Citation Overlap Across 4 AI Engines:
A 10,000-Record Independent Measurement

01Study Overview

Background and Objectives

"Does a strategy that works for ChatGPT also work for Gemini?" — this is the first question any company pursuing GEO (Generative Engine Optimization) must confront. Yet, to our knowledge, no Japanese-language study has measured the degree of citation source overlap across engines using large-scale data.

In this study, we measured the overlap rate of domains cited by four major AI search engines in response to identical queries, and conducted a structural analysis of "at which layer" and "to what degree" commonality exists.

Dataset

02Domain Citation Overlap — Approximately 10%

Identical queries were submitted to all four engines, and the Jaccard coefficient (intersection over union) of cited domain sets was calculated for each engine pair. The results are as follows.

Pairwise Overlap Rates

The Structural Isolation of ChatGPT

The heatmap reveals a striking pattern: ChatGPT references markedly different sources from the other three engines. The overlap between ChatGPT and every other engine falls in the 6–7% range, and the median for ChatGPT × Gemini is 0.0% — meaning that for more than half of all queries, not a single domain was shared between the two.

By contrast, Perplexity × AIO exhibits the highest overlap at 18.9%, which is attributable to their shared reliance on Google Search as a foundational index. Even so, 18.9% means that approximately 80% of cited domains remain distinct.

03Brand Recommendation Overlap — Approximately 14%

Beyond domains (source URLs), we also measured the overlap rate of brand names recommended by each engine. Brand names were extracted from 998 response texts, and Jaccard coefficients were calculated in the same manner.

This finding carries significant implications. Even when engines consult entirely different domains (information sources), they may arrive at recommending the same brands. In other words, AI recommendations are not contingent on any particular URL; rather, they appear to be shaped by the aggregate of information distributed across the web — what we refer to below as "perception."

04Recommendation Context Alignment — Approximately 37%

We deepened our analysis one further step, measuring the alignment rate of "why each engine recommends a given brand" — the recommendation context (reason tags). From 135 response texts, we extracted recommendation context tags (e.g., "cost-effective," "beginner-friendly," "proven track record," "highly reviewed") and calculated pairwise Jaccard coefficients.

ChatGPT's isolation trend persists at the context level (31–38%), but compared to the domain level (6–7%), the gap narrows considerably. This indicates that despite drawing from entirely different information sources, ChatGPT's rationale for recommendations shares a meaningful degree of commonality with other engines.

Industry-Specific Variation in Context Alignment

The degree of context alignment varies substantially depending on industry characteristics.

05The Perception 3-Layer Structure

The analyses above reveal that cross-engine "commonality" exists at three distinct layers. We term this the "Perception 3-Layer Structure."

What the 3-Layer Structure Means

This structure carries important practical implications.

• The Context Layer (37% shared) is formed from the web's aggregate cognitive structure. It is not dependent on any particular URL. Rather, it is shaped by the cumulative body of information distributed across comparison sites, reviews, news articles, and official websites — the collective assertion that "this brand excels at X" — which governs AI recommendation rationale.
• The Brand Layer (14% shared) is determined by contextual depth. Brands mentioned across a greater number and variety of web sources are more likely to be recommended by multiple engines.
• The Domain Layer (10% shared) is governed by each engine's internal logic. Merely appearing on a particular domain is effective only for engines that happen to cite that domain.

In summary, the strategies that are effective across all engines operate at the Context and Brand layers — that is, building a state in which the brand is discussed across diverse sources in varied contexts (perception reinforcement). Placement on any single domain is effective only for the engines that cite it.

06Comparison with International Research

We compare our findings with similar studies published internationally.

ConvertMate's English-language data (11% domain overlap) and our Japanese-language data (approximately 10%) are closely aligned. The finding that cross-engine domain citation overlap hovers around 10% appears to be a universal tendency, independent of language.

SparkToro's finding of "less than 1% exact match in brand lists" is also consistent with our results. Their study further noted that "top-tier brands appear consistently," which captures the same structural phenomenon as our finding that "brand overlap is higher than domain overlap" — observed from a different analytical perspective.

07Implications for Businesses

Cross-Engine Strategy: Perception Reinforcement

The most significant practical takeaway from the 3-Layer Structure is that "perception reinforcement" is the strategy that works across all engines.

Specifically, the following activities operate at the Context and Brand layers:

• Listing on comparison and review sites — The top domains cited in common across all four engines are predominantly comparison media (see our detailed report: Top 10 Domains Cited by All 4 AI Engines)
• Information dissemination via PR platforms and YouTube — These are also platforms cited in common across all engines (see also: 7 LLMO Strategies)
• Diversified brand mentions across multiple contexts — Rather than concentrating on a single outlet, cultivate mentions across multiple platforms from varied perspectives. (See also: How to Get Recommended by ChatGPT)

Engine-Specific Strategy: Addressing the Domain Layer

While perception reinforcement is essential as a baseline, it may be insufficient in certain cases. ChatGPT, in particular, exhibits extremely low domain overlap with the other three engines (6–7%), necessitating targeted responses to ChatGPT's distinctive citation tendencies (preference for authoritative sources, English-language data, etc.).

For a detailed analysis of engine-specific citation tendencies, see Why ChatGPT and Gemini Recommend Different Brands.

Prioritization by Industry Characteristics

• Industries with established leaders (finance, recruitment, SaaS, etc.): Context alignment is high (40–50%), meaning perception reinforcement is likely to propagate across all engines
• Industries without established leaders (niche services, local businesses, etc.): Context alignment is low (0–15%), warranting a higher allocation of effort toward engine-specific strategies

08Detailed Methodology

Data Collection

Japanese-language queries were submitted to each AI search engine, and both response texts and cited URLs were collected. Multiple measurements were conducted per query to construct a high-confidence dataset.

Overlap Measurement

• Domain overlap: The set of domains cited by each engine was constructed, and the Jaccard coefficient (|A∩B| / |A∪B|) was calculated for each engine pair
• Brand overlap: Brand names were extracted from response texts, and Jaccard coefficients were calculated in the same manner
• Context alignment: Recommendation context tags (e.g., "cost-effective," "beginner-friendly") were extracted from response texts, and Jaccard coefficients were calculated