How ChatGPT Search Finds and Chooses Its Sources

When does ChatGPT use live search vs memory?

ChatGPT decides, per question, whether it needs current information or can answer from what it already learned during training. A question like "what's a popular project management tool" can be answered from training data alone — no search required, no sources to choose. A question like "what's the current pricing for [tool]" or anything time-sensitive, local, or about something that changed after the model's training cutoff typically triggers a live web search.

This matters because only the search path involves source selection at all. If ChatGPT answers from memory, there's no crawl, no ranking, no citation — just whatever the model learned, which can be outdated, approximate, or simply wrong about specifics like pricing or features. If it searches, it runs a retrieval step, pulls candidate pages, and picks what to cite. Those are two different failure modes for a brand: being absent from a trained-in understanding, versus being absent from (or miscited in) a specific search result.

What do OAI-SearchBot, GPTBot and ChatGPT-User each do?

Three separate bots, three separate jobs, each controlled independently in robots.txt:

• OAI-SearchBot — crawls and indexes pages so ChatGPT's search feature can surface and cite them in answers. This is the one that matters for "will ChatGPT find and cite my page."
• GPTBot — crawls content that may be used to train future OpenAI models. It has nothing to do with search or citations.
• ChatGPT-User — fetches a single page live when someone asks ChatGPT to look at or act on a specific URL. OpenAI's own documentation now says robots.txt rules "may not apply" to it, since the fetch is triggered by a person, not an automated crawl.

The common mistake is treating GPTBot as a stand-in for "ChatGPT" generally and blocking it expecting to disappear from search answers — it won't do that, because OAI-SearchBot is the one actually powering citations. We cover this split, plus the equivalent bots for Anthropic and Perplexity, in full in Blocking GPTBot Isn't Blocking ChatGPT; this post focuses on how the search path itself works once a page is crawled.

How much does ChatGPT depend on Bing?

More than most brands assume, though the picture is dated and incomplete. OpenAI built ChatGPT's search feature on top of Microsoft's Bing index — a direct consequence of Microsoft's investment in and technical partnership with OpenAI. A Seer Interactive study from February 2025 (about 100 questions, 500-plus citations, looking at what was then called "SearchGPT") found that 87% of its citations matched a result somewhere in Bing's top 20 for the same question, versus 56% for Google's top 20, with Google matches landing at a median rank of 17. That's a meaningfully stronger correlation with Bing than with Google.

A caveat the study's own authors flagged directly: it's a roughly 100-question sample from February 2025, before ChatGPT search rolled out broadly, and the authors described it as a preliminary result — their own write-up noted they were already expanding the test to 10,000 queries. So "ChatGPT runs on Bing" was a strong, well-evidenced read in early 2025. Whether that correlation holds at the same strength by late 2026 is a fair question — OpenAI has kept developing ChatGPT search since — but nobody outside OpenAI has published fresher, larger-sample data to confirm one way or the other.

A separate, more recent Ahrefs study (August 2025, 15,000 long-tail queries) measured something adjacent but different: how often links cited by ChatGPT, Gemini and Copilot also rank in Google's own top 10. It found only about 8% overlap for ChatGPT specifically — consistent with the idea that ChatGPT isn't simply reflecting Google's rankings, even if it leans on Bing's index to retrieve candidates.

What does "visible in ChatGPT" actually mean?

It splits into layers that don't move together:

• Named — ChatGPT mentions your brand in the answer text, with or without a link.
• Cited — a source from your site (or about you) is linked as a reference.
• Recommended — the answer actively positions you as a good fit for the question, not just a name in a list.

A brand can be named without being cited, cited without being recommended, or recommended with the underlying source actually belonging to someone else. These aren't the same event, and conflating them is where most "are we visible in ChatGPT" checks go wrong. We define and illustrate all three with real examples in Mentions, Citations and Recommendations Are Three Different Things — treat this section as the short version.

Here's a real one, run live for this post. Asked a neutral buyer question — "what are the best project management tools for small teams?" — the answer named seven tools: five as top picks (monday.com, Asana, ClickUp, Trello, Wrike) plus Basecamp and a budget-focused pick, Plaky. Checking each one against the actual linked sources turned up three different patterns in a single answer: monday.com, Asana, ClickUp, Trello and Basecamp were named and recommended with no citation to their own site at all — every linked source was a third party. Wrike and Plaky were also cited, but only to their own blogs, which had themselves written "best of" roundups that ranked for the same query — a brand's own content citing itself, not an independent source vouching for it.

Evidence: AI engine (Claude, web search enabled), prompt: "What are the best project management tools for small teams?", 10 October 2026, logged-in research session. Named, not cited at all: monday.com, Asana, ClickUp, Trello, Basecamp. Named and cited, but only to the brand's own blog: Wrike (wrike.com's own "21 top picks" post), Plaky (plaky.com's own "best for small teams" post, the top-ranked source overall). No brand in this answer was named, recommended, and cited to a genuinely independent third party, all three at once.

How should a brand check its own presence?

Ask real buyer-style questions, without your own brand name in the prompt, in a fresh chat with search turned on — not a chat with prior context, which can bias the answer toward things you've already discussed. Do this on more than one day, since the same prompt on the same engine doesn't reliably return the same answer twice (see how many runs you actually need before treating any single check as settled). Log, per answer: named or not, cited or not, recommended or not, and whether the facts stated about you are accurate.

If you want a fast, single check rather than a manual log, the AI Citation Snapshot runs one real question through one engine and shows the live answer plus its cited sources. It's one run — a first look, not an audit.

FAQ

Does ChatGPT always search the web before answering? No. It decides per question whether the request needs current or specific information. Many general questions get answered from training data alone, with no search and nothing to cite.

Is GPTBot the bot that controls whether ChatGPT cites my site? No. GPTBot only controls training-data crawling. OAI-SearchBot is the one that crawls and indexes pages for ChatGPT's search and citation feature; they're independent settings in robots.txt.

Does ChatGPT just copy Google's rankings? No — the evidence points the other way. An Ahrefs study found only about 8% of ChatGPT's citations also rank in Google's top 10 for the same query. ChatGPT's search leans more on Bing's index, per the Seer Interactive data above, though that data is from February 2025 and may not reflect the current pipeline.

Can I tell whether a specific ChatGPT answer used live search or training memory? Not reliably from the outside. If the answer includes linked citations, search ran. If it doesn't, it may still have searched without surfacing links, or it may have answered from memory — OpenAI hasn't published a way to distinguish these from the user side.

What to do next

1. Check whether OAI-SearchBot can currently crawl your site using the AI Bot Access Checker — this is the bot that has to succeed before citation is even possible.
2. Ask five buyer-style questions in fresh, search-enabled ChatGPT chats, without naming your brand, and log named/cited/recommended for each.
3. Repeat on a different day before drawing any conclusion — one run tells you what happened once, not what's typical.

ContextOwl sells one-time AI visibility reports. The AI Citation Snapshot is a free, single-question version of the same check.