ChatGPT's source-selection process is no longer a black box. A detailed technical analysis reveals that the AI now exposes which websites competed for citations but lost, creating a new competitive intelligence tool for content creators and marketers. Additionally, OpenAI has begun testing Bing as a fifth data retrieval source for select user accounts, marking a significant shift in how ChatGPT gathers information.

What Changed in ChatGPT's Source Handling?

For months, ChatGPT's source citations appeared final and opaque. When the AI cited a webpage to support a claim, users saw only the winner. Now, ChatGPT's backend reveals a new field called "supporting_websites" that displays runner-up pages for every claim, complete with their own source attribution labels.

This shift happened quietly in late June 2026, and the implications are substantial. When ChatGPT answers a question about, say, Emirates baggage rules, the cited emirates.com page now carries metadata showing which other Emirates pages competed for that same citation slot but lost. Similarly, when multiple websites claim the same fact, ChatGPT now exposes which competitor's page won the visible slot and which ones sit invisibly beneath it.

The runner-up data comes in two distinct patterns. The first involves what researchers call the "domain fold," where multiple pages from the same website compete, and only one wins the visible citation. A Range Rover overview page, for example, won a specification claim while the brand's own electric-range page was demoted to supporting status. The second pattern is purely competitive: when three different websites claim the same fact, one appears in the citation, and the other two are logged as supporting evidence.

How Is OpenAI Testing Bing as a Data Source?

Parallel to the runner-up reveal, OpenAI has introduced Bing as a fifth retrieval pipeline for ChatGPT, though the rollout is highly selective. A researcher named David Konitzny discovered Bing appearing in his network traffic with a "result_source" label of "bing," a value that had never appeared before. When the same URL was tested across different user accounts in the same week, one account showed Bing as the source while another showed Bright Data, a web scraping service.

The Bing integration is cohort-gated, meaning OpenAI is testing it with specific user groups rather than rolling it out universally. A census of 30 recent conversations showed Bright Data appearing 558 times, Labrador 21 times, SERP 16 times, and Bing zero times in one account, while other accounts showed Bing traffic. This suggests OpenAI is running A/B tests to evaluate Bing's reliability and usefulness before wider deployment.

What Does This Mean for Content Strategy?

The transparency of runner-up data creates a new competitive audit tool that paid SEO platforms do not yet offer. Content creators can now study exactly how winning pages phrase claims compared to losing ones, down to specific word choice. When a competitor's page wins a citation for a particular claim, the runner-up data shows the exact language that succeeded.

For websites concerned about Bing indexation, the new Bing source integration adds urgency. If a user account falls into a Bing cohort, pages missing from Bing's index will be invisible to ChatGPT for those users. This makes Bing Webmaster Tools and sitemap verification essential maintenance tasks, though the feature could be withdrawn at any time since it remains experimental.

Steps to Optimize for ChatGPT's New Source Selection

• Consolidate Thin Content: If 20 pages cover the same topic, 19 of them will lose to the strongest one. Build one comprehensive page per claim rather than multiple weak pages competing against each other.
• Verify Bing Indexation: Check Bing Webmaster Tools and submit sitemaps to ensure your pages appear in Bing's index, since the AI is now testing Bing as a data source for some users.
• Study Runner-Up Data: Pull your own ChatGPT conversations and examine how winning pages phrase factual claims. Use this per-claim competitive analysis to refine your own language and structure.
• Own Factual Claims: Publish pricing, specifications, and documentation in plain HTML on pages that are unmistakably yours. Facts route to official sources, so position your brand as the authoritative source for your own data.
• Avoid Vague Language: Instead of "Contact sales" for pricing, state "Starts at $5000." ChatGPT's source-filtering prompt appears to favor specific, official information over placeholder text.

The Hidden Instruction Layer Behind Source Selection

Beneath ChatGPT's visible behavior lies a system-level instruction that OpenAI calls "sources_and_filters_prompt." This hidden prompt contains the actual rules for how ChatGPT selects and ranks sources, but the text never reaches users' browsers. OpenAI strips the content server-side, sending only the metadata envelope.

However, the prompt's behavior is readable in the open. Across 30 conversations, the word "official" appeared 17 times appended to factual queries, suggesting the hidden prompt instructs ChatGPT to seek official sources for facts. Opinion-based claims route to reviews and Reddit, while factual claims route to official pages. This pattern reveals the memo's logic without exposing its exact wording.

The implication is significant: source selection is a policy, not a fixed model behavior. OpenAI can rewrite these instructions any day without retraining the model or announcing the change. This is why the plumbing shifted within 10 days of the initial runner-up reveal, and why content strategies must remain adaptive rather than static.

For marketers and SEO professionals, the lesson is clear. Nobody can honestly claim to know ChatGPT's ranking factors with certainty, because the real instructions live in a sealed prompt on OpenAI's servers. What you can do is watch the behavior, treat each pattern as evidence of the underlying policy, and update your playbook when the behavior changes.