How ChatGPT Search Chooses Sources: The Mechanics Explained

When ChatGPT answers a question with live web results, it does not pick sources at random, and it does not read the whole web on the spot. Understanding how ChatGPT search chooses sources comes down to a handful of mechanics: which crawler it uses, where its search results come from, how it rewrites questions into searches, and which pages it can quote cleanly. This guide walks through each step, explains what site owners can and cannot control, and shows how to measure the traffic that comes back.

If you want the practical marketing checklist (reviews, directories, consistent business details), start with our guide on how to get your business mentioned by ChatGPT. This article goes underneath that advice and explains why it works.

The two routes into a ChatGPT answer

A model can mention your website in two different ways, and only one of them is a citation:

1. From training data. The model learned general knowledge from text gathered before its cut-off. It may know your brand, but it cannot link to a specific page from memory, and you cannot update it.
2. From a live search. When ChatGPT decides a question needs current information, it runs searches, reads the results and cites pages with clickable links. This is the only route that sends referral traffic, and it is the one this guide covers.

Other assistants work on the same principle. Google says AI Overviews and AI Mode use "query fan-out", issuing multiple related searches across subtopics, and only show pages that are already "indexed and eligible to be shown in Google Search with a snippet". The details differ, but the pipeline is broadly the same: rewrite, retrieve, read, cite.

Step 1: ChatGPT rewrites the question into searches

ChatGPT does not send the user's prompt to a search engine word for word. OpenAI's help article on ChatGPT search explains that when it works with search partners, it "typically rewrites your query into one or more targeted queries". OpenAI's own example turns a question about cancer drugs into the search "CCR8 immunotherapy drug development 2025", and then sends more specific follow-up queries after reviewing the first results.

Three things follow from that example:

• Your page competes for the rewritten query, not the original prompt. Long conversational prompts become short, specific search strings. Pages that use the precise terms people search for (product names, service names, places, model numbers) match those strings better than pages written in vague marketing language.
• Freshness is built into the query. The rewritten example includes a year. When the search itself asks for recent material, dated, recently updated pages have an obvious advantage over undated ones.
• Context shapes the search. OpenAI says approximate location from the user's IP address may be included, and saved memories can refine the query. A user in Manchester asking for "a good accountant" may trigger a search with a location attached, which is why clear location pages matter for local businesses.

Step 2: Results come from search partners and OpenAI's own crawler

When OpenAI launched ChatGPT search in October 2024, it said the feature "leverages third-party search providers, as well as content provided directly by our partners". Microsoft announced in May 2023 that Bing would be the default search experience inside ChatGPT, and being indexed by Bing remains one of the most sensible things you can do for ChatGPT visibility, even though OpenAI does not publish the full list of providers it uses today.

Alongside those providers, OpenAI runs its own crawlers. Their roles are documented:

OpenAI says each robots.txt setting works independently, and that "it can take ~24 hours from a site's robots.txt update" for search to adjust. If OAI-SearchBot is blocked, a strong Bing ranking will not get you a proper ChatGPT citation, because ChatGPT cannot fetch and summarise the page.

Why Bing still matters

In practice there are two gates. The page must be discoverable through the search results ChatGPT retrieves, and it must be readable by OAI-SearchBot. Bing covers the first gate for a large share of queries, which is why our Bing Webmaster Tools and IndexNow guide is worth reading alongside this one. Submitting your sitemap to Bing and using IndexNow to notify it of new or changed URLs helps your pages enter that pool sooner.

Step 3: The model reads the pages and decides what to quote

Once ChatGPT has a shortlist of results, the model reads the retrieved text and writes an answer, attaching citations to the claims they support. OpenAI does not publish a ranking formula, so be wary of anyone who claims to know the weighting. What we can say from the documented pipeline is which pages are easy to use and which are not.

Pages that are easy to cite tend to have:

• A direct answer near the top of the page, in a sentence that still makes sense when quoted on its own.
• Specific, checkable facts: prices or price ranges, opening hours, service areas, specifications, dates.
• Clear headings that match the question, so the relevant passage is easy to locate.
• Text that is present in the HTML. Content that only appears after JavaScript runs may not be seen by every crawler.
• A visible "last updated" date where freshness matters.

Pages that are hard to cite tend to:

• Bury the answer under long introductions or sliders.
• Use images of text, PDFs without a web version, or tabs that load content on click.
• Say different things on different pages, so the model cannot tell which figure is current.
• Restrict snippets. OpenAI's publisher guidance says summaries and snippets depend on OAI-SearchBot access, and Google's AI features honour nosnippet and max-snippet.

Structured data such as Organization, LocalBusiness or Product is not a ranking switch for ChatGPT, but it states your key facts in a consistent, machine-readable form. Our structured data guide explains which types are worth adding.

What site owners can control

You cannot buy or request a place in a ChatGPT answer. You can control whether you are eligible, and how easy you are to quote.

1. Crawler access

Check robots.txt first. A sensible policy for a business that wants citations but would rather not contribute to model training:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

User-agent: Bingbot
Allow: /

Then check your CDN and security plugins. Bot protection on Cloudflare, some WordPress security plugins and hosting firewalls can block AI crawlers at the network level, which robots.txt will never show. OpenAI publishes IP ranges for OAI-SearchBot at openai.com/searchbot.json if you need to allowlist it.

2. Keeping pages out deliberately

If there are pages you do not want summarised, such as internal price lists or old campaigns, robots.txt is not the right tool on its own. OpenAI's publisher FAQ notes that if a disallowed page is known to be relevant, ChatGPT may still show its link and title. To keep a page out, use a noindex meta tag and allow the crawler to read it, because a crawler that is blocked cannot see the tag.

3. Content built for retrieval

Write each important page so that it answers one clear question well. Use the words your customers use, include the place names you serve, put the answer in the first paragraph and keep facts consistent across your site, Google Business Profile, Bing Places and directories.

4. Speed and reliability

A crawler that times out or receives errors cannot cite you. Slow servers, 5xx errors and redirect chains quietly remove pages from consideration. Run a free website audit to check status codes, redirects, robots.txt, your sitemap and page speed in one scan.

You can also run our AI crawler checker (free account required). It reports what your robots.txt allows for Googlebot, Bingbot and each AI crawler, flags snippet controls such as nosnippet, measures how much text is in the HTML without JavaScript, and checks for Organization or LocalBusiness schema and an llms.txt file.

How to measure ChatGPT referrals

Measurement is patchy, but it is better than it was.

Google Analytics 4 and other analytics tools

OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral links, which makes ChatGPT clicks easy to separate. In GA4:

1. Go to Reports > Acquisition > Traffic acquisition.
2. Change the primary dimension to Session source / medium.
3. Search for chatgpt.com.
4. Add a comparison for landing page to see which pages ChatGPT sends people to.

Some visits arrive without the tag, for example when someone copies a link from an answer. Those show up as direct traffic, so treat your ChatGPT figure as a minimum.

Server logs

Your access logs show OAI-SearchBot requests. A rising crawl rate on a page is a reasonable sign it is being considered for answers. Look for the user agent string:

grep "OAI-SearchBot" /var/log/nginx/access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

Bing Webmaster Tools AI Performance

Microsoft added an AI Performance report to Bing Webmaster Tools as a public preview in February 2026. It shows how often your pages are cited in Microsoft Copilot, AI summaries in Bing and "select partner integrations", with total citations, cited pages and sample grounding queries. In June 2026 Bing added intents, topics, citation share and comparison views. It has no click data, and Bing says the figures are a sample, but it is the closest thing to a citation report any search engine offers for free.

Google Search Console

Google counts traffic from AI Overviews and AI Mode in the normal Performance report under the Web search type. You cannot isolate AI clicks there, but our Search Console dashboard makes it easier to watch for pages where impressions rise while clicks fall.

Frequently asked questions

Does ChatGPT search use Bing?

OpenAI says ChatGPT search uses third-party search providers alongside content from partners, and Microsoft announced Bing as ChatGPT's default search experience in 2023. OpenAI does not publish its current provider list, but being well indexed in Bing remains one of the most reliable ways to enter the pool of pages ChatGPT can cite.

If I block GPTBot, will I disappear from ChatGPT search?

No. GPTBot gathers training data. ChatGPT search relies on OAI-SearchBot, and OpenAI says each robots.txt setting works independently, so you can block GPTBot and allow OAI-SearchBot.

How long after changing robots.txt will ChatGPT notice?

OpenAI says it can take about 24 hours for its search systems to adjust after a robots.txt update. Being cited still depends on your pages being found and judged relevant, which can take longer.

Does ChatGPT prefer newer content?

OpenAI does not publish ranking factors, but its own example of a rewritten search query includes the current year, which favours recent material. Showing a clear publish or update date and keeping facts current is sensible for any time-sensitive topic.

How can I see traffic from ChatGPT in Google Analytics?

ChatGPT adds utm_source=chatgpt.com to referral links. Filter the Traffic acquisition report by session source for chatgpt.com, and remember that copied links can appear as direct traffic instead.

Can I pay to be cited in ChatGPT answers?

There is no paid route into search citations. OpenAI does now run ads in ChatGPT, and checks ad landing pages with a separate crawler, OAI-AdsBot, but an ad is not a citation in an answer.

Sources

• OpenAI: Overview of OpenAI crawlers
• OpenAI Help Center: ChatGPT search
• OpenAI Help Center: Publishers and developers FAQ
• OpenAI: Introducing ChatGPT search (October 2024)
• CNBC: Microsoft says Bing can be default search engine for ChatGPT users (May 2023)
• Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools (February 2026)
• Bing Search Blog: New AI visibility insights in Bing Webmaster Tools (June 2026)
• Google Search Central: AI features and your website