See how each source contributes to an answer.
How does ProRata attribute an AI answer?
ProRata’s claims tech breaks an AI answer into individual claims, verifies matching claims in the sources used to generate it, and calculates each source’s contribution. An attribution bar displays the results, with evidence you can inspect at each step.
What if the model cannot name its sources?
ProRata compares the claims in an answer with a defined source corpus. This makes the supporting evidence available for inspection independently of the model’s internal parameters.
- 01
Identify
Break the AI answer into discrete “claims”: atomic units of fact.
- 02
Match
Find and verify matching claims in the sources used to generate the answer.
- 03
Calculate
Determine the proportion of the answer’s total claims provided by each source and creator.
- 04
Display
Provide an attribution bar as a nutrition label: show your readers exactly where the answer came from.
A claim
A statement made by or about an entity, event, or concept in a given context. One sentence can contain several claims, each requiring separate verification.
How much did each source contribute?
ProRata determines the proportion of the answer’s claims provided by each source and creator, then makes those contributions visible in an attribution bar.
Connect the evidence
Match each claim to the supporting passages in the sources used to generate the answer.
Calculate contribution
Determine how much each source contributed across the answer’s claims.
Show the result
Display source shares in an attribution bar so readers can inspect where the answer came from.
Proportional contribution score
A measure of how much a source contributes to an answer, based on the claims it supports. It provides evidence that can inform compensation agreements.
What does attribution look like in practice?
Two citations can make two publishers look equally important. Follow a complete answer with ten claims to see why equal citation counts can represent very different contributions.
Illustrative example. The course, answer, publishers, and source passages are invented. Each claim counts equally here to make the comparison clear; this is not a universal scoring formula.
One answer. Ten claims. Two citations.
The course lasts six weeks and includes weekly practice sessions. [1] Classes meet on Tuesdays, and each class lasts ninety minutes. The course is taught online, and every session is recorded. Recordings are available for three months. Enrollment is limited to twenty learners. The course includes a final project, and completing the project earns a certificate. [2]
Publisher A supports 2 claims.
Course overview: “The course runs for six weeks. A practice session takes place every week.”
Publisher B supports 8 claims.
Course details: “Classes meet on Tuesdays for ninety minutes. Teaching is online. All sessions are recorded, with recordings available for three months. Each cohort has up to twenty learners. The course includes a final project. Learners who complete it receive a certificate.”
Inspect all ten claims and their sources
- Claim 1
The course lasts six weeks.
Publisher A - Claim 2
It includes weekly practice sessions.
Publisher A - Claim 3
Classes meet on Tuesdays.
Publisher B - Claim 4
Each class lasts ninety minutes.
Publisher B - Claim 5
The course is taught online.
Publisher B - Claim 6
Every session is recorded.
Publisher B - Claim 7
Recordings are available for three months.
Publisher B - Claim 8
Enrollment is limited to twenty learners.
Publisher B - Claim 9
The course includes a final project.
Publisher B - Claim 10
Completing the project earns a certificate.
Publisher B
Citation counts look equal.
Publisher A · 1 citation
Publisher B · 1 citation
A citation tells you a publisher was referenced. It does not measure how much of the answer that publisher supported.
Attribution reveals the difference.
Publisher A · 20%
Supports claims 1–2
Publisher B · 80%
Supports claims 3–10
What makes the result auditable?
Trace each score to its source.
Scores trace to specific claims and supporting passages. The indexed corpus records the publisher, document, metadata, and source claims. The result also reports the share of content that could not be supported.
Define the sources and review scope.
The team agrees which sources will be examined, how they can be accessed, and what source and permission records are available. This establishes the scope for attribution and any related review of permitted use.
Technical detail +
The service takes a prompt, a generated response, and, where available, the retrieved excerpts and their metadata. It returns attribution scores and support coverage. A claims index organizes source claims by document and publisher.
Outputs include attribution scorecards at claim, article, publication, and publisher levels. Because the method works on output, an evaluation can run offline on logged responses with no change to the customer’s product or user experience.
Input-side telemetry answers what was retrieved. ProRata Attribution answers what contributed to the result. These layers are complementary.
Future direction includes broader source coverage, developer APIs and frameworks, and extensions to audio, music, images, and multimodal content. No availability dates or commercial promises are implied.
FAQ
Is attribution really possible?
Yes. ProRata measures the output of an AI system, matching individual claims to supporting sources. The result connects an answer to supporting evidence and measures each source’s contribution.
How are attribution results checked?
Scores connect to individual claims and supporting source passages so reviewers can inspect the evidence. Evaluation scope and source coverage are established with the team.
Does it work with any model?
The method is model-independent because it evaluates generated output. Model independence is not a claim that every model and workflow has been validated; the team establishes the scope for your system.
What content is in the corpus?
ProRata examines an agreed set of sources. The source cohort for each engagement is agreed with the partner, including the content access and coverage needed for the evaluation.
How is compensation calculated?
Proportional contribution scores provide an input to payout calculations. Commercial agreements determine the payment terms. A contribution score is not itself a statement of money owed.
How do we integrate it?
An initial evaluation can run as an offline batch audit of logged responses, with retrieved excerpts and metadata where available. It does not require changes to the partner’s product or user experience.