How does ChatGPT choose sources for a search answer?
ChatGPT search can show links to sources when it uses web information, but a publisher does not control which page a response cites. The answer depends on the question, retrieval process and available material at that moment. A brand mention without a link, a linked citation to your page and a referral visit are different events. Before changing content, decide which one you intend to measure. Our AI visibility audit uses a repeatable prompt log to keep them separate.
For B2B teams, the useful question is whether a page contains facts a buyer needs to verify: the service scope, product requirements, methodology, price or a documented result. A thin article that restates a popular definition contributes little distinctive evidence. A detailed integration page or fair comparison may be more useful to a technical evaluator than a broad “best practices” post. Write for that evaluator first, and let the citation be a possible distribution outcome.
What OpenAI says about source links
OpenAI's ChatGPT search guidance explains that answers can include source links. That does not mean every source used will be linked in every answer, or that a publisher can reserve a place. Do not treat a screenshot of one answer as an enduring ranking. Save the prompt wording and date so a later review can compare the same question under similar conditions.
Can OpenAI's search crawler reach your pages?
OpenAI's crawler documentation distinguishes OAI-SearchBot, used for search, from GPTBot, used for possible model training. The controls are independent: a site can permit search discovery while restricting training. OpenAI also describes ChatGPT-User as a user-triggered fetcher, which is not the same as automatic search crawling. Check the actual robots.txt, CDN and firewall behaviour rather than assuming one “AI bot” setting controls all three.
A crawlable URL is a starting condition, not a citation promise. Make sure the important page returns a usable response, its main facts are available as text, internal links lead to it and the content is current. If the site uses a challenge page for all bot traffic, inspect server logs or run a controlled fetch test. A successful browser visit by a human does not prove the search crawler can read the page.
The AI crawler policy guide separates OAI-SearchBot, GPTBot and llms.txt into different decisions.
What information is worth citing on a B2B site?
Describe the thing your company uniquely knows and can support with evidence. A service page can explain the precise scope and workflow. A product page can list supported integrations and limitations. A case study can show baseline, work, period and measured result with permission. An expert guide can explain a decision using a worked example and cite primary sources for external facts. These details are useful even if an answer engine never links to them.
| Page type | Useful original detail | Review owner |
|---|---|---|
| Service | Scope, deliverable and exclusions | Delivery lead |
| Product | Supported version, setup and limits | Product manager |
| Comparison | Dated criteria and tradeoffs | Product and legal reviewers |
| Case study | Baseline, method, period and permission | Account owner |
| Expert guide | Worked decision and cited facts | Named specialist |
This is an editorial evidence matrix, not a claimed ranking factor. It helps the team avoid creating a hundred similar pages with no additional value. It also creates an update queue when a product, policy or price changes. A page that is clear about what the company cannot do may be more credible to a procurement team than one that claims to solve every problem.
How can independent sources support expertise?
Expertise should be verifiable outside the company site where possible. Relevant conference talks, standards contributions, customer reviews and editorial mentions can give a buyer context. Seek them because the work merits attention, not as a purchased citation tactic. Link to the original source and keep the claim modest. A B2B buyer can distinguish a named contribution from a list of logos with no explanation.
For specialist content, name the author or reviewer only when that person actually checked the material. Provide an author page with experience that can be verified. Do not invent qualifications or first-hand work to make a page sound authoritative. If a customer cannot be named, explain the limits of the anonymised evidence. B2B content marketing works best when it turns real subject expertise into buyer-facing answers rather than producing volume for its own sake.
How should you test whether the work helped?
Build a small set of prompts from real sales questions. Include unbranded discovery, category comparison, product-fit and branded verification questions. Run the same set in the same product mode, record the date and save the visible answer. Count whether the company was named, whether a page was linked, whether the description was correct and whether any referral visit or enquiry followed. Keep the denominator visible; a percentage from ten handpicked prompts is not market share.
Then inspect what can be improved. An inaccurate description may reveal unclear product copy or an outdated third-party page. An absent citation may reveal a crawl problem, weak relevance or simply that another source was selected. Do not infer the cause from one result. OpenAI says a robots.txt change can take roughly 24 hours to affect its systems, but it does not promise that the next answer will cite the page. Recheck the AI search service page and the current provider documentation when auditing the site.
What belongs in the prompt log?
Record the buyer persona, question, exact prompt, product and mode, date, location context, answer text, named brands, linked URLs and whether the answer was accurate. Keep screenshots or exports where the product permits them. For a follow-up run, use the same prompts first; add new prompts in a separate group so the denominator does not silently change. This makes a visibility trend more interpretable, although it still measures your selected sample rather than all buyer conversations.
How should you treat a new citation?
Open the cited URL and verify that it supports the statement the answer made. A citation to an old version of a pricing page can be harmful even when it increases apparent visibility. Fix inaccurate source material, check redirects and update public documentation. If the answer cites another site that misdescribes your product, seek a correction through that publisher's normal process rather than rewriting your own page to imitate the error. The commercial goal is an informed buyer, not a high count of links on a dashboard.
Set a review owner for the prompt log. Marketing can record what appeared, but a product or service specialist should judge whether the answer describes the offer correctly. Sales can say whether a cited page helped a real prospect. Those three views prevent a citation count from becoming a proxy for value. Keep the test small enough that the team can review each result, not just a score exported by a platform.