What Separates the Google Top 20 from Positions 21–50? On-Page, Topical, Intent and Technical Features of 7370 Result URLs, with a Comparison to Yandex

Search engine optimisation checklists ask for the exact query in the title and H1, a word count matched to competitors, full coverage of the user's intent and a light, fast page. This preprint tests those items against one snapshot of Google results: 160 Russian-language queries in 16 niches, positions 1–50, Moscow, desktop, 14 September 2026 (7401 result rows, 7370 URLs, 4432 domains). For every URL the study measured on-page text features, coverage of a query-specific list of topical terms ("LSI coverage"), a technical audit and host-level network properties; 1228 Google documents and 634 Yandex documents were scored for intent coverage by an LLM assessor blind to rank, domain and search engine. The main contrast is a page in the top 20 against a page from positions 21–50 of the same query (AUC with a query-level bootstrap and a within-query permutation test). No feature separates the groups strongly. LSI coverage, the hypothesis treated as confirmatory, is higher in the top 20 (AUC 55.3, 95% CI [53.6; 56.9], permutation p = 0.0002), holds when text length is controlled, in 15 of 16 niches and without aggregators; its within-domain gap is +0.026 with upper medians and +0.015 [0.000; 0.032] with conventional medians. Intent coverage is not reliably higher (AUC 52.7 [49.5; 55.6], p = 0.063) and adds nothing at equal length. The full query in the title or H1 performs at chance level (within-query pair share 50.6 and 50.0 of 100). Top-20 pages are larger (HTML 246.3 vs 207.1 KB), but only 15% of that gap remains between pages of the same site, and server response time does not differ. At a common boundary with a Yandex snapshot of the same queries (positions 1–10 vs odd 31–49), both engines favour larger pages; in Google the text separates the top more, in Yandex the query in the title and description does. The record contains the preprint (PDF) and a replication package: SERP snapshot (query, position, URL, domain), per-page on-page, LSI and technical features, host-level network data, stub and exclusion flags, LSI term lists, frozen intent maps, 0/1/2 scores of two assessor runs (weighted kappa 0.863), Yandex comparison features, original analysis logs, the pre-analysis plan with amendments (Russian) and code. code/verify_headline_numbers.py recomputes 88 headline numbers from the CSV files and exits with code 0 when they match the original logs. Raw HTML and page texts are third-party content and are not included. Text, figures and data: CC BY 4.0; code: MIT. Observational design, one date, one region, desktop only; no causal claims. A Russian-language article for practitioners based on the same data: sk-seo.ru/blog/chto-otlichaet-top-20-google.html.