Grounding Depth Index

Grounding Depth Index (GDI) scores how deeply an AI output is anchored to source material at the claim level. Not citation count. Not link stuffing. It is a quality measure for whether each factual statement can be traced to a specific passage, table, product spec, policy page, or dataset row.

For GEO teams, that matters because generative answers get trusted or rejected in seconds. If your model cites a homepage for a pricing claim that actually lives three clicks deeper in a PDF, your grounding is weak even if the answer looks polished.

What GDI actually measures

A useful GDI model usually evaluates three things: claim coverage, source specificity, and match accuracy. Coverage asks how many factual statements have support. Specificity asks whether the support points to an exact section, URL fragment, table, or quote. Accuracy checks whether the cited source really supports the claim instead of merely sharing keywords.

In practice, teams score GDI on a 0-1 or 0-100 scale. A rough framework:

• 0.00-0.40: weak grounding; broad citations, unsupported claims, obvious retrieval misses
• 0.41-0.70: usable for low-risk content, but still needs review
• 0.71-0.85: solid operational range for most editorial and product content
• 0.86+: strong grounding, usually required for medical, legal, or finance workflows

Why GEO teams should care

GDI is not a Google ranking factor. Let's be precise. Google Search does not publish a "Grounding Depth Index," and Google has not said it uses this metric directly. But the underlying behavior matters because unsupported AI content tends to fail on the signals that do matter: trust, accuracy, editorial review, and user satisfaction.

Google's John Mueller has repeatedly pushed the point that content quality is not rescued by the production method. In 2025, he again emphasized that useful, accurate content is what matters, not whether AI helped write it. Same standard. Different workflow.

For SEO operations, GDI is a control metric. Use it to compare prompt versions, RAG configurations, or model vendors. If one setup moves from 0.52 to 0.79 while keeping answer completeness stable, that is a real improvement. Track it next to manual fact-check pass rate, citation error rate, and downstream engagement in GSC.

How to measure it in the real world

Most teams do not need a research-grade framework. They need a repeatable one. Extract factual claims, map each claim to a source fragment, then weight the match quality. Ahrefs, Semrush, and Moz will not calculate GDI for you. This is closer to an internal QA metric than a standard SEO platform KPI.

Use Screaming Frog to verify cited URLs return 200 status codes and are indexable. Use GSC to monitor whether pages with stronger source transparency hold clicks and impressions better after publication. If you are testing answer formats, Surfer SEO can help standardize on-page structure, but it will not validate factual grounding.

The caveat: GDI can be gamed. A model can attach lots of citations and still misread the source. High citation density is not high truth. You still need human review on sampled outputs, especially in YMYL topics and anywhere source documents change weekly.

What good implementation looks like

• Require claim-level citations, not one source list at the bottom
• Prefer exact source fragments over domain-level references
• Set thresholds by risk: 0.75 for product content, 0.90+ for regulated content
• Audit 20-50 outputs per prompt or model change before rollout
• Track false-support rate, not just citation presence

Bottom line: GDI is useful because it forces a hard question. Can this answer be checked quickly and defended confidently? If not, the content is not ready, no matter how fluent it sounds.