike the UI parameter. For teams running hundreds of prompts, commercial platforms like Nectiv AI Tracker and Spotlight handle extraction at scale.

Rate limits: Plan for about 100-150 search-triggering prompts per browser session before rotating sessions or user agents. The parameter is rate-limited by standard session throttling, not a hard cap. For larger runs, the API is more reliable.

What Does ChatGPT's Source Selection Reveal?

ChatGPT's citation patterns follow clear, measurable rules that differ from Google's ranking signals.

Domain authority is a hard gate. Ahrefs' analysis of ChatGPT's top 1,000 cited pages found that 65.3% sit on domains with DR 81+, with a median domain rating of 90. But here's the surprise: page-level authority barely matters. The median URL Rating for those top-cited pages was just 6. ChatGPT trusts domains, not individual pages. If your domain passes the trust threshold, even new pages with no backlinks can get cited.

Freshness is the strongest content-level signal. SE Ranking and ALM Corp data shows content updated within 30 days receives about 3.2x more citations than content older than 90 days. Among Ahrefs' top-cited URLs, 76.4% were updated within the last 30 days. Monthly content refreshes are associated with roughly 23% higher AI coverage.

Position on the page matters. ALM Corp's analysis of 1.2M ChatGPT responses found that 44.2% of citations come from the first 30% of content. Conclusions at the bottom of articles are largely ignored - the bottom 10% contributes just 2.4-4.4% of citations.

Format preferences are consistent: Listicles and "best X" comparisons dominate commercial citations. FAQ sections and structured Q&A earn ~40% more citation weight than unstructured text. Pages above ~2,900 words are 59% more likely to be cited than those under 800 words.

Review platforms amplify trust. Domains with active profiles on G2, Trustpilot, Capterra, or Yelp have roughly 3x higher citation probability. ChatGPT appears to treat review platform presence as a "real brand" verification signal.

For the metadata layer that complements these content signals, see Schema Markup for AI Citations. For body-text structure that improves extraction, see How to Structure Content That AI Systems Actually Cite.

How Do You Turn Fan-Out Data Into Content Strategy?

Run 100 representative prompts for your industry through ?hints=search. Extract all QFOs and cited sources. Then analyze four dimensions.

1. Modifiers and patterns. Tally the most frequent qualifiers in the QFOs: "best," "top," "2026," "for startups," "for enterprise," "pricing," "integrations," "alternatives," "vs." This gives you the commercial and segment breakdown ChatGPT uses behind the scenes. If QFOs consistently include "for mid-market" but your content only addresses "enterprise," you've found a targeting gap.

2. Task clusters. Group QFOs into decision stages: evaluation ("best X for Y"), comparison ("X vs Y"), implementation ("how to set up X with Z"), risk ("limitations," "downsides"), and optimization ("best practices for X"). Each cluster should map to at least one strong content asset.

3. Entity coverage. Check which brands appear in QFOs and citations and which don't. If your brand never shows up in QFOs around your own category, you likely have weak entity clarity on-site or insufficient cross-web presence (few review profiles, media mentions, or community references).

4. Source archetypes. Classify cited URLs: reference sites, media, review platforms, competitor content, community threads. This shows where ChatGPT currently "sees" your category. If most citations come from G2 and Forbes, that's where your authority-building should focus.

Case studies show this works. A B2B SaaS company tracked by Dataslayer discovered through AI logs that ChatGPT recommended competitors for "integrations with Salesforce" prompts despite having strong Salesforce support. They published a detailed integration guide and started appearing in AI answers within 30 days.

A fintech company reported by Mersel AI implemented structured "answer objects" across core pages, coordinated with review and media signals. AI visibility rose from 2.4% to 12.9% across tracked prompts in 92 days, generating 94 citations and contributing to about 20% of demo requests.

Search Engine Journal covered a new business running weekly AI visibility sprints. After six weeks: visibility in 16.5% of relevant AI responses, 39 of 150 targeted questions, 74 mentions with 42 cited. From zero to measurable AI presence in under two months.

CompetLab's AI Visibility tracking monitors how ChatGPT, Claude, and Gemini mention your brand across queries - so you can see whether your content changes actually move citation rates, not just hope they do. To set up baseline measurement, see How to Measure Your AI Visibility.