Geo-Citation-Lab Dataset

• Geo-Citation-Lab Dataset is a public resource that supports a two-stage analysis distinguishing citation selection from absorption in AI search outputs.
• It links 602 controlled prompts with detailed citation-level features across platforms like ChatGPT, Google, and Perplexity to measure source impact.
• Empirical findings reveal a divergence between citation breadth and answer-level influence, challenging reliance on raw citation counts for GEO evaluation.

The Geo-Citation-Lab dataset is a public dataset and analysis pipeline for studying how AI search engines select and use citations. It is analyzed as the empirical basis of a two-stage measurement framework for Generative Engine Optimization (GEO) that separates citation selection, where a platform triggers search and chooses sources, from citation absorption, where a cited page contributes language, evidence, structure, or factual support to the final answer (Kai et al., 28 Apr 2026). In this framing, the dataset is notable because it links 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity to 21,143 valid search-layer citations, 23,745 citation-level feature records, 18,151 successfully fetched pages, and 72 extracted features, thereby supporting joint analysis of discoverability, source selection, and answer-level influence (Kai et al., 28 Apr 2026).

1. Dataset identity and analytical scope

Geo-Citation-Lab is presented as a public dataset and analysis pipeline for studying how AI search engines select and use citations, hosted at https://github.com/yaojingang/geo-citation-lab, with an associated public report at https://yaojingang.github.io/geo-citation-lab/04-repet/final_report.html (Kai et al., 28 Apr 2026). The dataset is not introduced as an internal benchmark with private collection conditions; rather, it is treated as a public empirical substrate for evaluating citation behavior in AI search systems.

Its central analytical role is to support a distinction between three outcomes that are often conflated in GEO discussions: a page may be discoverable, cited as a source, or absorbed into the generated answer. The paper’s main methodological claim is that GEO should be measured as a two-stage process. The first stage is citation selection, operationalized through search triggering and citation counts. The second stage is citation absorption, operationalized through a citation-level influence proxy. This separation is the core reason the dataset is consequential: it contains linked prompt-level, citation-level, and fetched-page records sufficient to study both stages in one pipeline (Kai et al., 28 Apr 2026).

A common misconception is that citation count alone is an adequate measure of GEO performance. The dataset is explicitly used to argue otherwise. Per the reported analyses, platforms that cite more sources do not necessarily exhibit deeper use of those sources in generated answers. This suggests that raw citation breadth and answer-level citation depth are analytically distinct quantities rather than interchangeable proxies.

2. Dataset composition and prompt architecture

The dataset statistics reported in the paper are concise and central to its identity.

Component	Reported count	Notes
Prompts	602	Controlled prompts
Platforms	3	ChatGPT, Google AI Overview/Gemini, Perplexity
Cleaned search-layer rows	21,181	Search-layer data
Valid search-layer citations	21,143	Citation-domain rows
Citation-level feature records	23,745	Feature table
Successfully fetched citation pages	18,151	fetch_ok pages
Extracted features	72	Feature dimensions

The 602 prompts are organized into four layers: A/B/C/D = 432/60/60/50 (Kai et al., 28 Apr 2026). The paper characterizes them as follows. A layer contains the main experimental prompts. B layer introduces style contrast. C layer introduces language contrast. D layer introduces realistic and extreme scenarios. The dataset therefore encodes not only citation outcomes but also a designed prompt space with controlled contrasts.

The prompt layers support several specific experimental dimensions. In Layer A, the paper reports controlled or observed factors including task type, trigger strength, time sensitivity, industry, and subtask. In Layer B, prompt styles include natural phrasing, explicit source request, and expert-role prompt. In Layer C, the design uses a Chinese-English pair contrast. In Layer D, scenario types include high-risk, ambiguous, multi-constraint, long-decision, and macro-trend (Kai et al., 28 Apr 2026).

The paper also reports platform-specific observed prompt counts after cleaning: ChatGPT: 587 observed prompts, Google: 602 observed prompts, and Perplexity: 602 observed prompts. This difference reflects cleaning issues and missing outputs rather than a different experimental design.

3. Collection workflow, schema, and feature families

The dataset is described as the result of a controlled prompting workflow across the three platforms. At a high level, the process is: define a controlled prompt set, issue prompts to each platform, record whether search was triggered, capture citation rows, clean and normalize those citation records, fetch cited pages, extract page content and metadata, compute citation/page-level features, and construct an answer-level influence proxy (Kai et al., 28 Apr 2026).

At the prompt-platform level, the paper defines search triggering as:

$$S_{p,e}=\mathbb{1}\{\text{any cleaned row for prompt }p\text{ on platform }e\text{ indicates search triggered}\}$$

and citation count as:

$C_{p,e}=\sum_i ValidCitation_{i,p,e}$

These quantities make the selection stage explicit. A prompt-platform pair may trigger search without yielding the same number or type of cited sources across systems.

The paper provides a compact data dictionary in terms of field families rather than a full published schema. These families include:

• Prompt metadata: layer, industry, question type, language, style
• Platform metadata: platform, prompt id, response id
• Search-layer citation: citation domain, citation [URL](https://www.emergentmind.com/topics/unidirectional-reflection-lasing-url), search triggered
• Source metadata: website type, country, language, Final_DR
• Fetch status: fetch_ok, fetched_html, error state
• Page structure: word count, headings, paragraphs, list density
• Evidence genre: definitions, numbers, comparisons, how-to, code, Q&A
• Semantic alignment: embedding similarity, LLM relevance, LLM quality
• Influence components: ref_count, first_position_ratio, paragraph_coverage_ratio, tfidf_cosine, bigram_overlap, trigram_overlap (Kai et al., 28 Apr 2026)

The paper states that the public feature table contains 72 extracted features, but it does not enumerate all 72 individually. This suggests that the schema is richer than the subset of named variables discussed in the manuscript. A plausible implication is that Geo-Citation-Lab should be understood as a linked observational dataset with several granularities: prompt-level records, search-layer citation rows, fetched-page records, and citation-level feature rows.

The collection and preprocessing discussion also identifies concrete cleaning issues: repeated header rows in the ChatGPT CSV, normalization issues for ChatGPT A_news and A_technology naming, 15 missing ChatGPT prompt outputs after cleaning, and unknown values in country and language fields (Kai et al., 28 Apr 2026). These are not peripheral details; they materially affect denominators and cross-platform comparability.

4. Citation selection and citation absorption framework

The defining methodological contribution associated with Geo-Citation-Lab is the separation between citation selection and citation absorption (Kai et al., 28 Apr 2026). Citation selection concerns whether a system triggers search and how many sources it cites. Citation absorption concerns whether a cited source substantively shapes the generated answer.

The paper operationalizes absorption with a citation-level influence score. For each citation-page $i$, the public feature table computes:

$$\begin{aligned} Influence_i = &\;0.20\cdot \min(ref\_count_i/3,1) \ &+0.15\cdot (1-first\_position\_ratio_i) \ &+0.20\cdot paragraph\_coverage\_ratio_i \ &+0.25\cdot tfidf\_cosine_i \ &+0.20\cdot \frac{bigram\_overlap_i+trigram\_overlap_i}{2}. \end{aligned}$$

The variables named in this formula are themselves present as feature components in the dataset. The score rewards a source that is referenced multiple times, appears earlier, is used across more answer paragraphs, and shows stronger lexical similarity to the answer. The paper is careful to treat this as an observational proxy, not as direct evidence of hidden model attention.

The framework also motivates two model classes for future confirmatory analysis. For citation breadth, the paper proposes a negative binomial model:

$C_{p,e}\sim NegativeBinomial(\mu_{p,e},\alpha)$

$$\log\mu_{p,e}=\beta_0+\beta_{platform[e]}+\beta_{layer[p]}+\beta_{style[p]}+\beta_{language[p]}+\beta_{industry[p]}+\beta_{interactions}+u_p.$$

For absorption, it proposes a fractional-logit or beta-regression style specification:

$$\operatorname{logit}(\mathbb{E}[Influence_i])=\gamma_0+\gamma_{platform[e_i]}+\gamma_{domain\_type[d_i]}+\gamma_{structure}^T X_i+\gamma_{genre}^T G_i+\gamma_{semantic}^T Z_i+\epsilon_i.$$

These models are not the dataset itself, but they clarify what the dataset was built to support: platform comparison, prompt-condition comparison, source-type analysis, and source-influence analysis under explicit variable families.

5. Empirical patterns reported from the dataset

The central descriptive finding is that citation breadth and citation depth diverge (Kai et al., 28 Apr 2026). At the search-selection level, mean citations per prompt are reported as:

• ChatGPT: 6.88
• Google AI Overview/Gemini: 12.06
• Perplexity: 16.35

The corresponding medians are 6, 12, and 17. On breadth alone, Perplexity and Google appear to cite substantially more sources than ChatGPT.

At the absorption stage, among fetch-ok citations, the reported influence means and medians are:

• ChatGPT: $N = 3,323$, mean influence 0.2713, median 0.2611
• Google: $N = 6,385$, mean influence 0.0584, median 0.0515
• Perplexity: $N = 8,443$, mean influence 0.0646, median 0.0333

This is the empirical basis for the paper’s claim that ChatGPT cites fewer sources but shows substantially higher average citation influence among fetched pages.

Search triggering is reported as near-universal:

• ChatGPT: 579/587, 98.64%
• Google: 600/602, 99.67%
• Perplexity: 602/602, 100%

This shifts analytical emphasis away from whether search occurs and toward what happens after it occurs.

The dataset also supports several controlled comparisons. In the B-layer style contrast, citation means differ by style and by platform. For example, ChatGPT shows 7.30 for natural phrasing, 6.15 for explicit source request, and 7.95 for expert-role prompting, whereas Google shows 14.05, 15.90, and 10.40, and Perplexity shows 15.70, 17.15, and 16.70. In the C-layer language contrast, the reported means are 7.77 vs 7.03 for ChatGPT, 7.53 vs 11.57 for Google, and 15.93 vs 16.43 for Perplexity when comparing Chinese and English. In D-layer multi-constraint scenarios, the means are 3.4 for ChatGPT, 12.6 for Google, and 17.7 for Perplexity (Kai et al., 28 Apr 2026).

Source composition is also quantified. Official, news, and vertical sources account for 87.52% of ChatGPT selections, 87.34% of Google selections, and 79.12% of Perplexity selections. Among identifiable samples, the reported US share is 82.70%–86.76%, and the English share is 82.90%–95.07%. Frequently selected domains include youtube.com (560), en.wikipedia.org (352), reddit.com (315), reuters.com (287), and linkedin.com (187), among others (Kai et al., 28 Apr 2026).

A further substantive finding concerns the kinds of pages associated with higher absorption. Comparing top and bottom influence quartiles, the paper reports large differences in word count, heading total, paragraph count, list density, answer-citation semantic similarity, LLM relevance score, and LLM content quality. It also reports positive influence associations for pages that contain code, contain numbers/statistics, contain definition markers, contain comparison content, and contain how-to content, while Q&A format shows a negative association (Kai et al., 28 Apr 2026). This suggests that GEO success in answer absorption is tied not merely to being cited, but to providing modular, semantically aligned, extractable evidence.

6. Position within adjacent citation datasets and methodological limitations

Geo-Citation-Lab belongs to a broader emerging family of citation-centered resources, but its scope is distinctive. Unlike the AA Citation Corpus, which is a paper-level NLP citation corpus with inferred geographic annotations derived from the ACL Anthology (Rungta et al., 2022), Geo-Citation-Lab is centered on AI search platforms, prompt-controlled outputs, and answer-level citation influence. Unlike the MIMIQ benchmark, which is document-centric and designed for held-out-query optimization and citation-failure diagnosis (Tian et al., 10 Mar 2026), Geo-Citation-Lab emphasizes descriptive cross-platform measurement of source selection and source absorption. Unlike the GEO-16 audit corpus, which studies URL citation behavior in English-language B2B SaaS and scores pages with a 16-pillar framework (Kumar et al., 13 Sep 2025), Geo-Citation-Lab joins prompt architecture, citation records, fetched content, and answer-linked influence in one observational dataset.

The paper is explicit about several limitations. It treats the dataset as a static snapshot and notes the lack of unified record-level timestamps. It also cautions that fetch failures affect absorption analysis because only fetch_ok pages are included there. Taxonomies such as website type are noisy, and country/language fields include unknown values. The 15 missing ChatGPT prompt outputs after cleaning mean platform denominators differ. Most importantly, the influence score is a constructed proxy, not direct model internals (Kai et al., 28 Apr 2026).

A plausible implication is that Geo-Citation-Lab is strongest as a measurement substrate and descriptive benchmark, rather than as a finished causal dataset. The paper itself recommends a final confirmatory version that reruns raw CSVs with locked scripts and versioned scoring models. This implies that the present release should be read as analytically rich but still sensitive to platform drift, fetch availability, and model-version dependence.

In sum, Geo-Citation-Lab is best understood as a structured observational resource for measuring how AI search platforms move from search triggering, to source selection, to answer-level source use. Its importance lies less in any single count than in the way it makes these stages jointly observable, and in the accompanying argument that GEO should be evaluated beyond citation counts alone (Kai et al., 28 Apr 2026).