The Grounding Report

Every generated artifact — audio overviews, video slideshows, and infographics — includes a Grounding Report. This is your factual audit trail: a record of every claim the AI made and where it came from in your documents.

What the Grounding Report Shows

When a generation job finishes, the platform runs a two-step verification process on the output:

1. It checks each factual statement in the generated content against the evidence retrieved from your documents.
2. It flags or removes any claim that cannot be tied back to a specific passage in your files.

The result is a summary that tells you:

• How many factual claims appeared in the generated content.
• How many were verified — meaning a supporting passage was found in your documents.
• What was removed — statements the system could not substantiate, along with a brief note on why.
• Where each verified claim came from — the source document name and page reference.

How to Read the Report

The Grounding Report is available in the artifact detail view after a job completes. Here’s how to use it:

1. Check for removed content. Any statements listed as removed could not be traced to your source files. Review these manually before publishing — they are the highest-risk parts of the output.
2. Verify key claims. For your most important statements (statistics, product names, financial figures), use the source document and page references to confirm accuracy against the original.
3. Use the verified ratio as a confidence indicator. A high proportion of verified claims means the output is closely grounded in your data. A low ratio warrants careful manual review.

Best Practice

Before publishing any AI-generated content, always check the Grounding Report. Pay particular attention to any statements that were flagged or removed, as these are where inaccuracies are most likely to occur.

Advanced Usage Notes

Use these patterns to get more predictable results:

• Keep inputs specific and scoped to one intent per generation request.
• Save intermediate outputs often, especially before running major edits or regenerations.
• Use review and assignment workflows early to reduce rework late in the lifecycle.

Troubleshooting

If results are not what you expected:

• Verify source quality first (missing context, inconsistent structure, or conflicting inputs reduce output quality).
• Re-check selected options and audience/profile settings before re-running.
• If a job fails repeatedly, wait briefly and retry once; then escalate with the artifact or job context.

Frequently Asked Questions

Why did quality change between runs?

Different input context, settings, and model/runtime conditions can produce different outputs.

Should I regenerate or manually edit?

Regenerate when the core structure is wrong; manually edit when only wording, ordering, or style needs adjustment.

What is the best way to reduce iteration cycles?

Start with clear scope, run smaller batches, and use review checkpoints after each major step.