Skip to content

AI search · Public sources · Citations

AI Search Citations Compared: ChatGPT, Perplexity, Gemini, Brave and Kruqond

Compare what five AI search tools document about citations, research modes and source access—and use a repeatable checklist to test the evidence behind an answer.

By Kruqond · Published

A citation is a lead, not a fact-check

An answer can show source links and still cite a page that is off-topic, outdated or too weak to support the sentence beside it. A citation also does not, by itself, tell you whether a tool opened the original page, read a Reddit reply, retrieved a video transcript or only found a search result.

The useful comparison is not simply how many links appear. Ask what each product says it can search, how it exposes sources, and what you can verify yourself. Then test the same real question under comparable access conditions.

This guide compares public product documentation for ChatGPT, Perplexity, Gemini, Brave Search and Kruqond, checked on 17 November 2026. It is a guide to documented workflows—not an independent product test, benchmark or ranking. Features and access can vary by plan, account, region and rollout.

What each product documents about sources

The products describe different ways to surface research. These summaries report what their own help pages say; they do not independently prove that every answer will follow the described workflow. Open the original help page for current plan and availability details.

  1. ChatGPT Search — OpenAI documents automatic or manually selected web search, citations and a Sources view when available. Its help page warns that search citations can be incomplete, outdated or incorrect. ChatGPT Deep Research adds a proposed research plan, source controls and a report with citations or source links.
  2. Perplexity — Its Help Center describes conversational answers with numbered citations linking to original sources, plus a Research mode for more involved tasks. Search depth and available features can depend on the user’s plan; the documentation is not a head-to-head quality result.
  3. Gemini — Google says Gemini may show sources or related links in a Sources panel, but not every response includes them. Its Deep Research instructions describe a report workflow that uses Google Search by default and lets users adjust sources; the cited instructions say users must be signed in to use Deep Research.
  4. Brave Search — Brave documents concise AI Answers with source references. It also describes Ask Brave as a longer conversational answer with follow-up and Deep Research options. Those descriptions do not establish how well a particular answer supports a particular claim.
  5. Kruqond — Kruqond documents public-web, code and social-platform research alongside platform-specific post and comment tools. For example, comment workflows use a public post or video URL, or discover a bounded set of posts before retrieving available comments. Kruqond says the result is a sample, not a complete or representative record of a platform.
  1. OpenAI describes web search, citations, a Sources view when available, and cautions that citations can be incomplete, outdated or incorrect.

    OpenAI Help Center: Searching the web with ChatGPT

    What it does not establish: This is OpenAI’s description of its own interface. It does not establish which sources a particular answer uses or whether those sources support each claim.

  2. OpenAI describes a proposed research plan, controls for choosing or prioritizing websites, and reports with citations or source links.

    OpenAI Help Center: Deep research in ChatGPT

    What it does not establish: The documentation describes the workflow, not comparative accuracy, coverage or performance on a fixed set of research tasks.

  3. Perplexity describes cited conversational answers and a Research mode that performs multi-step searches and produces reports.

    Perplexity Help Center: How does Perplexity work?

    What it does not establish: These are vendor-documented capabilities; they do not show that an answer is correct, that every citation is relevant, or that all plans behave identically.

  4. Google says Gemini may show source links in a Sources panel or inline, and explicitly notes that not all responses include sources.

    Google Gemini Apps Help: View related sources

    What it does not establish: This help page does not promise a source list for every answer or explain which sources will appear in any particular response.

  5. Google documents Deep Research with Google Search included by default, selectable sources and a research plan before report generation.

    Google Gemini Apps Help: Use Deep Research

    What it does not establish: The instructions say sign-in is required; available sources, models and request limits may vary. This is not a test of Gemini’s answers.

  6. Brave describes AI Answers with references to sources and Ask Brave for longer, follow-up research, including a Deep Research mode.

    Brave Search Help: AI in Brave Search

    What it does not establish: Brave’s own feature description does not independently assess citation relevance, factual accuracy or the results of a matched prompt.

  7. Kruqond documents separate discovery and comment-retrieval tools, public-source restrictions and bounded comment workflows across supported platforms.

    Kruqond: Source coverage and access limits

    What it does not establish: This is Kruqond’s own product documentation, not an independent review; provider access and the material retrieved can change.

  8. Kruqond says citations do not guarantee correctness and that a retrieved comment sample is not a complete record or representative opinion poll.

    Kruqond FAQ: citations and public-source limits

    What it does not establish: These statements describe Kruqond’s product and limits. They do not establish how another service will behave on the same prompt.

Match the research mode to the question

For a quick, current fact, start with a search mode and open the cited page. For a complex question that needs several sources, use a research or deep-research workflow and inspect its plan, scope and source list. A longer report is not automatically a stronger answer; it simply gives you more material to check.

For social research, separate finding a post from reading its replies. A search result or post title can help you locate a discussion, but it cannot establish what commenters said. If the question depends on comments, check whether the tool returned actual comment text, what post it came from and what may be missing.

For a factual product claim, look for primary documentation or an original announcement. For a user's experience, the original post may be the best evidence of what that person reported, but it does not establish how common the experience is. Different questions need different evidence.

  1. Quick web lookup: inspect the source next to the claim and check its date.
  2. Multi-source research: record the chosen mode, plan, sites and access limits before comparing reports.
  3. Social discussion: save the original public post or video URL and confirm whether the comment content itself was retrieved.
  4. Product or policy fact: follow the claim to the vendor’s current documentation, terms or release notes.
  5. Conflicting evidence: retain the disagreement and explain why one source is more relevant; do not average incompatible claims into a confident answer.

A fair comparison you can repeat

Pick a question you actually need answered, then send the same wording to each product. Record the date, account state, plan, selected mode and any source controls. If one product asks you to sign in or cannot access a source, record that as unavailable; do not score it as a low-quality answer or let the other product win by default.

Keep the raw answers and citations. For every important claim, note whether you opened the original page, inspected comment text, obtained a transcript or only saw a search listing. Repeat across several distinct tasks before making a broad performance claim. One result is a case study, not a benchmark.

For a small citation audit, decide the sampling rule before comparing answers: check every central factual claim, or take the same fixed sample from each answer, such as its first five factual claims. For each sampled claim, record whether the citation opens and whether the page directly supports the statement. Label support separately from source type and factual correctness; an on-topic link is not proof that the claim is true.

A simple, transparent support code can help readers reproduce the review: 0 means the opened source does not support the claim, 1 means it supports only part of it or does so indirectly, and 2 means it directly supports it. Keep a failed or inaccessible URL in a separate unavailable category instead of treating it as zero. Report the number of claims checked and the number unavailable. If you assess factual correctness, verify it separately against suitable independent evidence.

  1. Write the task and success criteria before opening the products.
  2. Use the same prompt, date window and source constraints for every tool.
  3. Record the plan, selected model or mode, plan tier and sign-in state when visible.
  4. Save the complete answer and source URLs; do not count links without checking them.
  5. Mark each source as opened, snippet-only, comment text, transcript or unavailable.
  6. Set the claim-sampling rule and denominator before reviewing; apply the same rule to every product.
  7. Score citation support separately from reachability, source type and factual correctness; publish the coding rule.
  8. Deduplicate sources that repeat the same underlying announcement or report.
  9. Repeat on varied tasks before using words such as best, faster, more accurate or broader.

Check citations one claim at a time

A tidy source list can hide weak evidence. Read the sentence beside each citation, then open the link and check the section that supposedly supports it. A citation to a page about the same topic is not enough if the page never makes the claim.

Count independent evidence, not repeated pages. Several summaries may trace back to the same announcement. For a social-media claim, also preserve the post date, parent or reply context, and retrieval limits. For a video claim, distinguish a visible title or metadata from a transcript or content you actually watched.

  1. Relevance: does the source address the exact product, version, place and time in the claim?
  2. Support: can you point to the passage that establishes it, or is the answer making an inference?
  3. Origin: is this a primary document, a first-hand report or a summary repeating another source?
  4. Freshness: when was it published or updated, and could a later change matter?
  5. Context: did you inspect the surrounding text, parent comment or video content needed to interpret it?
  6. Limits: what could not be opened, retrieved or compared? Keep unknown distinct from false.

What this comparison does not establish

Public help pages let you compare documented features, but they do not show how well the products answer the same question today. They cannot establish a universal winner, prove that citations are correct, or show that every service can retrieve comments from every social platform. Those claims need controlled tests with saved outputs and comparable access.

Disclosure: Kruqond publishes this guide and is included in the comparison. The summaries were prepared with AI assistance from the official documentation linked above. This is not independent coverage, and no product-performance test was run for this article. The related Kruqond vs Perplexity source audit records a separate single task and explains why its incomplete Perplexity session was not scored.

Before you finish

  • Record the exact prompt, date, account state, plan and selected research mode.
  • Open each citation and match it to the specific claim it should support.
  • Distinguish original sources from summaries that repeat the same announcement.
  • Label whether a source was opened, snippet-only, comment text, transcript or unavailable.
  • Preserve dates, reply context, source limits and conflicting evidence.
  • Treat undocumented behavior as unknown, not as proof a feature is absent.
  • Do not name a broad winner from one task or from documentation alone.

Related guides and product information

Follow the evidence, not just the answer.

Try five free searches without signing up, then open the cited sources yourself.

Try Kruqond