SE Ranking AI Mode 20 Ranking Factors Study (2025-12)

Source: seranking-50-ai-mode-factors-2026-05-19.md — Yulia Deda (reviewed by Svitlana Tomko). Published 2025-12-15 on SE Ranking Blog.

SE Ranking ran an original study analyzing 50+ ranking factors that influence Google AI Mode citations and ranked the top 20 by SHAP (SHapley Additive exPlanations) attribution weight. The headline finding: global domain traffic from regular Google organic search is roughly 3x more predictive of AI Mode citation than any single on-page content factor — high-traffic domains (>1.16M visitors) average 6.4 citations vs 2.4 for low-traffic domains (<2.7K). SE Ranking also published side-by-side SHAP weights for AI Mode vs ChatGPT, which is the most useful primary-research artifact in this cluster for diagnosing engine-specific differences. Counterpart to the cross-engine Zyppy meta-analysis — primary-research, AI-Mode-specific, with measurable SHAP attributions rather than weighted-consensus scores.

Key Takeaways

• Domain traffic dominates (SHAP 0.63 AI Mode / 0.62 ChatGPT). Sites with >1.16M visitors earn ~6.4 citations vs 2.4 for sites with <2.7K visitors. Each subsequent traffic tier roughly adds one extra citation in log-transformed space, which corresponds to exponential growth in raw citation counts. This is the most reliable and consistent single-factor signal in the dataset.
• Referring domains is #2, but the AI-Mode-vs-ChatGPT split is dramatic. Referring-domain count carries SHAP 0.56 for AI Mode but 1.21 for ChatGPT — backlinks matter almost 2x more for ChatGPT citation than for AI Mode citation. Sites with >24K referring domains average 6.8 citations vs 2.5 for <300.
• AI Mode favors longer content more than ChatGPT does. Content length SHAP: 0.20 for AI Mode vs 0.06 for ChatGPT — over 3x weight differential. Sweet spot is ~1,500 words; pages under 500 words average 4.1 citations vs 5.1 for 2,300+ words.
• FAQ content beats FAQ schema. Pages with real FAQ sections in the main content average 4.9 citations vs 4.4 without (+11%). FAQ schema markup showed no meaningful AI Mode citation impact — about 3x weaker than having actual FAQ content. This is a structural-substance-over-markup finding that empirically aligns with the Ahrefs schema causal null.
• Content freshness matters more for AI Mode than for ChatGPT. Freshness SHAP: 0.071 + 0.026 for AI Mode vs 0.018 for ChatGPT. Articles updated within 2 months average 5.0 citations vs 3.9 for those untouched for >2 years. Refresh cadence of every 2-3 months maintains visibility.
• Quora and Reddit brand mentions show an inverted-U. Quora optimal at 3.8K-93K mentions (5.3 citations); >93K mentions causes slight drop. Reddit optimal at 35K-718K mentions (5.5 citations); >718K causes slight drop. Too few does not help; too many can backfire. AI Mode weighs social signals (SHAP 0.087 + 0.048) slightly more than ChatGPT (0.085 + 0.033).
• Core Web Vitals (INP/LCP) still matter, but moderately fast beats blazing fast. INP sweet spot is 0.59-1.07s (4.8-5.6 citations). Very fast pages (<0.43s INP) underperform at 2.8 citations — likely because they correlate with thinner, less-comprehensive pages. LCP >1.85s can negatively affect citations.
• Page Trust threshold is 19-23. Pages scoring <7 on Page Trust average 2.7 citations; pages scoring 24+ hit 6.2 citations. The critical inflection is in the 19-23 range — get above this band to compete for citations at all.
• Section length matters for parseability. Sections of 100-150 words between headings earn ~4.7 citations vs 4.3 for sections under 35 words (too thin) and 4.6 for sections over 150 words (slight drop, AI Mode prefers chunkable text).
• Methodology caveat: query fan-out is not modeled. The study does not account for the fact that AI Mode expands the user query into multiple sub-queries (the fan-out effect documented by AirOps). This limits completeness — SE Ranking explicitly flags this as a known limitation.

Methodology

SE Ranking used SHAP (SHapley Additive exPlanations) — a game-theoretic feature-attribution technique — to assign weights to 50+ candidate ranking factors based on how much each factor contributed to predicting AI Mode citation outcomes across the sampled domain corpus.

The target variable is citation count per domain. Because citation counts vary across multiple orders of magnitude (some domains get 1 citation, others get thousands), counts are log-transformed before modeling. When the article says “citations increased from 1.6 to 8,” those are log-transformed values — corresponding to roughly 5 raw citations vs over 3,000 raw citations. Readers should interpret all reported citation deltas as exponential growth in raw terms.

SHAP weights in the AI-Mode-vs-ChatGPT comparison table are the average absolute SHAP value per feature — higher value means the feature contributed more to the model’s prediction. SHAP is a correlational measure layered on a predictive model, not a causal estimator — it identifies factors that predict citation outcomes given the observed data, not factors that cause citation outcomes if changed. ^[inferred — SE Ranking does not explicitly frame the causal-vs-correlational distinction, but the SHAP methodology is inherently correlational]

Acknowledged limitation: query fan-out is not modeled. AI Mode expands a single user query into multiple related sub-queries before retrieval. SE Ranking’s analysis treats each query→citation observation as independent, which underestimates the impact of factors that help a page win across the expanded query space (covered by AirOps’s Fan-Out Effect study as a separate methodology).

Sample size and corpus composition (number of domains, queries, time window) are not explicitly disclosed in the SE Ranking post — readers verifying the study at depth should request these from the authors. ^[inferred — source does not disclose explicit N]

Top 20 Factors (Summary)

SE Ranking’s top 20 organizes by SHAP attribution but groups conceptually into five families:

1. Authority and traffic dominance — Domain organic traffic, homepage traffic, referring domains, Page Trust score, Domain Trust score. Together these account for the largest share of citation prediction weight.
2. Brand strength — Brand search volume (Google), Quora brand mentions, Reddit brand mentions. Inverted-U relationships for the social platforms.
3. Content depth and structure — Content length (1,500+ words sweet spot), section length between headings (100-150 words), FAQ section presence, question-formatted titles/H1s, content freshness (last 2 months optimal).
4. Semantic and meta relevance — Descriptive URL slugs, meta description cosine similarity to query, Flesch-Kincaid readability (Grade 6-8 optimal).
5. Technical performance — INP (0.59-1.07s sweet spot), LCP (<1.02s ideal, >1.85s harmful).

The full ranked SHAP weights are in the comparison table below.

AI Mode vs ChatGPT — Where Citation Factors Diverge

SE Ranking’s most valuable contribution to the cluster is this side-by-side SHAP weight table. It is the cleanest primary-research answer to the question “do AI Mode and ChatGPT cite differently?”

Factor	AI Mode (SHAP)	ChatGPT (SHAP)	Key differences
Domain traffic	0.63	0.62	Both matter; slightly more for AI Mode
Referring domains	0.56	1.21	Dominant for ChatGPT, secondary for AI Mode — 2x weight differential
Content length	0.20	0.06	AI Mode favors longer content more — 3x weight differential
Page Trust / Domain Trust	0.12	0.15	ChatGPT weighs trust slightly higher
Semantic URL & meta relevance	0.06-0.07	0.018	AI Mode emphasizes semantic relevance more
Brand signals	0.11-0.09	0.13-0.14	Both matter; AI Mode slightly more on recognition
Technical metrics	INP 0.085, LCP 0.051	INP 0.199, FCP 0.066, LCP 0.064, SI 0.025	ChatGPT uses broader technical metrics
Social signals (Quora/Reddit)	0.087, 0.048	0.085, 0.033	AI Mode slightly more influenced by social presence
Content freshness	0.071, 0.026	0.018	AI Mode favors more recently updated content

Interpretation:

• AI Mode leans more heavily on domain traffic, brand recognition, content length, semantic relevance, and freshness. It behaves like a Google-Search-anchored retrieval layer that uses an expanded set of on-page signals to rank within the already-retrieved candidate set.
• ChatGPT leans more heavily on referring domains, domain trust, and overall authority. It behaves more like a classical link-graph authority filter, with technical performance also weighted higher across a broader set of metrics (INP + FCP + LCP + SI vs AI Mode’s INP + LCP only).

The practical implication: a single SEO strategy targeting both engines is workable but not optimal. Pages targeting ChatGPT citation should over-invest in backlink acquisition and domain trust; pages targeting AI Mode citation should over-invest in content depth, freshness, and semantic URL/meta-description alignment.

Where This Sits in the Cluster

This study slots into the AI-citation ranking-factors thesis cluster as AI-Mode-specific primary research with quantified SHAP attributions — complementing rather than replacing the cross-engine work already in the cluster.

• vs Zyppy AI Citation Ranking Factors Meta-Analysis: Zyppy aggregates 54 studies into weighted 0-10 scores across ChatGPT, Gemini, and Perplexity. SE Ranking is one primary study with SHAP-attributed weights, AI-Mode-specific (Zyppy does not separately rank AI Mode). The two are highly compatible: SE Ranking’s “domain traffic dominates” finding directly mirrors Zyppy’s #2 Search Rank score (9.7/10), and the FAQ-content-beats-FAQ-schema finding mirrors Zyppy’s #20 Structured Data score (5.6/10).
• vs Ahrefs Schema → AI Citations Causal Study: Ahrefs is a matched difference-in-differences causal study finding no statistically meaningful citation lift from adding JSON-LD schema. SE Ranking’s FAQ-schema null is the correlational corollary: when SHAP attribution is computed over a passive observation corpus, FAQ schema markup shows about 3x weaker AI Mode predictive weight than actual FAQ content. The two methodologies converge on schema is not a citation lever.
• vs AirOps Fan-Out Effect ChatGPT Study: AirOps is the deepest single-engine empirical study (16,851 queries / 353,799 pages) with retrieval-rank as the dominant predictor. SE Ranking explicitly flags query fan-out as a known limitation of its methodology, so AirOps’s fan-out findings extend SE Ranking’s coverage rather than contradict it.
• vs Digital Applied 1,000 AIO Citation Pattern Study: Digital Applied measures Google AIO citations (not AI Mode) with a domain authority control. Both surfaces are Google-owned; the studies are complementary regional probes of the same underlying ranking layer.
• vs GEO-16 Framework: GEO-16 is the academic cross-sectional study covering Brave + AIO + Perplexity. SE Ranking covers AI Mode + ChatGPT — non-overlapping surfaces, complementary coverage.

Cross-method reconciliation. The cluster now reads: domain traffic + referring domains correlate strongly across every study; on-page content depth correlates more for AI Mode than ChatGPT; schema correlates weakly across studies and shows no causal lift in the one matched-DiD test. SE Ranking strengthens the engine-specific divergence finding that was harder to pin down from Zyppy alone.

Open Questions

• Sample size and corpus composition. SE Ranking does not disclose explicit N (number of domains, number of queries, time window of citation observations). The SHAP-based methodology is sound, but downstream verification — and any attempt to replicate or extend — requires these disclosures. ^[inferred]
• Stability of SHAP weights over time. AI Mode launched in May 2025 and is still rapidly evolving. SHAP weights computed in late 2025 may not generalize to mid-2026 if Google has materially changed the AI Mode retrieval architecture. Re-running the study quarterly would surface drift.
• The inverted-U on Quora/Reddit mentions. The slight citation drop at very high brand-mention volumes (>93K Quora, >718K Reddit) is interesting but underexplained in the source. Hypothesis: spam-flagging by AI Mode’s classifier on hyper-saturated brand signals — would need a separate study to confirm. ^[inferred]

Related

• Zyppy AI Citation Ranking Factors Meta-Analysis — Cross-engine 23-factor meta-analysis of 54 studies. SE Ranking is one primary study; Zyppy is the consensus rollup. SE Ranking’s per-factor SHAP weights are the kind of underlying input that gets aggregated into Zyppy-style scores.
• Ahrefs Schema → AI Citations Causal Study — Matched DiD causal study finding no schema-driven citation lift. SE Ranking’s FAQ-schema null (3x weaker than FAQ content) is the correlational corollary.
• AirOps Fan-Out Effect ChatGPT Study — Covers the query-fan-out dimension SE Ranking explicitly flags as a known limitation. AirOps’s retrieval-rank dominance finding (4.1x over all on-page signals) is the deepest empirical anchor for the “Search Rank rules everything” thesis.
• Digital Applied 1,000 AIO Citation Pattern Study — Google AIO citation patterns (sibling Google surface to AI Mode). Top 1% of domains capture 47% of citations — the same domain-traffic-dominates pattern in a different sample.
• GEO-16 Framework — Academic cross-engine study covering Brave + AIO + Perplexity. Non-overlapping coverage with SE Ranking; complementary academic anchor for the cluster.
• FLUQs Framework — Practitioner framework for AI search optimization. SE Ranking’s domain-traffic-dominates finding empirically backs FLUQs’s “win SEO first” thesis.
• Google’s Generative AI Search Optimization Guide — Google’s official position. The SE Ranking findings on Page Trust thresholds, content length, and freshness sweet spots map directly to Google’s E-E-A-T and helpful-content guidance.

Try It

1. Pull domain traffic from SE Ranking, Similarweb, or Ahrefs. Find your global organic traffic tier. If below ~10K monthly visitors, AI Mode citation is a near-impossibility regardless of other factors — domain traffic dominates by 3x. Address that first; nothing else here matters until you do.
2. Audit your highest-priority pages for FAQ content, not FAQ schema. Add real FAQ sections in the main body of pages targeting question-style queries. If you already have FAQ schema, keep it (still helps classical Google surfaces), but don’t expect AI Mode lift from it. The lift comes from the prose answers, not the JSON-LD.
3. Set a 60-90 day content refresh cadence. Articles updated within 2 months average 5.0 citations vs 3.9 for stale ones. Pick your top 20 highest-traffic pages and put them on a quarterly refresh rotation. Even minor updates with new dates count.
4. Differentiate AI Mode vs ChatGPT strategy. For ChatGPT-citation-priority pages, over-invest in backlink acquisition and domain trust (SHAP 1.21 for referring domains vs 0.56 for AI Mode). For AI-Mode-priority pages, over-invest in content depth (3x SHAP weight), freshness, and semantic URL/meta-description alignment.
5. Target the 100-150 word section length. Break long sections at headings every 100-150 words. This is the SHAP-validated sweet spot — sections under 35 words score worse (4.3 citations) and over 150 words score slightly worse (4.6 vs 4.7 at 100-150).
6. Audit Core Web Vitals for the SE Ranking sweet spot, not the Google PageSpeed maximum. Target INP between 0.59-1.07s and LCP under 1.02s. Blazing-fast pages (INP <0.43s) correlate with thin content and underperform on citations. Speed is necessary but not sufficient — comprehensiveness beats raw speed.