Source: zyppy-ai-citation-ranking-factors-2026-05-07.md + raw/newsletter-zyppy-signal-203893f8bc.md — Cyrus Shepard, Signal by Zyppy. Published 2026-05-07.

Corrected 2026-07-24

The 23-factor table and several takeaways were rebuilt from the recovered primary source. A prior version of this article carried fabricated middle-tier factors (e.g. “Direct Quote Density,” “Page Speed,” “Mobile UX,” “Anchor Text,” “Image Alt Text,” “Backlinks”) and invented per-engine ChatGPT/Gemini/Perplexity score columns that do not appear in Shepard’s post. Shepard publishes a single score per factor (9.5 down to 2.0), not per-engine scores. The table below now matches the source exactly.

Cyrus Shepard (Zyppy SEO) downloaded nearly every published AI-citation experiment, study, explainer, and patent from the prior ~2 years (across ChatGPT, Gemini, and Perplexity), narrowed to the 54 most salient sources, cross-referenced their findings, and hand-scored a 23-factor ranking on a single 0-10 scale. The strongest signal: AI citation engines re-rank on top of classical search relevance, so winning organic SEO is the precondition. The most-hyped 2025 tactics — schema and LLMs.txt — score near the bottom (5.6 and 2.0). Shepard’s thesis: “win SEO, win AI citations (most of the time, with extra steps).”

This is a meta-analysis / secondary synthesis, not new primary data. Scores are Shepard’s manual judgment, weighted by his three criteria (below); he states plainly “these aren’t ‘Ranking Factors’ in the traditional sense… Correlation is not causation.” Treat it as an evidence map over the cluster, not as an independent study.

Key Takeaways

  • Top tier (9.0+): URL Accessibility (9.5 — can the bot reach the page?), Search Rank (9.4 — does it already rank on Google?), Fan-out Rank (9.3 — does it rank for the query’s fan-out expansions?), Preview Control (9.2 — nosnippet/data-nosnippet can suppress your own visibility), Query-Answer Match (9.2 — page content is semantically close to the query and the answer), Intent-Format Match (9.0 — listicle for “best,” step-by-step for “how-to”).
  • Upper-mid tier (8.0-8.9): Topic Cluster Ranking (8.9), Answer Near the Top (8.8 — Gemini applies a strict per-URL retrieval cap, so top-of-page content is likelier to be cited), AI-ready Structure (8.6 — headings/sections/tables), Factually Specific (8.3), Explicit Phrasing (8.1 — definitive claims beat hedged ones), Cites Sources (8.0), Self-Contained Passages (8.0).
  • Lower-mid tier (6.3-7.6): Content Visibility (7.6 — visible HTML text, not JS-hidden), Freshness (7.0), Brand / Entity Trust (6.8), Length (6.7 — longer tended to help but evidence was inconsistent, and length reduces the share of a page that gets retrieved), Language (6.3 — engines bias toward the query’s language/locale).
  • Bottom tier (under 6): Entity Consistency (5.8), Structured Data (5.6), Known Source (5.4 — already in the model’s training data), Domain Authority (5.0 — relationship found but “often weak”), LLMs.txt (2.0 — no credible evidence it influences citations at all; the most overhyped tactic of 2025-26 by Shepard’s read).
  • Structured Data scores 5.6 (#20 of 23). Shepard’s read: “practically every study that looks at schema and AI citations finds a positive relationship. The effect is typically small, but it’s amazingly consistent.” Empirically corroborated by the Ahrefs schema causal study (2026-05-11), which found no statistically meaningful AI-citation lift from adding JSON-LD — consistent with schema being a small, correlational marker rather than a lever.
  • The through-line is classical SEO. Shepard’s own summary: the factors reduce to Relevance, Trust, Topical Authority, and Extractability — signals that “should align with current SEO thinking.” “Win SEO, win AI citations (most of the time, with extra steps).”
  • Why a meta-analysis and not a single study. Each of the 54 underlying studies has small-sample or selection issues; cross-referencing which findings recur across studies, surfaces, and methodologies is more robust than trusting any one case study — at the cost of the scores being subjective aggregate judgments rather than measured effect sizes.

The 23 Ranking Factors

Scored on three criteria: Repeatability (how often a similar finding recurs across studies + consistency of direction), Strength of Evidence (a 50-million-query study outweighs a 10-query case study), and Official Support (docs, technical specs, patents). Shepard assigned each score by hand, using AI to fine-tune. A single score per factor; no per-engine breakdown exists in the source.

#FactorScoreDefinition
1URL Accessibility9.5Page is available and crawlable during training/grounding
2Search Rank9.4How the URL ranks for the exact query
3Fan-out Rank9.3How the URL ranks for related fan-out queries
4Preview Control9.2Preview directives (nosnippet, data-nosnippet) can suppress visibility
5Query-Answer Match9.2Page content closely matches the query (primary/fan-out)
6Intent-Format Match9.0Page type matches query intent (listicle for “best,” etc.)
7Topic Cluster Ranking8.9Site ranks for multiple related queries (primary + fan-out)
8Answer Near the Top8.8Content near the top of the page is likelier to be cited
9AI-ready Structure8.6Formatted so AI can extract sections (headings, tables)
10Factually Specific8.3Shows specific, verifiable facts
11Explicit Phrasing8.1Definitive claims over vague/hedged statements
12Cites Sources8.0Facts backed with referenced sources
13Self-Contained Passages8.0Key statements stand alone without extra context
14Content Visibility7.6Important text in visible HTML, not hidden/JS-gated
15Freshness7.0How current the information is (varies by query)
16Brand / Entity Trust6.8How much the engine knows about and trusts the brand
17Length6.7Word count (longer tended to help, but inconsistent)
18Language6.3Language/locale of the content vs the query
19Entity Consistency5.8Consistent naming for brands, people, products
20Structured Data5.6Schema to identify entities and support content
21Known Source5.4URL already known to the engine via training data
22Domain Authority5.0Link-based popularity measure (relationship “often weak”)
23LLMs.txt2.0Hosting an LLMs.txt file (no credible supporting evidence)

Tactical Implications

  • Win SEO first. Search Rank is #2 (9.4). If you aren’t ranking organically, AI citations are near-impossible regardless of everything else. This collapses the “AEO is separate from SEO” narrative. Supporting data Shepard cites: Ahrefs found 38% of AI Overview citations come from Google’s top 10; AirOps found a strong retrieval-rank→ChatGPT-citation relationship; Semrush found Perplexity answers had ~82% overlap with Google’s top 10.
  • Optimize for query fan-out. Fan-out Rank is #3 (9.3). Engines expand the query into sub-queries; pages ranking for the head term and its expansions get cited more. Cluster content around topical concepts, not just exact-match keywords. (See the companion Zyppy AI Citation Playbook for Shepard’s step-by-step fan-out workflow.)
  • Control your preview. Preview Control (#4, 9.2) — a nosnippet/data-nosnippet on important text can suppress your own AI visibility. Audit for accidental suppression.
  • Put the answer near the top and make passages self-contained. Answer Near the Top (8.8) + Self-Contained Passages (8.0): engines don’t retrieve the whole page (Dan Petrovic’s Gemini retrieval-cap research), so front-load the citable claim and make it stand alone.
  • Format-match the intent. Intent-Format Match (#6, 9.0). “How to” → numbered steps; “what is” → definition + example; “compare” → table. Wrong format = no citation even when the content is correct.
  • Skip LLMs.txt. Score 2.0 (#23). No credible evidence. Reallocate the hours to factors 1-9.
  • Schema for the right reasons. Score 5.6 (#20). Add schema because Google still rewards it on classical surfaces — not for AI-citation lift. The Ahrefs causal study is the independent confirmation.

Open Questions

  • The exact 54 underlying studies. Shepard links a public spreadsheet of all 54; worth pulling for citation-chain verification.
  • How scores decay over time. Engine retrieval changes monthly. The 2.0 for LLMs.txt assumes engines aren’t using it as of May 2026 — if any engine starts honoring it, this jumps.
  • Weighting by vertical. All 23 factors are averaged across studies. Per-vertical weights (medical, legal, e-commerce) would likely shift the ranking — medical content is cited more conservatively, so Brand/Entity Trust and Topical Authority probably weight higher there.

Try It

  1. Pull your own top-10 ranking pages from GSC. Cross-reference against the 23 factors. The 1-2 lowest-scoring factors on your pages are your highest-leverage fixes.
  2. Audit title tags and meta descriptions for Preview Control (#4). Rewrite any preview that doesn’t literally answer the page’s primary query, and check for accidental nosnippet/data-nosnippet on important text.
  3. Check fan-out coverage. Run your head term through Google AI Mode / a fan-out tool and capture the expanded sub-queries. Does your page address them? If not, expand sections or add an FAQ block. (Full workflow in the playbook.)
  4. Stop new LLMs.txt projects (score 2.0). Reallocate to factors 1-9.
  5. Re-frame schema work. Keep existing schema (still helps classical Google). Don’t expand it chasing AI citations.