A source-role count cannot tell us whether Google rank predicts AI citations. The Machine Relations Index records which domains appeared as sources across answer runs. A separate study compares individual pages by Google search-rank tier for matched queries. Both observations can be true: a source class can accumulate many citations across a market, while a highly ranked page can have better citation odds within one question. Treating either denominator as a verdict on the other would misstate the evidence.
What the Index cohort measured #
In the MRI v2 cohort ending July 19, 2026, six answer engines returned 74,597 citation events across 8,702 answer runs. The Index observed 15,756 cited domains. Its role classification groups domains by the kind of source they provide; it does not join each cited URL to its Google position for the same question. The numbers below describe a dated cohort, not the current release shown in the updating data card above.
| Source role | Domains | Runs cited across domains | Runs cited per domain |
|---|---|---|---|
| Community and social platform | 27 | 2,518 | 93.3 |
| Vendor-owned source | 823 | 8,933 | 10.9 |
| Editorial publication | 1,021 | 9,355 | 9.2 |
| Market and company database | 580 | 5,026 | 8.7 |
| Analyst and consulting research | 364 | 2,979 | 8.2 |
| Academic and government source | 346 | 2,636 | 7.6 |
| Search or media platform | 10 | 757 | 75.7 |
| Wire and press-release distribution | 9 | 200 | 22.2 |
| Other observed source | 12,576 | 34,918 | 2.8 |
Source: Machine Relations Index v2, cohort ending July 19, 2026. A domain can be cited in many runs. These are summed domain-run counts, not the 74,597 individual citation events.
Community and social platforms averaged 93.3 cited runs per domain in this cohort; the large, mixed unclassified group averaged 2.8. That approximately 33-fold difference is a description of two unequal groups, not an experiment in source role or a comparison with search rank. The unclassified group contains most of the observed domains, and its average is not a benchmark for an otherwise comparable community platform. A site's category does not, by itself, cause an engine to cite it.
Confidence grades answer another question: how much repeated evidence stands behind a domain's observed rate. In this cohort seven domains had A-tier evidence, 38 B-tier and 223 C-tier; most were still collecting. A grade is confidence in measurement, not a quality award or an SEO score. Cross-engine presence and repeat observations help describe breadth and persistence; they were not tested here as predictors against search ranking.
What page-level search-rank evidence measured #
Anthony Lee's April 2026 preprint assembled 100,411 citation events from ChatGPT, Perplexity, Claude and Google AI Mode across 2,000 queries. Its mixed-effects model used 114,034 URL–query observations after excluding the de-indexed tier. For a page ranked in Google's top three on the study's reformulated search-intent query, the odds of citation were 7.82 times those of a page ranked 11–30 (95% CI 7.28–8.39). An odds ratio is not a promise that moving an existing page will multiply its citations.
The study also reports that 75.4% of citation events went to pages beyond Google's top 30. There are far more pages in that deeper pool, so a high share of all citation events there can coexist with lower per-page odds. The preprint is observational and not peer-reviewed. Its Google rank is a proxy for the question's search intent, not a measurement of every answer engine's retrieval path. The authors note unmeasured publisher differences and dependence on query reformulation. The Index neither reproduces this regression nor measures Google positions for its domain cohort.
| Question | Unit and denominator | What this evidence can say |
|---|---|---|
| Which source classes appeared across the Index's answer runs? | Cited runs summed over domains within each role, July 2026 cohort | The observed concentration by class, with unequal class sizes |
| Does a page's Google rank tier relate to its citation odds for a matched query? | URL–query observations in Lee's four-engine preprint | A conditional association by rank tier, not an intervention's effect |
| Does source role outperform search rank as a predictor? | Would require the two variables measured on comparable pages and questions | Not established by either dataset here |
A measurement rule for publishers #
Choose the denominator that matches the decision. To evaluate a page, pair its search position and citation record on the same question, engine and dates, and compare against other pages in that rank tier. To evaluate a source domain's standing, use repeated citation rates, question segments and engine coverage in the Index. To test whether source role adds predictive power beyond rank, collect role and rank on the same URL–query observations and control for question and publisher differences. Neither a high domain citation rate nor a good Google position alone identifies why the page was selected.
The practical lesson is to maintain accessible, useful pages while publishing evidence an answer needs. It is not to abandon SEO or to buy an untested “source architecture” formula. The Index tells us where sources were selected; a search-rank study tells us about a different association. Keep them separate until a joined study can test the comparison.
FAQ #
Does Google ranking affect AI citations? #
Lee's four-engine preprint found a strong association between a page's Google rank tier and citation odds for its sampled queries. It did not randomize rank changes or measure the causal effect of moving a page. The Index cohort in this article did not measure search rank.
Do source roles predict citations better than SEO rank? #
Not from these data. The Index shows cited-run concentration by source class, with most domains in a mixed unclassified group. The rank study estimates a page-level conditional association. A fair predictive comparison needs both variables on the same observations.
Why can most citations land outside the top results? #
The pool of lower-ranked pages is much larger. Lee's preprint reports 75.4% of citation events beyond Google's top 30 while finding much higher per-page odds for top-three pages than for pages ranked 11–30. Event share and per-page odds are different denominators.
What does an Index confidence grade mean? #
It marks the evidence supporting a domain's observed citation rate, including repeated observations over distinct dates. It does not compare the domain with Google's search results, prove that its source category caused citations or certify a publisher's quality.
Correction, October 16, 2026. An earlier version claimed that source role, cross-engine recognition and temporal consistency predicted citation authority instead of search rank. The Index cohort had no matched search-rank variable and could not support that comparison. The revised article keeps its dated source-class observations, separates them from the original page-level preprint and removes vendor-relayed percentages that were not verified at their originating studies.
Dated Index cohort: May 10–July 19, 2026, MRI v2.0. The live data card above is a newer release and does not retroactively change this analysis. Current Index and methodology.
Attribution #
Published by Machine Relations Research, the research publisher of machinerelations.ai. Provenance and editorial standards.