What Should I Do When AI Answers Cite No Sources?

Why do some AI answers cite no sources?

• No search was run. Assistants such as ChatGPT decide when to search. For questions they consider general knowledge, they answer from what the model learned in training.
• The platform shows sources differently. Some assistants name brands without linking to their sites. Semrush and Growth Memo found Gemini named brands in about 84% of appearances but cited their sites in only about 21%.
• The question does not need current information, such as definitions, history or broad advice.
• Search is turned off in the app, the account or the plan being used.

Perplexity is the main exception, since it shows sources for almost every answer.

How can you tell whether an answer used search?

• Look for a search indicator, a sources panel or inline links in the interface.
• Ask the same question again with search explicitly on, or add "search the web" to the prompt, and compare.
• Use APIs for testing. When OpenAI's API runs a web search, it returns the sources used. Google's Gemini API does too, but its terms allow grounded results to be shown only to the person who asked, not stored or analysed for testing.
• Check dates. Answers that miss recent changes, such as a new price or a new location, usually come from training.

What should you do when an uncited answer is wrong about you?

1. Record the error with the prompt, date and assistant.
2. Find the likely origin. Search the web for the wrong fact: old pages on your own site, directory listings, press coverage, review sites, Wikipedia or Wikidata. Errors in training usually come from information that was widely published at some point.
3. Correct what you can: update or annotate old pages, fix listings, ask publishers to update articles.
4. Publish the correct fact clearly on your own site, in text, on the page most likely to be retrieved.
5. Check the search-enabled answer. If it is right, most users who need current facts will get them. If it is wrong too, the problem is in current sources, as covered in how to correct false or outdated information that AI repeats from its sources.
6. Report the error with the feedback option.
7. Recheck after model updates, since training-based answers change mainly when models are retrained.

What should you do when uncited answers leave you out?

Answers from training reflect what was widely written about a category when the model's training data was collected. To appear in future models' answers:

• Be written about widely and consistently: press coverage, industry articles, directories, reviews and reference sources that describe what you do.
• Keep one clear description of your company and category everywhere.
• Consider training crawler access. OpenAI uses GPTBot to collect training data, separately from OAI-SearchBot for search. Blocking GPTBot keeps your site out of future OpenAI training data but does not remove you from ChatGPT search. Anthropic's ClaudeBot plays the same training-only role for Anthropic. Google-Extended covers training of Gemini models and also grounding in Gemini apps, so blocking it also keeps your pages out of grounding in the Gemini apps. Allowing them lets future models learn from your own pages. Blocking them is a legitimate choice, but it has this cost.
• Focus on search-enabled answers now, since they respond to changes within weeks rather than model cycles.

How assistants decide whether to cite sources is covered in whether ChatGPT always provides sources and how many it cites.

Which sources shape training-based answers most?

Model makers rarely publish full training data, but the pattern is well known. Large web crawls such as Common Crawl, reference works, books, news and public discussion sites make up much of it. Some sources are weighted above their size. The paper describing GPT-3 reported that Wikipedia was sampled more heavily than its share of the data, at about 3% of the training mix and several passes over the text. For a business, that means widely copied descriptions matter: a Wikipedia or Wikidata entry if you qualify, major directories, and articles repeated across many sites. One accurate page on your own site is rarely enough to outweigh many older copies elsewhere.

When should an uncited error be escalated?

Escalate when the error could cause harm: wrong safety or dosage information, a false claim that you are under investigation or out of business, or wrong contact details that expose customers to fraud. Use the assistant's feedback option with a clear explanation. Keep records, and take legal advice where the claim is defamatory.

Should you ask the assistant for its sources?

You can, but treat the reply with care. When asked after the fact, an assistant may run a search and list pages that support its answer. Those pages did not necessarily shape the original answer, and it may even invent sources. Use the reply as a set of leads to check, not as proof of where the answer came from.

How should you track uncited answers?

Record the share of answers about you, or in your category, that show no sources, per assistant. Track accuracy for cited and uncited answers separately. If uncited answers are consistently wrong while cited ones are right, the problem is outdated training data, which only broad, consistent information and time will fix. If both are wrong, fix current sources first.

What might an illustrative case look like?

Consider an illustrative food delivery software company that changed its name two years ago. Asked "What is [old name]?", ChatGPT without search describes it as a restaurant review site, with no sources. Searching the web for that description finds a directory listing and two old press articles from the company's first product. The company asks the directory and publishers to update or annotate the pages. It adds a "formerly [old name]" section to its About page and confirms that GPTBot and OAI-SearchBot can read the site. With search on, ChatGPT already describes the company correctly. The uncited answer remains wrong until a later model update, after which it too reflects the current business.

What do people ask about uncited AI answers?

Can I make ChatGPT always cite sources about my brand?

No. The assistant decides when to search. Users can ask it to search, and many do for current information.

Are uncited answers less trustworthy?

They cannot be checked as easily, and they may be out of date. For current facts, search-backed answers are usually more reliable.

How long until training-based answers change?

Until the next model update that includes newer data, which can take months and varies by company.

Does blocking AI training hurt my visibility?

It can reduce what future models know about you from your own site, though other sources still describe you. Blocking GPTBot or ClaudeBot does not affect search-based answers if those companies' search crawlers are allowed. Google-Extended is different: blocking it also stops your pages being used to ground Gemini app answers, though Google Search and AI Overviews are unaffected.