ProductHow it WorksBrief TypesUse CasesPricingFAQ

Use caseI have an AI answer

Check an AI answer against sources you can read

Turn the claim you’re unsure about into a question and read what independent sources say about it — or bring the whole conversation in, labelled as context, with the extension.

Check a claimFree plan · No card required
Get the extension

The extension runs in desktop Chrome — open this page on a computer to add it.

Bringing the whole ChatGPT or Claude conversation in as context needs the extension.

Two kinds of input, kept apart

  • AI chat · context

    • AI context 1ChatGPT conversation

  • Web sources · evidence

    • source 2Vendor pricing page
    • source 3Independent review
    • source 4Practitioner write-up

Source-backed brief

The tier structure the chat described matches the pricing page and the independent review source 2 source 3; the claim that it gets cheaper past twenty seats rests on the conversation alone AI context 1.

No source coverage · cost past 20 seats

C marks the AI chat, cited as context. S marks a web source, cited as evidence. One numbering, never merged.

The question you came with

Can you see which sources an AI used?

Only partly — and no tool can show you all of it. Here is what is visible, what isn’t, and what you can do instead.

  • What some AI tools show

    When an assistant searches the web for a reply, some list the links it retrieved. That tells you what it fetched for that answer — not which sentence leans on which link, or what else shaped the wording.

  • What no tool shows

    What a model absorbed in training, or why it phrased a claim the way it did. SourceFrame can’t reveal that either: it never sees a model’s hidden sources or reasoning.

  • What you can do

    Read the answer against sources you can open yourself — which parts they cover, where they differ, and which confident claims nothing in them supports.

The honest boundary

It reads the answer against sources. It doesn’t grade it.

What SourceFrame does not produce

  • Claim → VerifiedNo such output
  • Claim → FalseNo such output
  • Supported claims / Unsupported claimsNo such output
  • Hallucination scoreNo such output

There is no sentence-by-sentence verifier. A brief that claimed to be one would be the same unearned confidence you came here to get away from.

Instead

What it does instead

  • Labels the material

    The conversation is cited as AI-generated context with its own C citations — included in the set, not counted as independent evidence.

  • Reads it inside a source set

    The brief is built from the sources, so you read the topic as they describe it, with the chat as one labelled voice.

  • Names what isn’t covered

    Where the chat was confident about something no source documents, that shows up as an evidence gap rather than as a verdict.

Start here

Check one claim, or bring the whole conversation

Two different jobs. Checking a claim needs nothing installed; reading the conversation itself beside your sources needs the extension, because a web prompt can’t take a conversation.

Web app · Check one claim

“Is Linear cheaper than Jira past twenty seats?”

Find sources & createMode · Summary

The link fills in the prompt only — choose Summary in the composer before you create. A free brief without an account runs in Auto, which picks the brief type itself.

Route in
  1. The claim, as a question
  2. Summary
  3. Source Brief

Write the claim the AI made as a question. SourceFrame finds independent sources, answers from them, and names what none of them cover. The conversation itself is not part of this brief.

Extension · Bring the conversation

  1. Capture the chat
  2. Add sources
  3. variant 2

Capture the conversation — from the side panel, or the Create brief button SourceFrame adds inside ChatGPT and Claude — on its own or with the pages you already have open. The chat is cited as AI-generated context, never as evidence. A brief built only from the chat says so; add sources and the evidence-backed version arrives as a second version beside it.

The extension runs in desktop Chrome.

What you get

The conversation is one labelled voice

Two cuts from one Source Brief. What’s inside a Source Brief

  • Where the chat sits

    Sources rail

    • source 2

      Compliance platform guide

    • source 6

      Vendor blog

      Perspective · single perspective
    • AI context 7

      Captured ChatGPT conversation · AI-generated context

    S independent evidence · C AI context

    The chat is in the set and marked as AI-generated context; each web source carries its perspective.

  • Where nobody has anything to say

    Evidence gaps & caveats

    • No source covers EU or UK buyer requirements.
    • Two of six sources are vendor-published; the pricing of audit work is single-perspective.

    Evidence gaps name what the sources don’t cover — which is where a confident claim with nothing under it shows up.

A real run

One confident AI claim, researched as a question

This doesn’t show which sources an AI used — no tool can. It shows what the pages SourceFrame read say about the question behind a claim, and where they stop. The claim was written for this example; the brief under it is a real run, quoted as written.

Illustrative AI claim
“If an AI answer includes citations to reputable websites, its claims have already been verified against those sources, so you can generally trust the answer.”

Written for this example. It isn’t from a captured conversation, and it wasn’t an input to the brief.

The question the brief answered

Does an AI-generated answer become trustworthy simply because it cites reputable websites, and do those citations reliably establish the specific claims being made?

Web composer · the claim, researched as a question
Real briefSource BriefGenerated 23 Oct 2026 · 13 sourcesVerbatim excerpts

Whether AI citations make answers trustworthy

The brief’s answer · from Answer

  • No. An AI answer does not become trustworthy just because it cites reputable websites. Trust depends on whether each cited source actually supports the exact proposition the answer makes, and benchmarks show that models often cite relevant or valid sources that still leave many statements unsupported or even wrong.source 10source 12source 8source 4

What the sources establish · from Answer and What the evidence shows

  • Relevant retrieval is not the same as proposition-level supportsource 10

  • Benchmarks show that cited or retrieved evidence still often fails to support the answer… In SourceCheckup's medical benchmark of 800 questions and 58,000 statement-source pairs, GPT-4o with Web Search still had approximately 30% unsupported statements, and its response-level support was 55%. …source 12

What remains uncertain · from Caveats

  • The strongest direct claim-level benchmark evidence here is medical, and the strongest domain framework is legal. Those results show the core problem clearly, but they do not by themselves establish identical rates in every other field.source 12source 10source 3

Sources these excerpts cite

  • source 3

    Evaluating large language models for evidence-based clinical question answering - PMC

    pmc.ncbi.nlm.nih.gov · Found by web search

  • source 4

    Can AI Hallucinations Ever Be Eliminated Completely?

    manuelgarcia.info · Found by web search

  • source 8

    What are AI hallucination evaluations? Metrics and methods that work in 2026 - Articles - Braintrust

    braintrust.dev · Found by web search

  • source 10

    A Survey of Large Language Models for Law: Task Capabilities, Authority Grounding, and System Governance

    arxiv.org · Found by web search

  • source 12

    An automated framework for assessing how well LLMs cite relevant medical references - PMC

    pmc.ncbi.nlm.nih.gov · Selected by the author

AI-generated from 13 web pages found by search and selected by the author. The titles are as the brief lists them; the full set is in the brief.

Open the real brief: Source Brief, opens in a new tabRun this question yourself →

Doesn’t reveal an AI’s sources. A new run finds its own.

The brief was built from web pages only. No conversation was captured, so it has no C citations; bringing a real ChatGPT or Claude conversation in as labelled context is the extension route above, and a different brief.

Where you can go next

  • Primary brief type

    Source Brief

    When your focus is a question about the topic: an answer from the sources, with the conversation cited as context beside it.

  • When the focus points at the sources

    Source Comparison Brief

    Agreement and divergence across the chat and the pages, point by point, with both readings kept.

  • Related job

    Compare what sources say

    When the disagreement is between web sources, with no conversation in the set.

  • Same capture-first start

    Summarize a source set

    What the whole set says, rather than how one answer holds up.

Read the answer with the evidence next to it.

Check a claimFree plan · No card required
Get the extension

The extension runs in desktop Chrome — open this page on a computer to add it.

Bringing the whole ChatGPT or Claude conversation in as context needs the extension.