How ChatGPT Chooses Which Websites to Cite

Two different things people call "ChatGPT knowing about you"

Before getting into crawlers, separate two ways ChatGPT can talk about your business. The first is from its training data: whatever the model learned before its knowledge cutoff. The second is from live search: when ChatGPT searches the web during a conversation and cites the pages it found.

Training data shapes the model's background impression of your brand, and you can only influence it slowly, by being widely and consistently described on the public web. Live search is where citations come from, and it responds much faster to changes you make. Most practical AEO work for ChatGPT targets live search.

The three OpenAI user agents

OpenAI documents separate crawlers for separate jobs, and you control each one in robots.txt. Getting this wrong is the single most common reason a site never appears in ChatGPT citations.

A common setup for brands that do not want their content used for training but do want to be cited:

Also check your CDN and firewall. Cloudflare and similar services have bot management settings that can block AI crawlers regardless of what robots.txt says. We regularly find sites with a perfect robots.txt and a firewall rule quietly returning 403 to every AI bot.

Where the search results come from

ChatGPT search does not work from a single index of its own. OpenAI has said it uses third-party search providers alongside its own crawling, and Microsoft's Bing has long been reported as one of the main providers. The exact mix is not public and can change, so treat any precise claim about it with caution.

The practical takeaway is straightforward. If your important pages are not indexed in Bing, you are making ChatGPT's job harder. Many teams have never opened Bing Webmaster Tools. It takes ten minutes to verify a site there, submit a sitemap and check for crawl errors, and it is one of the cheapest AEO tasks available.

What happens between the question and the citation

1. The user asks something like "best CRM for a 20-person real estate brokerage in Texas".
2. ChatGPT decides whether to search. Time-sensitive, local, comparative or niche questions are more likely to trigger a search.
3. It rewrites the question into one or more search queries, often more specific than what the user typed.
4. It retrieves candidate pages and reads the relevant passages.
5. It writes an answer, citing the pages whose passages it relied on. Pages that merely mention the topic, without a clear answer, tend to be left out.

Step three is easy to overlook. The engine is often searching for something narrower than the user's words: "CRM real estate brokerage small team comparison 2026", for example. Pages that cover a topic with specific sub-sections (pricing, team size, integrations, region) match more of these rewritten queries.

What makes a page citable

Passages that stand on their own

The model quotes passages, not whole pages. A paragraph that starts with "As mentioned above, this is why it matters" cannot be lifted out cleanly. A paragraph that starts with "Most real estate brokerages under 30 agents choose between Follow Up Boss, kvCORE and HubSpot" can.

Specifics over adjectives

Prices, timelines, thresholds, version numbers, named tools, named places. "Pricing starts at $49 per user per month on annual billing" gives the model something to work with. "Flexible pricing for every business" does not.

Readable HTML

If your main content only appears after client-side JavaScript runs, some crawlers will see an empty shell. Server-side rendering or static generation for important pages removes that risk. Check by viewing the raw HTML source, not the browser's rendered DOM.

Agreement with other sources

When ChatGPT recommends a product or company, it is safer for the model to pick one that several independent sources agree on. Comparison articles, review sites, community threads and industry directories all count. Your own site is one voice; you need others saying compatible things.

Freshness where it matters

For topics that change, such as pricing, regulations or "best of" lists, visible updated dates and current information make a difference. A 2023 comparison page competing against a 2026 one is at a disadvantage, even if it was better written at the time.

A checklist you can work through this week

• robots.txt allows OAI-SearchBot on every page you want cited.
• CDN and firewall settings do not block OpenAI user agents (check your server logs for 403s).
• Site verified in Bing Webmaster Tools, sitemap submitted, key pages indexed.
• Key pages render their main content in the initial HTML.
• Each important page answers its main question in the first two sentences.
• Headings are phrased as the questions buyers ask.
• Organization schema with sameAs links to LinkedIn, Crunchbase and other official profiles.
• Pricing, locations, team size and other facts are stated plainly and match across your site and profiles.
• At least a few independent, credible pages describe your company accurately.
• Updated dates are visible on pages with time-sensitive content.

The first four items are technical and can be checked quickly. Our free AI Visibility Score tests crawler access, structure and schema for any URL if you want a starting point. The rest is editorial and reputation work that takes longer but compounds.

How to check whether it is working

Build a list of 30 to 100 questions your buyers ask. Run them in ChatGPT with search enabled, ideally several times each, from a logged-out or clean session. Record whether you are mentioned, whether you are cited, and which URL is cited. Repeat every two to four weeks. Server logs are a second signal: rising OAI-SearchBot and ChatGPT-User requests to your key pages usually come before rising citations.