A source is where information came from. A citation is the marker that points you to it.

When you read a good news article or a school textbook, you often see small numbers or links beside certain sentences. Those marks lead to the place where the claim was found: a report, a web page, a book, or a study. They let readers verify the facts for themselves. The same idea applies to AI answers. A source might be a web page, a document you uploaded, or an entry from an app you connected. A citation is the visible pointer tying a statement in the answer to that source.

Citations matter because AI models can sometimes produce statements that sound right but are not. Without a pointer to a source, you have only the model’s word. With one, you can open the original and judge for yourself. That is why sources are one of the most practical tools for trusting AI answers wisely, rather than blindly.

flowchart LR
    A["A claim in Claude's answer"] --> B["Citation marker"]
    B --> C["Source: web page, document, or app"]
    C --> D["You open it and check"]
        

A few simple words appear again and again in this topic. Here they are in plain English.

AI answers can sound equally confident whether they are right or wrong.

A language model writes fluent, well-organized text. That fluency is a strength, but it also hides weakness. A correct answer and an incorrect one can look identical on the screen. The model does not raise a flag when it is unsure. This is why checking matters, and why sources are so valuable. They give you something concrete to compare against.

Sources also help in another way. Claude learned most of what it knows during training, up to a certain date. Anything that happened later, such as new laws, prices, or news, is not part of that training. When Claude can search the web and cite what it finds, it can answer questions about current matters and show you exactly where the information came from.

Claude can draw on several kinds of source, and the way it shows them depends on the kind.

The last row of knowledge, Claude’s own training, is the one most likely to be uncited. When Claude answers from memory alone, there is no source to point to, and that is exactly when verification matters most. If a fact is important, it is worth asking Claude to search or to show a source.

flowchart TD
    A["Your question"] --> B{"Does it need current or checkable facts?"}
    B -- "Yes" --> C["Claude searches the web or reads your documents"]
    C --> D["Answer with citations"]
    B -- "No, general knowledge" --> E["Answer from training"]
    E --> F["No specific source, so verify if it matters"]
        

When Claude searches the web, it shows where its information came from.

According to Anthropic’s Help Center, web search lets Claude look up current information. When Claude uses the live web to help answer you, the article says every response includes citations so you can verify sources yourself. During a search, you will see an indicator that Claude is searching. Claude looks at multiple sources, and then writes a conversational answer. That answer includes direct citations to sources, source links for further reading, and relevant quotes when appropriate.

When web search is on, Claude can also open a specific web page if you give it a link. This is useful when you want to ask about an article, a policy page, or a report. The article also notes that search results can include images, each with a source link so you can visit the original page.

A small change in how you ask can make a big difference in how well Claude shows its sources.

The Help Center offers practical tips. If you want to be sure Claude uses web search, you can include a phrase such as “search the web” in your request. You can also ask for several sources, because Claude can gather and combine information from various places to give you a fuller picture. And you can tell Claude not to use web search if you do not want it to.

When you share a document, Claude can base its answer on it and show you where.

Often the best source is a document you already have: a contract, a report, meeting notes, or a manual. You can upload it or paste it into the conversation and ask questions. Because the document is in front of Claude, it can answer from the real text instead of guessing. You can strengthen this by asking Claude to point to the exact place in the document that supports each part of its answer.

People who build apps with Claude can switch on a Citations feature for precise references.

Anthropic offers a Citations feature in its developer interface. Developers can give Claude source documents, turn citations on, and receive answers in which claims are tied to specific passages in those documents. The documentation explains that the supplied documents are processed so that Claude can point to exact sentences or locations. For plain text, the pointers refer to character positions. For PDFs, they include page numbers. For custom content, they refer to the pieces the developer provided.

The documentation also says that this built-in approach has advantages over simply asking Claude in a prompt to quote its sources. The citations are parsed into a standard format and are guaranteed to contain valid pointers to the provided documents, and in Anthropic’s own testing the feature was more likely to cite the most relevant quotes than prompt-only approaches. The cited text does not count toward output tokens.

A source appearing in the answer does not automatically make the answer correct.

It is easy to feel reassured when you see links and references. But a citation only tells you where Claude says the information came from. It does not guarantee that Claude read the source correctly, that the source itself is accurate, or that the source says exactly what the answer claims. The Help Center itself advises users to verify important information and cross-check cited sources.

A short routine turns a citation into real confidence.

flowchart TD
    A["Citation in the answer"] --> B{"Does the source open and exist?"}
    B -- "No" --> C["Treat as unverified"]
    B -- "Yes" --> D{"Does it say what the answer claims?"}
    D -- "No" --> E["Correct the claim or ask Claude to recheck"]
    D -- "Yes" --> F{"Is the source reliable and current?"}
    F -- "No" --> G["Find a better source"]
    F -- "Yes" --> H["Confident enough to use"]
        

Not all sources deserve the same trust.

When Claude reads from your own apps, the source is something you can check yourself.

Connectors let Claude read from services such as calendars, documents, and project trackers, with your permission. When Claude answers using information from a connected app, the useful habit is to ask which item it used. For example, “Which document did you take that from?” or “Which meeting does that refer to?” You can then open that item in the app and confirm.

This is especially important for work information. A figure from a shared spreadsheet, a deadline from a project tracker, or a decision from meeting notes may be exactly what you need, but small misreadings can have real consequences. Treat the connected source the same way you would treat a web source: open it and check.

If you use Claude’s information in a report, essay, or article, cite the real source, not the AI.

When you write something for others, the standard practice is to credit the original source of facts, such as the report, study, or official page, so readers can check it. Claude is a helpful guide to those sources, but it is usually not the original origin of a fact. A good approach is to use Claude to find and understand sources, then read the sources yourself and cite them directly.

Different schools, workplaces, and publishers have their own citation styles and rules about AI use. Some require you to say that you used an AI tool. Check the rules that apply to you, and follow them honestly.

Let us follow an imagined student named Farah who wants to understand a claim she saw online.

Farah used sources the right way. She asked for strong sources, opened them, compared the wording, caught a small overstatement, and treated a health question with extra care.

A few habits cause most problems with AI sources.

Sources add trust, and also add a bit of work.

Keep these habits in mind whenever you rely on a sourced answer.

Quick answers to the questions beginners ask most often.

A visible pointer that connects a statement in an answer to the source it came from, such as a web page or a passage in a document, so you can check it yourself.

Not always. When Claude uses web search, the Help Center says responses include citations. When it answers from general knowledge learned in training, there may be no specific source, so important facts should be checked.

In some versions you switch web search on from the controls by the message box, and in the newer experience Claude searches when it helps. You can also write “search the web” in your request. On organization plans, an administrator may need to enable it first.

Yes. According to the Help Center, when web search is on, Claude can also retrieve content from a specific page when you provide its link.

Not necessarily. Citations show where the information came from, but Claude can still misread a source, and the source itself may be weak or outdated. Always open and check the important ones.

AI models can sometimes produce references that sound real but are not, especially when asked to list sources without being able to look anything up. Open every reference to confirm that it exists.

Share the document and ask Claude to name the section or page for each claim, then check the original. Developers can also use the Citations feature in the Claude API for structured references.

According to the documentation, citations point to specific places in the supplied documents, such as character positions for plain text and page numbers for PDFs.

Generally you should cite the original sources that you have read. Check your school, employer, or publisher’s rules about whether and how to disclose AI assistance.

Yes. The Help Center notes that web search and web fetch count toward your usage limits, so turn it off or tell Claude not to search when you do not need current information.

Links can occasionally fail or change. Search for the page title, try another source, or ask Claude to find a working source and check again.

In the Claude Help Center article on enabling and using web search, and for developers, in the Citations page of the Claude developer documentation.

Sources tell you where information came from, and citations are the pointers that let you check it. When Claude uses web search, Anthropic says its responses include citations, source links, and relevant quotes where appropriate. You can also base answers on your own documents, ask Claude to point to the section or page it used, and, if you are a developer, use the Citations feature to tie claims to exact passages. Citations make answers more checkable, but they are not proof. Claude can misread a source, a source can be weak or outdated, and AI can sometimes produce references that do not exist. The reliable habit is to open the source, compare the exact wording, judge the publisher, and cross-check anything important, with more care for health, money, safety, and legal matters. In your own writing, cite the original source, not the AI.