Ukrainian SEO Specialists Ranking

1. Finding candidates

Sixteen independent passes searched from different angles: agencies, conference speakers, tool builders, the diaspora, AI search, research, education, technical SEO, e-commerce and in-house, affiliate, trade media and awards, and separately the people the market under-represents. The passes could not see one another. Together they produced 564 records, which reduced to 272 people.

2. Screening and deduplication

Transliteration is this market's main trap: one person easily splits into Oleh and Oleg, Mykola and Nikolai, Kseniia and Ksenia. Names were normalised, merged by fuzzy match and checked by hand. The 182 candidates found by only one pass were screened separately: 90 promoted, 58 held, 34 rejected including 5 duplicates. The 58 held were later researched in full as well.

3. Evidence dossiers

A separate researcher per person: 6–15 dated items, each with a URL the researcher actually opened, a claim type and a confidence level. “Top 10 SEO specialists” listicles counted at most once and only as opinion. An award published by the person's own employer is not evidence for them. A talk proves the talk happened, not that its content was original.

4. Scope adjudication

Contested candidates went to a panel of three judges with different lenses: did this person do search work themselves; did they change how the field works even without running campaigns; and a prosecutor whose job was to prove they do not belong. Removal required two votes of three. Tooling, education and research are in-scope disciplines in their own right; owning a company that sells SEO is not the same as being an SEO specialist.

5. Fact-check and re-scoring

Every one of the 224 records was checked by a separate agent, and exactly one passed clean: 920 corrections, including invented transliterated names and several invented absences, where a researcher checked the wrong URL path and recorded “no publications”. Where a correction changed the evidence base, the scores were recomputed — in both directions.

6. Rater calibration

Three people received deep anchor dossiers, which every researcher then scored — 714 independent ratings of the same three people. The mean inter-rater spread was 3.87 points per dimension. Each rater's severity was measured as the median deviation from consensus, and half of it was removed from that rater's scores.

7. Normalisation and the evidence modifier

percentile(d)  rank of dimension d across all
results, influence = 50% global + 50% within discipline
m         = 0.55 + 0.45 · EQ / 100
p'(d)     = 0.5 + (p(d) − 0.5)·m   if p > 0.5
Base      = 1000 + 1000 · Σ w·p' / Σ w
weights      results 20, research 15, AI 15, innovation 12.5,
          technical 10, influence 10, commercial 7.5,
          peer 5, momentum 5

Evidence quality works as a confidence modifier, not a tenth score. Standing above the median has to be earned with evidence, so thinly evidenced strengths are pulled toward the middle. Nothing below the median is pushed down further — there, only the uncertainty band widens.

8. Virtual matches, ELO and intervals

pairs      all candidates within ±250 base points
gap(d)    = (p'A − p'B) × 100
outcome = 1.00 if ≥30 | 0.75 | 0.60 | 0.50 | 0.40 | 0.25 | 0.00
S_A       = Σ w · outcome / Σ w
E_A       = 1 / (1 + 10^((R_B − R_A)/400))
K         = 32 → 6, 30 passes, 20 shuffles
bands  300 bootstrap, σ = 3 + 12·(1 − EQ/100)

No match result is invented: everything derives from the normalised evidence. Spearman between the base and final rating is 0.9996, so the ELO stage adds robustness rather than new information. 14993 pairs were played.