How ChatGPT Picks Its Sources: A Practical Guide for PR and SEO Teams

If you publish for a Western or China-facing audience, understanding how ChatGPT picks its sources is no longer optional. Its answers now drive referral traffic that Google's organic results never delivered, and the rules it follows are different from the ones most SEO teams have spent a decade learning. Treat this as a practical field note, not a manifesto: here is what is actually documented about the pipeline, what it means for your pages, and what you can control.

How ChatGPT Picks Its Sources, Step by Step

The mechanics are simpler than the hype suggests. When a question needs current information, ChatGPT does not crawl the open web at query time. It sends the query to Bing, reads the ranked results from Bing's index, and then OpenAI's own retrieval crawler fetches the most relevant pages to read them in full. The model synthesises an answer from those pages and attaches clickable source links. So the chain is: your page must be in Bing's index, Bing must rank it well for the rewritten query, and OpenAI's crawler must be allowed to fetch it. Miss any link and you are invisible to live answers, no matter how good the content is.

Why Bing, not Google, feeds ChatGPT's answers

This is the part Western teams miss. OpenAI's web search is built on Microsoft's Bing infrastructure, a consequence of the deep Microsoft investment in OpenAI. The practical upshot is that your Bing presence now has a second layer of value on top of Bing's own traffic. A page that ranks on page two of Bing is less likely to be cited than one on page one, and a page Bing has never indexed simply cannot enter the candidate set. Google ranking is still the larger traffic channel, but it does almost nothing for ChatGPT visibility on its own. The same Bing index also feeds Microsoft Copilot and, in part, DuckDuckGo, so fixing Bing discoverability is a multiplier across several AI surfaces at once.

The three OpenAI crawlers you should know

OpenAI operates more than one crawler, and confusing them is the most common self-inflicted mistake. GPTBot collects public content to train future models; blocking it opts you out of training data but does not stop live citations. OAI-SearchBot is the one that powers ChatGPT Search citations: it fetches pages from Bing results so ChatGPT can quote them with links, and it respects robots.txt. Block OAI-SearchBot and you remove yourself from live citation entirely. ChatGPT-User is triggered only when a user asks ChatGPT to visit a specific URL, and OpenAI notes that robots.txt rules may not apply to those user-initiated fetches. For citation purposes, the lever you control is simple: make sure your robots.txt allows both GPTBot and OAI-SearchBot.

What kinds of pages ChatGPT tends to cite

Looking at what actually gets cited, a few patterns hold. ChatGPT leans on authoritative reference material, big editorial brands, and community discussion, and it favours pages that lead with a direct, extractable answer rather than burying the point under context. Pages that open with a short, factual statement, use clear definition patterns, break processes into numbered steps, and carry concrete data points tend to be quoted more often. Structured data helps too: Article and FAQPage schema give the model clean fields to pull from. This is less about gaming a ranking signal and more about making your facts easy to lift accurately, because the model is extracting specific claims, not just linking to a page.

Structuring content so ChatGPT can extract it

The single highest-leverage habit is to lead with the answer. If the definitive sentence is in the third paragraph, the model may summarise from a weaker sentence or skip you for a competitor who stated it plainly. Use the question as a heading, then answer it in the first sentence beneath. Add definition lines, numbered how-tos, and specific numbers wherever you can. Implement Article and FAQPage schema so the structured version of your page matches the prose. Keep pages focused: a cluster of tightly related articles signals topical authority far better than one bloated pillar page. And remember freshness, because live retrieval is precisely what kicks in for anything dated after the model's training cutoff.

Does llms.txt change anything yet?

llms.txt is a proposed convention for giving crawlers a map of your most important content, and it is worth setting up as a low-cost signal. But be honest about its status: as of now, no mainstream answer engine has committed to honouring it the way search engines honour a sitemap. Writing a glowing article about llms.txt as if it were a ranking lever would be overclaiming. The dependable levers remain Bing indexing, open crawler access, clean extraction-friendly structure, and genuine authority. Treat llms.txt as a tidy extra, not the strategy.

A practical citable-content checklist

Before you publish, walk this list. Is the page indexed by Bing, not just Google? Have you verified that in Bing Webmaster Tools? Does your robots.txt allow GPTBot and OAI-SearchBot? Does the page open with a direct answer to a question a real person asks? Is there schema markup, and do the FAQ items match the prose? Is the content specific and current, with numbers and dates rather than vague claims? If the answer is yes across the board, you have done the things that are actually within your control. Everything else is model behaviour you cannot dictate.

Frequently asked questions

Will ChatGPT cite my company by name? Naming tracks how often your brand appears across the web and in authoritative contexts. Building a consistent body of well-structured, cited-worthy content matters more than any single page.

Is blocking GPTBot a problem for citations? It affects training-data inclusion, not live OAI-SearchBot citations, but many teams allow both to keep every door open.

Does this replace traditional SEO? No. The fundamentals, Bing indexing, authority, structure, and freshness, are exactly what AI retrieval rewards. You are doing the same work with one more consumer of it.

The closing point is the one worth remembering: ChatGPT does not read the whole internet live, and it does not use Google. It reads Bing's index through OpenAI's crawlers, and it quotes the pages that state answers cleanly. Services that place announcements across Chinese and global outlets, including CN NEWSWIRE, normally submit sitemaps to Bing and Baidu so the content is crawlable before any answer engine sees it, which is the unglamorous prerequisite every citation depends on.