ChatGPT Search Risks, Limits and Governance Rules to Set First

OpenAI added search to ChatGPT so the assistant can look up recent information instead of relying only on its training data. When a question needs fresh facts, ChatGPT runs a search, reads a set of pages and writes an answer with links to the sources it used. Users can also trigger it on purpose from the search option in the message box.

The answer is a synthesis, not a list of ten blue links. That makes it fast to read, but it also means the model decides which sources matter and how to phrase them. Users see the conclusion first and the evidence second, which changes how people verify information.

The risks fall into three groups: accuracy, data protection and dependency. Each needs its own rule, because a single “use AI carefully” policy is too vague to follow.

• Accuracy: the summary may merge two sources, drop a qualifier or cite a page that does not say what the answer claims
• Freshness: some topics move faster than the index, so prices, regulations or product specs may be out of date
• Source quality: forum posts, affiliate pages and official documents can be weighted similarly in one answer
• Confidential input: a prompt such as “compare our unreleased pricing with competitor X” exposes internal data
• Dependency: staff stop checking primary sources once answers look reliable

It is weak where the question depends on exact wording, private context or legal precision. A tax rate, a contract clause or a medical dosage needs the primary source, read directly. A summary that is 95 percent correct is not good enough there.

It also cannot see paywalled or login-protected material in most cases, and it does not know your internal systems unless connected to them. Research on niche B2B topics can come back thin, padded with general pages because better sources are not reachable.

Citations are not verification

A link next to a sentence feels like proof, but it only shows where the model looked. Teams should click through for any claim that will be repeated in client work, published or used in a decision. In spot checks it is common to find a cited page that supports only part of the claim.

Write rules that a busy employee can remember. Five or six rules on one page work better than a long policy nobody reads. A practical starting set:

• Approved uses: early research, competitor overviews, finding sources, drafting questions for experts
• Banned uses: final legal, financial or medical answers without a qualified human check
• Data rule: no client names, personal data, unreleased figures or contract text in prompts on consumer accounts
• Verification rule: every fact that goes into a deliverable is checked against the cited primary page
• Labelling rule: internal documents note when research started from an AI search answer
• Account rule: use business or enterprise workspaces with admin controls where available

Assign an owner for the policy, usually someone in operations or IT, and review it every quarter. Features change, and a rule written for last year’s version may be too loose or too strict now.

For marketers, ChatGPT search is also a channel. People ask it which agency, tool or product to choose, and the answer names some brands and not others. Visibility depends on being present in the sources it reads: clear pages that answer specific questions, accurate third-party listings and consistent facts across the web.

OpenAI publishes information about its crawlers, including OAI-SearchBot for search. Blocking it in robots.txt can keep your pages out of ChatGPT search results, so check your robots.txt before assuming you are invisible for content reasons. Referral traffic from chatgpt.com shows up in analytics and is worth tracking as a separate source.

Errors will appear, so plan for them. Keep a shared log with the prompt, the wrong claim, the cited source and the correct fact. Over a few months this log shows which topics the tool handles badly, and those topics can be moved to the banned list or given a mandatory expert check.