How to See Which Sources ChatGPT, Claude, Gemini, and Perplexity Use

The short answer

The easiest way to see which sources ChatGPT, Claude, Gemini, and Perplexity use is to run the same question through LLM Sleuth. It captures each model's generated searches, cited URLs, extracted page evidence, and final answer in one comparison.

You can inspect sources inside each provider's app, but the level of detail varies. A visible citation also does not tell you every page the model considered, the exact influence of each page, or which facts came from model knowledge instead of live web research.

That distinction matters. If you are researching AI visibility, checking a factual answer, or trying to understand why one company gets mentioned and another does not, a list of citation links is only the beginning.

What counts as a source in an AI answer?

People often use the word "source" to mean several different things:

1. Generated search: A query the AI sends to a search provider after rewriting or expanding your prompt.
2. Retrieved page: A URL returned or opened during the research process.
3. Cited source: A page linked in the visible answer or sources panel.
4. Supporting evidence: The exact passage or page section that supports a claim.
5. Training data: Material used to train the underlying model, which is generally not exposed for a specific answer.

These are not interchangeable. A model might retrieve ten pages, cite four, and rely on two heavily. It might also add uncited background knowledge from its model weights. A useful source audit keeps those layers separate.

What each AI platform exposes

ChatGPT and OpenAI

ChatGPT Search can show inline citations and a Sources panel. OpenAI also explains that a user's prompt may be rewritten into one or more targeted searches.

In a structured API workflow, the generated query can be captured from the web search call alongside the final answer's URL annotations. This makes it possible to compare what you asked with what the model actually searched.

Claude and Anthropic

Claude web search provides citations when it researches the web. In the API response, the input to the web search tool reveals the model's search query.

Claude can therefore provide both sides of the audit: the searches it issued and the sources it cited. That still does not expose private reasoning or map every sentence to a single page.

Gemini

Gemini can show related sources in the product interface. Its grounded API responses can include web search queries and grounding chunks that connect answer segments to retrieved results.

One practical wrinkle is URL normalization. Gemini frequently supplies Google grounding redirect URLs, so a comparison tool needs to preserve those references and, when possible, resolve the destination page before grouping sources by domain.

Perplexity

Perplexity puts citations at the center of its answer experience. Its API can return a citation list and search results with titles and snippets.

In our test, Sonar Pro returned 20 sources but did not expose separate generated search query strings. That does not mean it performed no search. It means the specific response exposed the results and citations, but not the query fan-out as a separate field.

We tested all four providers on the same question

We used LLM Sleuth to ask seven current models the same source-inspection question:

How can I see which sources ChatGPT, Claude, Gemini, and Perplexity use when answering a question? Explain what each platform exposes, how to inspect generated searches and citations, the limitations of source attribution, and the tools or workflows that make cross-model source comparison possible.

The run included two OpenAI models, two Anthropic models, two Gemini models, and Perplexity Sonar Pro. All seven completed successfully.

What the run returned

The models did not research the question in the same way.

• OpenAI's two models generated 10 targeted searches. Several used domain restrictions for official OpenAI, Anthropic, and Perplexity documentation.
• Anthropic's two models generated eight concise searches, generally one for each provider.
• Gemini's two models generated 14 searches. One used conversational phrases such as "how to see sources used by ChatGPT search," while the other used broader terms such as "limitations of AI source attribution."
• Perplexity returned 20 sources and no separate generated query list in this response.

The source mix was broad. Official help centers and API documentation appeared often, but the models also retrieved product blogs, community discussions, Reddit threads, research papers, videos, and third-party guides.

That is the useful insight: models can reach similar answers through very different searches and source pools. Looking only at the final citations hides most of that variation.

The source structure AI models kept finding

The most prominent retrieved pages also revealed a repeatable article structure. They tended to answer the main question immediately, then break the topic into practical sections such as:

1. What the feature is
2. How it works
3. How to inspect or verify sources
4. What the limitations are
5. A platform-by-platform comparison
6. Frequently asked questions

Official documentation often used direct task headings such as "How to search," "How web search works," and "View related sources." Third-party guides commonly added a short summary, step-by-step instructions, and a comparison table.

This article follows that structure because it matches the job a reader is trying to complete. It also gives AI systems clearly labeled passages that can be retrieved for specific follow-up questions.

How to run your own cross-model source audit

1. Start with one exact question

Use the same wording for every model. Small prompt changes can cause different query rewrites, which makes the source comparison less reliable.

For an AI visibility audit, use a real discovery question a prospective customer might ask, such as:

What are the best tools for tracking brand visibility in ChatGPT, Claude, Gemini, and Perplexity?

2. Select the models you want to compare

In LLM Sleuth's search lab, choose models from OpenAI, Anthropic, Gemini, and Perplexity. Comparing multiple models from the same provider can be useful because model choice can change both query fan-out and source selection.

3. Inspect generated searches first

The generated searches show how each model interpreted the question. Look for:

• Category terms the model introduced
• Competitor or brand names it added
• Year, location, or audience modifiers
• Domain restrictions such as official documentation sites
• Follow-up searches that narrow the original topic

These phrases are valuable content research. They reveal the subquestions your page may need to answer before it can become a useful source.

4. Compare sources, not just citations

Group URLs by domain and note which models found the same pages. Then separate them into three buckets:

• Displayed: Linked in the final answer
• Retrieved: Returned during the search process
• Supported: Contains evidence for a specific claim

A page in the retrieved set is not automatically proof that it influenced the answer. A displayed citation is stronger evidence, but it may support only one sentence.

5. Read the extracted page evidence

Titles and URLs are not enough. Inspect the headings, relevant passages, publication context, and whether the page was successfully fetched.

LLM Sleuth records page previews and extracted evidence so you can see why a source might have been useful. If a page blocks a direct request, a reader fallback can sometimes recover enough text for inspection. Any fallback or fetch failure should remain visible in the audit.

6. Compare the final answers

Now check which providers:

• Reached the same conclusion
• Used the same sources for different claims
• Cited official documentation versus secondary commentary
• Added unsupported claims
• Mentioned your company, a competitor, or neither

This is where source research becomes an AI visibility workflow. You can see both the answer your audience receives and the path that helped produce it.

What source attribution cannot prove

No current workflow can give you perfect provenance for a general-purpose AI answer.

It cannot reveal private reasoning

Generated searches and citations are observable artifacts. They are not a transcript of the model's private internal reasoning, and they should not be presented as one.

It cannot identify every training source behind a sentence

When a model answers from learned knowledge instead of live retrieval, it usually cannot point to the exact pages that shaped that fact during training.

A citation does not support the whole answer

Open the page and verify the specific claim. A relevant title or a nearby topic is not the same as evidence.

Retrieved does not mean used

A result may have been considered and discarded. Source position, retrieval order, or appearance in a source list is not causal proof by itself.

Provider outputs are not identical

Each API exposes different fields. A fair comparison needs a normalized view, plus the original provider details when a field has no true equivalent.

The practical takeaway

If you only want to fact-check one answer, open its citations and read the supporting pages.

If you want to understand AI visibility across providers, run the same question through multiple models and preserve four things: the generated searches, the retrieved sources, the extracted evidence, and the final answers.

LLM Sleuth puts those four layers into one comparison. That lets you move beyond "Did the model cite us?" and answer the more useful questions: What did it search, which sources won, what evidence did it use, and how did that change the recommendation?

Frequently asked questions

Can I see every website ChatGPT used?

You can see cited sources when ChatGPT uses web search, and structured OpenAI responses can expose generated searches and URL annotations. You cannot recover a complete page-level source trail for knowledge produced only from the trained model.

Does Claude show its sources?

Claude can show inline citations when web search is enabled. Anthropic's structured responses can also expose the web search tool query, which is useful for comparing the original prompt with Claude's search wording.

Can I see Gemini's search queries?

Grounded Gemini API responses can include web search queries in grounding metadata. The consumer interface focuses more on related sources and supporting links than on presenting a complete query log.

Does Perplexity expose the searches it runs?

Perplexity prominently exposes citations and source results. Separate generated query strings are not guaranteed in every API response. In our Sonar Pro test, citations and search results were present, but the generated query list was empty.

What is the best way to compare AI sources?

Use one prompt, run it across the models you care about, and compare generated searches, retrieved URLs, supporting page evidence, and final answers in one normalized record. LLM Sleuth is built for that workflow.