Demand-Side Measurement for Generative Engine Optimization: Constructing and Validating a Million-Persona, Intent-Annotated Buyer Corpus
PersonaGen-1M uses MinHash LSH and semantic deduplication to build 1.03M intent-annotated buyer personas.
Key Findings
Methodology
The authors aggregate roughly 40 million descriptions from NVIDIA Nemotron-Personas-USA, BSC-LT m-Personas, Orange PersonasForSalesbot, and Tencent PersonaHub. They apply GPU MinHash LSH with 128 permutations, 8 bands of 16 rows, and Jaccard threshold 0.9, followed by semantic deduplication using KaLM-Embedding-Gemma3-12B embeddings and cosine threshold 0.9. Grok-4-1-fast then enriches records into enforced JSON at temperature 0.7.
Key Results
- PersonaGen-1M contains 1,031,732 personas across 511 industries and four market contexts, with 19,416,821 behavioral attributes: 5,160,046 queries, 5,132,320 information needs, 3,124,904 goals, 3,115,862 pain points, and 2,883,689 uncovered needs.
- Primary intent is 78.3% informational, 17.4% commercial, and 4.3% transactional; the commercial slice is about 179,600 personas and nearly 0.9 million queries. In a FinTech pilot, 150 queries produced brand mentions in 97.3% of answers, averaging 5.01 brands.
- Normalized entropy is 0.37 for industry, 0.58 for market, 0.73 for gender, and 0.57 for intent. Industry × intent has Cramér’s V=0.299. Two brand extractors agree exactly on 36.0% of answers, with mean Jaccard agreement 0.737.
Significance
The paper shifts GEO measurement from observing generated answers alone toward structured demand-side sampling. Query language explains 26.5% of single-response variance, whereas brand identity explains only 1.5%; therefore buyer-framed, industry-specific queries are a more defensible probe frame than analyst-written prompts. The preferred_sources field also enables comparison between buyer trust and engine citation behavior. This addresses a persistent gap between what users seek and what generative systems recommend.
Technical Contribution
The contribution is measurement infrastructure rather than a new language model. Each record joins primary_intent, typical_queries, and named preferred_sources. MinHash LSH provides scalable lexical candidate generation; KaLM-Embedding-Gemma3-12B removes semantically equivalent records; Grok-4-1-fast performs schema-constrained enrichment. First-pass schema compliance is 98.5%, rising to 99.8% after one retry, and quality filtering leaves 100% complete final rows. The design makes demand-to-supply joins operationally possible.
Novelty
Compared with PersonaHub, FinePersonas, and Nemotron-Personas-USA, PersonaGen-1M combines buyer queries, staged commercial intent, and named source preferences in one record. MatrAIx Persona 1M has intent and source preference, but uses conversational intent labels, a six-value media-channel enum, and no stored queries. The fundamental novelty is therefore the integrated, join-oriented schema for GEO demand measurement, not scale alone.
Limitations
- The corpus is entirely synthetic: queries are plausible constructions rather than observed logs. It can support hypothesis generation and system testing, but cannot directly estimate real buyer intent, market demand, or causal brand visibility.
- The sample is highly concentrated: General, EdTech, and Consulting account for 60.2% of personas, the top 15 industries account for 94.9%, and B2C represents 70.6%. These source-distribution effects limit enterprise and long-tail generalization.
- One primary intent covers a persona’s query set, usually about five queries, so the representation cannot fully capture multi-intent, multi-turn sessions. Effective age also includes caregiver-age imputation for 112,850 sub-13 records.
Future Work
The authors leave the central controlled estimate for future work: join commercial-intent queries with supply-side brand recommendation shares and citation provenance. Strong follow-up studies should benchmark against real commercial search logs, use repeated and paraphrased probes, perform cross-vendor brand extraction, and stratify by language, industry, and market. Rebalanced sampling is also needed before making population-level claims.
AI Executive Summary
Generative engines are turning search into a direct recommendation. When a buyer asks ChatGPT, Gemini, or Perplexity which tool to purchase, the system often names a short list of brands. For brands, visibility is therefore no longer equivalent to ranking a web page: it means being selected inside a synthesized answer. Yet large persona resources such as PersonaHub were designed mainly to diversify training data. They generally lack buyer queries, staged purchasing intent, and trusted-source preferences—the variables needed to explain why a brand is recommended or omitted.
Żatuchin and Dzemesjuk introduce PersonaGen-1M, a corpus of 1,031,732 synthetic buyer personas. Starting from approximately 40 million public descriptions, the pipeline applies GPU-accelerated MinHash LSH for lexical deduplication, KaLM-Embedding-Gemma3-12B for semantic deduplication, and Grok-4-1-fast for schema-constrained enrichment. The result spans 511 industries and four market contexts, and contains 5,160,046 search queries. Each persona receives a primary intent label and a named preferred_sources list, allowing researchers to build industry-specific commercial query banks and compare user trust with engine citations.
The corpus is an instrument, not a population census. In a FinTech pilot, 150 queries caused 97.3% of answers to mention at least one brand, with 5.01 brands per answer on average. But the data are synthetic, industry coverage is highly concentrated, and two brand extractors agreed exactly on only 36.0% of responses. The paper’s main achievement is thus a reusable demand-side measurement framework; whether persona alignment changes recommendation share remains an explicit empirical question.
Deep Analysis
Background
GEO studies how brands appear in generated answers. Aggarwal et al. named GEO as an optimization problem, while related supply-side work measures citations and brand concentration. A study of 12,933 answers found query language explained 26.5% of response variance, versus 1.5% for brand identity. This makes analyst-authored prompts a consequential source of bias. Traditional personas are small; PersonaHub scales to a billion but lacks staged buyer intent, trusted sources, and query sets.
Core Problem
The problem is to create a scalable, stratified demand-side query frame that can be joined with supply-side brand recommendations and citations. Key bottlenecks are massive redundancy, schema consistency, adaptation of keyword-era intent labels to conversational behavior, and response variation caused by wording, language, turn position, and sampling randomness. A useful corpus must preserve buyer context while remaining computationally tractable and auditable.
Innovation
- �� Joint schema: primary_intent, typical_queries, and named preferred_sources are stored per persona.
- �� Stratification: 511 industries and four market contexts enable vertical-specific sampling.
- �� Hybrid deduplication: MinHash LSH handles lexical similarity; embeddings handle semantic equivalence.
- �� Measurement alignment: named source lists can be paired with citation-provenance data, unlike a coarse media-channel enum.
- �� Reproducibility: the 14,955-row PersonaGen-15K subset exposes the protocol and validation.
Methodology
- �� Input: four public HuggingFace corpora, totaling about 40 million descriptions.
- �� Lexical filtering: GPU MinHash with 128 permutations, 8 bands × 16 rows, Jaccard θ=0.9, processed in 500,000-record chunks; output about 4.2 million records.
- �� Semantic filtering: KaLM-Embedding-Gemma3-12B, 4-bit quantization, and Flash Attention 2; pairs above cosine similarity 0.9 are removed, leaving roughly one million descriptions.
- �� Enrichment: Grok-4-1-fast in enforced-JSON mode, temperature 0.7, batches of 100 with checkpointing. Initial compliance is 98.5%; one retry raises it to 99.8%.
- �� Validation: schema completeness, normalized Shannon entropy, chi-square tests, and Cramér’s V. Consistency filtering removes 1.5%, yielding 1,031,732 complete rows.
Experiments
All descriptive statistics use the full 1,031,732-persona corpus. The authors report entropy across industry, market, gender, and intent, and test categorical associations. For an instrument pilot, they select 7,912 commercial-intent FinTech personas with three to five queries, randomly sample 30, and submit 150 queries to one commercial engine at temperature 0.7. Answers are capped at 120 words. Gemini-3.5-flash and Gemini-2.5-flash independently extract brands, with consensus used for reporting. No real-log benchmark or controlled causal recommendation-share experiment is performed.
Results
Industry entropy is only 0.37, confirming severe head concentration; gender entropy is 0.73. Industry × intent gives Cramér’s V=0.299, while commercial intent is 32.0% in B2B versus 12.4% in B2C, suggesting industry-appropriate enrichment. The pilot yields 751 brand mentions across 530 distinct brands. Brand extractors agree exactly on 36.0% of answers and have mean Jaccard 0.737, demonstrating both instrument utility and nontrivial measurement uncertainty.
Applications
Researchers can filter the 17.4% commercial slice to construct industry-specific buyer-framed probes, then measure recommendation share with repeated calls. Content and communications teams can compare preferred_sources distributions with actual engine citations to identify trust gaps. Proper use requires fixed language, model, turn position, repetition, and paraphrase protocols; corpus outputs should be treated as synthetic test stimuli, not direct market-share estimates.
Limitations & Outlook
Synthetic queries lack behavioral calibration against real logs, while upstream corpora create strong industry, demographic, and B2C biases. A single intent label compresses multi-stage conversational behavior. Enrichment depends on Grok-4-1-fast, and brand extraction uses two models from the same vendor family. Future work should benchmark real commercial logs, add cross-vendor extraction, rebalance long-tail industries, and preregister repeated, paraphrase-controlled recommendation-share studies.
Plain Language Accessible to non-experts
Imagine a giant shopping mall building a practice crowd before testing its recommendation screens. The researchers collect about 40 million rough customer cards from four public collections. They throw away cards that are nearly identical, then remove cards that use different words but describe the same kind of customer. A computer fills the remaining cards with likely questions, concerns, goals, buying stage, and the kinds of information each customer trusts.
The finished box contains 1.03 million cards covering 511 industries. Each card says whether the customer is learning, comparing products, or ready to buy. Researchers can therefore ask an AI the kinds of questions different customers might actually ask, rather than inventing one favorite prompt, and observe which brands appear in the answer.
But this is a practice crowd, not a real survey. The customers were generated by software, some industries have far more cards than others, and even two software readers sometimes disagree about brand names. The box is useful for designing fairer tests and finding gaps, but it cannot prove what all real shoppers believe.
ELI14 Explained like you're 14
Think of a game where you type a shopping question and an AI recommends brands. You ask, “Which project-management app should a 30-person design agency use?” and the AI gives a few names. But if you change the wording, language, or situation, the answer might change. The researchers built a huge virtual crowd so they could test many realistic customer questions.
They started with about 40 million character cards, deleted repeats, removed cards that meant the same thing, and ended with 1.03 million. Each virtual person has a job, problems, goals, possible questions, and a note about whether they trust reviews, communities, analyst reports, or publications. The card also says whether they are learning, comparing, or ready to buy.
Then they tested 150 FinTech questions. Almost every answer—97.3%—mentioned at least one brand, and each answer named about five brands. Cool, right? But these people are computer-made, not real shoppers. Also, two programs agreed perfectly about the brands in only 36% of answers. So this is like a training arena for testing AI recommendations, not a magic crystal ball for the real world.
Glossary
Generative Engine Optimization
The practice of increasing a brand’s probability of appearing in a generated answer from systems such as ChatGPT or Gemini. Its outcome is recommendation inclusion rather than a ranked web URL.
The paper frames PersonaGen-1M as a demand-side instrument for GEO.
MinHash LSH
An approximate similarity-search method that uses signatures and buckets to find texts with high Jaccard similarity efficiently. The paper uses 128 permutations and 8 bands of 16 rows at threshold 0.9.
It performs the first, lexical deduplication stage.
Semantic deduplication
Removal of records that are meaningfully equivalent even when their wording differs, using embeddings and cosine similarity. It complements lexical matching.
KaLM-Embedding-Gemma3-12B embeddings and cosine threshold 0.9 are used.
Primary intent
One dominant stage label assigned to a persona’s query set: informational, commercial, transactional, or navigational. It is not a fully multi-label annotation of every individual query.
The 17.4% commercial subset is intended for recommendation-oriented probing.
Preferred sources
A named list of source types that a persona would trust, such as analyst reports, peer reviews, communities, or publications. It represents demand-side source preference.
It is designed to join with supply-side citation provenance.
Cramér’s V
An effect-size statistic measuring association between categorical variables, especially useful when huge samples make tiny effects statistically significant. Larger values indicate stronger association.
Industry × intent has V=0.299; market × intent has V=0.164.
Open Questions Unanswered questions from this research
- 1 Do synthetic queries resemble real commercial searches? The paper does not yet provide a systematic comparison with observed logs for the same industries, languages, and purchase stages.
- 2 Can preferred_sources predict citation behavior? This requires joining persona preferences with industry-level provenance data and controlling for model, language, repetition, and prompt wording.
Applications
Immediate Applications
Industry-specific GEO probe banks
GEO researchers can select commercial-intent personas in a target vertical, submit their three-to-ten buyer-framed queries to several engines, and measure brand inclusion, variability, and citations. Model, language, turn position, repetitions, and paraphrases should be fixed in advance.
Trust–citation gap analysis
Content teams can aggregate preferred_sources for an industry and compare them with the sites engines actually cite. The resulting gaps can guide review outreach, community participation, analyst relations, and evidence-oriented content development.
Long-term Vision
Integrated demand–supply measurement
A mature platform could connect persona queries, recommendation share, brand identity, and citation provenance in a stratified experiment. With preregistered repetition and paraphrase budgets, it could test whether alignment with buyer intent changes the probability of recommendation.
Abstract
Generative engines such as ChatGPT, Gemini, and Perplexity answer buyer questions directly and name a shortlist of brands inside the answer. Studying how brands enter or fail to enter that shortlist requires demand-side data: what buyers in a category ask, what information they need, and which sources they trust. Existing large persona corpora are built for training-data diversity and carry neither a staged search-intent label nor a preferred-sources field, so they cannot be joined to supply-side recommendation measurements. We built and validated PersonaGen-1M, a corpus of 1,031,732 synthetic buyer personas spanning 511 industry labels and 4 market contexts, carrying 19,416,821 structured behavioral attributes, 5,160,046 of them search queries. Each persona carries a single primary_intent label covering its query set (78.3% informational, 17.4% commercial, 4.3% transactional) and a preferred_sources field naming the source types that buyer would trust. The corpus was built from roughly 40 million raw persona descriptions drawn from four public datasets through GPU-accelerated MinHash LSH plus semantic deduplication, then enriched to a fixed schema. The intent field selects the commercial-evaluation personas whose queries drive recommendation, and the preferred_sources field pairs against citation-provenance data; that join is the primary intended use, and its controlled empirical estimate is future work. Among million-scale persona corpora surveyed in August 2026, one other carries a source-preference attribute, as a six-value media-channel enum; PersonaGen-1M pairs named per-persona source lists with a staged commercial search-intent label and an attached query set. The full corpus is shared on request for non-commercial research; a stratified subset is published openly so the protocol, the schema and the validation can be inspected and reused without asking us.