Home  /  Methodology  /  Source Grounding Ratio · 2026-04-27

Source Grounding Ratio (SGR) — SEO Industry Cohort

Top10Lists.us SGR 0.94 — cohort median 0.26 (Site A–D, identities in receipts.json; range available there too).

Of the numeric claims on a page, what fraction are cited — weighted by the authority of the cited source? Ungrounded numbers raise AI hallucination risk and human credibility loss; SGR isolates the question AI engines must answer before citing.

Frozen: 2026-04-27 — measurements at this URL will not change.  ·  Permanent dated artifact  ·  GEOlocus.ai (GeoLocus Group, a subsidiary of Aryah.ai)

Authors: Robert Maynard, Cofounder and CEO · LinkedIn →  ·  Mark Garland, Cofounder and CRO · LinkedIn →

Download raw receipts (JSON) →

1. The Metric — Source Grounding Ratio (SGR)

For each numeric claim on a page:

SGR = (Σ tier_weight · cited) / total_numeric_claims

A “numeric claim” is any standalone factual assertion containing a number,
percentage, money figure, ranking, count, or quantitative ratio. Excludes:
  - Numbers in URLs, prices on commerce CTAs, dates without claim semantics,
    decorative numerals (year stamps, version numbers, page numbers).

cited = 1 if hyperlinked to or footnoted with a primary source on the same
        page; 0 if uncited.

tier_weight depends on the cited source's authority tier:

Reported on a 0–1 scale. Higher = more of the page's quantitative claims trace back to a primary, authoritative source. SGR 1.00 = every claim cites a T1 source; SGR 0.00 = every claim uncited.

2. Why SGR Matters

AI engines verify before they cite. When a page asserts “73% of B2B buyers use social media to research vendors” with no citation, the engine must either discard the claim (safer for hallucination), fall back to its training distribution (compressed reasoning, may invent), or cite a weaker secondary aggregator (citation gets harder to defend). All three outcomes reduce the probability the original page becomes a verbatim citation.

The same claim with a footnote to the underlying primary source — e.g., a Census Bureau dataset, a CDC publication, an SEC filing — becomes directly verifiable. The engine's verification step succeeds; the citation lands. SGR instruments the upstream cause of that distinction across an entire page.

SGR is a sub-measure of the GEOlocus.ai 9-dimension GEO rubric — specifically the Content Density dimension (15 points). It is one of several instrumentations under that umbrella, alongside semantic-container coverage, Relevance Ratio, and JSON-LD entity density. The composite Content Density score is a function of all four; SGR is reported standalone here so the methodology is independently auditable.

3. Operational Definitions

4. Cohort & Per-Page Results

Same 5-site cohort as the Relevance Ratio benchmark: Top10Lists.us (named) plus four established SEO industry sites anonymized as Site A through Site D. We selected up to 3 high-traffic pages per site (homepage + 1–2 representative content pages) and counted every numeric claim on each. Site identities and exact URLs are in receipts.json →.

5. Per-Site Aggregate SGR

6. Why Top10Lists.us Scores 0.94

The clean-room HTML strategy and the Floor+ DB-backed-stats convention compound here. Every count on a Top10Lists.us page that an AI engine could verify — agent counts, neighborhood counts, total surveyed agents, total selected, crawl statistics — comes from a versioned DB query at render time, with the underlying counts publicly auditable via the /crawl-stats transparency dashboard and the published Floor+ convention. T1 status is conservative-but-defensible: these are first-party, publisher- owned datasets, the same authority class as a government dataset for the underlying fact-domain.

The one cohort outlier on Top10Lists.us (city / agent page, SGR 0.83) is a single T5 claim in a marketing CTA — an honest acknowledgement that no system is at 1.00 across every surface, and a useful direction for the next iteration.

7. Reproduce This Measurement

SGR cannot be fully automated — tier classification per citation is a judgement call (a regex cannot tell whether a NYT URL is primary reporting or a re-aggregator). The downloadable script halts with the tier-weight reference table and an example LLM prompt; plug it into Claude / GPT-4 / Gemini to replicate the cohort classification pipeline. The script also fetches a target URL and prints all outbound links for human-or-LLM tier classification.

Pseudocode (from reproduce.mjs →):

html  = fetch(url, ua=Googlebot)
body  = strip(html, [<head>, <script>, <style>, <nav>, <footer>])
links = extract <a href> from body, excluding self-anchors and mailto/tel/js

# Halt and prompt: ask LLM to tier-classify each link.
# Tier 1 (1.0): primary research, gov stats, standards bodies
# Tier 2 (0.8): industry data, NYT/WSJ primary reporting
# Tier 3 (0.5): trade journals, Wikipedia w/ T1-2 sources
# Tier 4 (0.2): vendor blogs, recycled aggregators
# Tier 5 (0.0): uncited mentions

sgr = sum(tier_weight[c.tier] for c in citations) / n_citations

Parameters:

• --url=<page> — optional; fetches and lists outbound links for tier classification
• (omit --url to print just the tier-weight reference + LLM prompt)

Run it:

# Print reference + prompt only:
curl -O https://geolocus.ai/methodology/source-grounding/reproduce.mjs
node reproduce.mjs

# Fetch a URL and list its outbound links for classification:
node reproduce.mjs --url=https://www.top10lists.us/methodology
# then paste each link into the LLM prompt to classify by tier

Download the script: reproduce.mjs →

Manual cohort rubric:

1. Render the page with a browser or bot UA; capture the visible-text + outbound-link map.
2. Walk every paragraph; flag every standalone numeric claim per the operational definition above.
3. For each flagged claim, check whether it is hyperlinked, footnoted, or named-and-cited on the page.
4. Classify each citation by tier using the URL host: .gov / .edu / peer-reviewed = T1; major institutional = T2; etc.
5. SGR = (sum of tier_weights for cited claims) / (count of all numeric claims).

For audit defensibility, publish the per-claim worksheet (claim text, classification, cited Y/N, source URL if cited, tier). The receipts.json artifact for this page does exactly that for the published cohort.

8. Limitations

1. Numeric-claim detection is human-judgment-bounded. We provide the operational definition above; reasonable analysts will disagree at the margins. Inter-rater agreement on the published cohort: ~0.91 Cohen's kappa across two independent passes.
2. Tier classification has a conservative bias. Borderline citations get the weaker tier; biases SGR down compared to a generous extractor.
3. Sample size. 2–3 pages per cohort site. Per-site rank ordering is unlikely to flip between adjacent bands; absolute SGR has ~0.05 standard error.
4. SGR doesn't measure factual correctness. A page can score 1.00 by citing a wrong primary source. SGR measures verifiability and citation discipline, not truth. For truth, the AIFS Probe layer (4-platform citation acceptance rate) remains authoritative.

Conclusion

Across a 5-site SEO industry cohort, Top10Lists.us delivers a 0.94 mean SGR — every numeric claim verifiable to a primary source — while the cohort median (n=3, Site C excluded as bot-blocked) lands at 0.26. Top10Lists.us scores ~3.6× the cohort median. The mechanism is publishable: DB-backed counts surfaced on a transparency dashboard with permalink receipts, plus a Floor+ convention preventing hardcoded stat drift. SGR is a sub-measure of the GEOlocus.ai Content Density rubric dimension; reported standalone here so the rubric is independently auditable. The full cohort range is available in receipts.json →.

Read the methodology overview at geolocus.ai/methodology →