BRAG Evidence Gates: How AiVIS Cite Ledger Uses Proof-First Citation Filtering | AiVIS Cite Ledger Blogs
By R. Mason · · 13 min read · INFRASTRUCTURE
BRAG is the filtering layer that stops AI visibility tools from reporting findings they cannot prove. Here is what the system is, how it was customized for AiVIS, and why it changes citation audit reliability.
Key Takeaways
- BRAG (Based, Retrieval, Auditable, Grading) is a filtering standard that prevents tools from reporting findings they cannot prove.
- AiVIS Cite Ledger adapted base BRAG with multi-model consensus, temporal stability gates, and query-intent grounding for production citation auditing.
- The Cite Ledger dataset contains 50,000+ findings across 8,000+ sites, informing all fix recommendations with evidence of what works.
- Every finding must pass four gates before reaching your audit report: proof in crawl data, fetchability by AI crawlers, reproducibility over time, and impact scoring.
- Post-fix re-audits validate that changes actually resolved the finding, closing the repair trust loop that other tools leave broken.
Article
# BRAG Evidence Gates: How AiVIS Cite Ledger Uses Proof-First Citation Filtering
What BRAG actually is
BRAG stands for Based, Retrieval, Auditable, Grading.
It is not an acronym someone invented for marketing.
It is a filtering standard that emerged from analyzing thousands of AI visibility audits and identifying the failure point where tools report findings they cannot defend.
The original repository at github.com/dobleduche/brag is a reference implementation that enforces one principle: if a finding cannot be traced back to a specific piece of page data, the finding does not exist in the output.
Most AI visibility tools do not have this gate.
They run a model against your page, the model returns a score, and they report the score without verifying that each input signal is real and provable.
BRAG closes that gap.
---
The four gates: Based, Retrieval, Auditable, Grading
Based
Based means the finding must map to specific, observable content from your live page.
Not an estimate. Not a model guess. Not a "your site probably lacks this."
An actual crawl observation.
Example of based:
"Your Organization schema is missing the sameAs field (crawl evidence: line 247 of page JSON-LD)"
Example of not based:
"Your brand authority might be weak"
Retrieval
Retrieval means the finding must describe content that answer engines can actually access at query time.
A page can have perfect schema and still be invisible if robots.txt blocks the crawler or if the page is behind a login.
BRAG checks this separately.
Example of retrieval:
"Your FAQ page is crawlable and includes 12 Q+A pairs"
Example of not retrieval:
"Your FAQ exists" (without checking whether an AI crawler can fetch it)
Auditable
Auditable means every finding must remain stable and repeatable.
Run the same page crawl twice. Run the same prompt twice. The finding should emerge both times.
If something is a one-time blip, it is not auditable. It is noise.
Example of auditable:
Enable JavaScript for the full interactive reading experience with related articles and discussion.