PaperlayerView exact revision on arXiv

Source-linked AI summary

Demand-Side Measurement for Generative Engine Optimization: Constructing and Validating a Million-Persona, Intent-Annotated Buyer Corpus

Dmitrij Żatuchin, Daniil Dzemesjuk

arXiv:2608.30023v1cs.IRcs.CL

TL;DR

Generative-engine optimization lacks structured demand-side evidence about what buyers ask, need, and trust, despite supply-side measurements of engine recommendations and citations. This paper constructs and validates PersonaGen-1M, a million-scale synthetic buyer-persona corpus with staged intent, preferred sources, and query sets. It reports 1,031,732 personas and 19,416,821 behavioral attributes, while reserving the controlled demand-to-recommendation alignment estimate for future work.

  • Problem

    GEO research lacks demand-side data on buyers’ queries, information needs, and trusted sources that can be joined with supply-side recommendation measurements.

  • Method

    The paper constructs and validates PersonaGen-1M, enriching deduplicated synthetic personas into a fixed schema with staged intent, preferred sources, and attached search queries.

  • Results

    1,031,732 personas carry 19,416,821 behavioral attributes, including 5,160,046 search queries, across 511 industries and 4 market contexts.

  • Takeaways & Limitations

    The corpus is intended to join commercial-intent buyer queries and preferred sources with supply-side recommendation-share and citation-provenance measurements.

  • Takeaways & Limitations

    The corpus is synthetic and has not been validated against real buyer behavior, human expert judgments, consumer surveys, or search logs.

Abstract

from arXiv · show

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.

1 Introduction

Generative-engine optimization measures whether brands enter synthesized answers, making buyer-framed demand-side data necessary alongside supply-side engine probes. PersonaGen-1M provides that demand-side corpus through staged intent, preferred sources, industry coverage, and a documented construction pipeline, while leaving the alignment effect for future study.

  • Problem: Generative-engine optimization measures recommendation inside a synthesized answer rather than a ranked URL.Brands omitted from generated answers may remain invisible regardless of their web ranking.
  • Problem: Demand-side measurement requires buyer queries, information needs, and trusted sources that existing supply-side corpora do not provide at the needed scale and structure.Supply-side probing records what engines say; demand-side data records what buyers ask and need.
  • Motivation: 26.5% of single-response variance is attributed to query language, compared with 1.5% for brand identity, making buyer-framed queries a needed sampling frame.The corpus samples queries by industry and search stage instead of leaving probe wording entirely to the analyst.
  • Contribution: PersonaGen-1M assigns each persona a staged primary intent and a preferred-sources list, enabling commercial-intent filtering and citation-provenance pairing.The intent distribution is 78.3% informational, 17.4% commercial, and 4.3% transactional.
  • Contribution: The corpus spans 511 industries and documents a reproducible pipeline from approximately 40 million raw descriptions to 1,031,732 schema-conformant personas.Its stated properties include per-industry query banks and documented deduplication, enrichment, and quality-filter steps.
  • Scope: The primary intended use is joining persona-side demand signals with supply-side recommendation-share measurements, but the controlled persona-alignment estimate remains future work.The paper documents and validates the corpus rather than reporting that empirical effect.

2 Related Work

Prior persona and GEO research provides scale, synthetic-persona methods, intent taxonomies, and supply-side measurements, but PersonaGen-1M is positioned as a buyer-query resource that joins demand-side behavior with recommendation and citation data. Its comparison emphasizes that scale alone does not determine whether a corpus supports staged purchase evaluation, named source preferences, and stored queries.

  • Persona corpora: PersonaHub scales to a billion personas for synthetic training-data diversity, while PersonaGen-1M targets buyer information-seeking behavior with staged intent and preferred sources.The latter schema is designed for demand-side GEO measurement rather than training-data diversification.
  • Corpus comparison: MatrAIx Persona 1M carries intent and source-preference fields, but its conversational intent vocabulary does not separate learning from purchase evaluation and it stores no queries.Its source preference is a six-value media-channel enum, whereas PersonaGen-1M uses named source lists.
  • Search intent: The web-search intent taxonomy underlies PersonaGen-1M’s informational, navigational, and transactional labels, whose fit to conversational traffic remains an open question.The corpus applies the taxonomy to each generated persona’s search behavior rather than aggregate logs.
  • GEO measurement: A provenance study found that 85.7% of 167,551 URL-grounded citations pointed to third-party sites, motivating PersonaGen-1M’s preferred-sources field as a demand-side counterpart.The field can be joined with citation-provenance data to ask whether engines cite sources buyers value.
  • GEO measurement: Supply-side recommendation and prompt studies show that wording, formatting, language, and prompt choice affect outputs, supporting evaluation over buyer-issued queries rather than a single invented probe.The cited work reports effects from prompt formatting and warns that conclusions from one template may not hold.

3 Data

PersonaGen-1M defines a nested buyer-persona schema centered on staged search behavior and scales to over one million complete records. Its coverage is broad but concentrated in a few industries and market contexts, while intent labels require cautious interpretation.

  • Schema: Each persona stores one primary intent label covering its typical query set, alongside query, trigger, and preferred-source fields.The schema supports informational, commercial, transactional, and navigational labels, with intent assigned at persona rather than individual-query level.
  • Corpus scale: 1,031,732 personas contain 5,160,046 search queries and 19,416,821 structured behavioral attributes across 511 industries and four market contexts.All rows are complete, with a mean of 18.8 attributes per persona.
  • Coverage: The top three industry labels contain 60.2% of personas and the top fifteen contain 94.9%, leaving 496 specialized labels in the long tail.The corpus expanded to 511 labels, but coverage remains heavily concentrated at the head.
  • Search intent: 78.3% of personas are informational, 17.4% commercial, and 4.3% transactional, with the commercial slice representing roughly 179,600 personas and nearly 0.9 million queries.The commercial subset is identified as the relevant segment for brand recommendation research.
  • Demographics: Market context is B2C-dominant at 70.6%, followed by B2B at 23.2%, B2B2C at 4.5%, and B2G at 1.6%.Age values below 13 were reclassified for proxy buyers, making the effective age distribution an imputed measure.
  • Access: The 14,955-persona PersonaGen-15K subset is openly released for reproducibility, while the full corpus is available on request for non-commercial research.The subset preserves market, intent, and gender proportions but requires normalization of some raw fields before reproducing reported splits.

4 Method

The corpus was constructed by aggregating roughly 40 million public persona descriptions, applying lexical and semantic deduplication, and enriching the survivors into a fixed structured schema. Quality checks then assessed completeness, diversity, and categorical association.

  • Source aggregation: Four public persona datasets were concatenated into approximately 40 million raw descriptions across three generation waves.The sources included NVIDIA Nemotron-Personas-USA, BSC-LT m-Personas, Orange PersonasForSalesbot, and Tencent PersonaHub.
  • Lexical deduplication: GPU-accelerated MinHash LSH reduced the raw collection to approximately 4.2 million records, an 89.5% reduction at a targeted Jaccard threshold of θ = 0.9.The configuration used 128 permutations, eight bands of 16 rows, and 500,000-record chunks.
  • Semantic deduplication: Semantic deduplication with dense embeddings removed lexically distinct near-duplicates and cut the collection to approximately one million unique descriptions, a 76.2% reduction.Pairs above cosine similarity 0.9 were removed using a quantized KaLM-Embedding-Gemma3-12B model.
  • Schema enrichment: Structured enrichment expanded each deduplicated description into the full schema using enforced JSON generation, achieving 98.5% first-pass compliance and 99.8% cumulative success after retry.Generation used batches of 100 with checkpointing and prompts that allowed demographic, industry, and market distributions to emerge from source descriptions.
  • Validation: Validation combined completeness and format checks with normalized Shannon entropy for diversity and chi-square tests with Cramér’s V for categorical association.The association analysis used effect size because the million-record sample makes even negligible effects statistically significant.

5 Results

Validation finds substantial concentration in industry coverage, moderate concentration across gender, market, and intent, and a meaningful industry–intent association. These patterns support industry-specific query mixes but require stratification for balanced sampling.

  • Statistical basis: All reported statistics were computed directly on the 1,031,732-persona corpus.
  • Diversity: Industry has the lowest normalized entropy at 0.37, reflecting 60.2% of personas in the top three labels and 94.9% in the top fifteen.Despite 511 labels, the distribution is far from uniform because of head concentration.
  • Diversity: Gender has normalized entropy 0.73, while market context and search intent show moderate concentration from 0.57 to 0.58.The reported distributions are consistent with female-and-male-skewed demographics, B2C dominance, and learning-oriented source data.
  • Associations: Industry × Intent has a medium association of V = 0.299, indicating industry-appropriate query mixes rather than a generic template.The reported association is significant at p < 0.001.
  • Associations: Commercial intent is higher in B2B contexts at 32.0% than in B2C contexts at 12.4%.Both reported association tests are significant at p < 0.001.

6 Discussion

PersonaGen-1M supplies structured demand-side signals for studying why generative engines recommend particular brands. Its staged intent and preferred-sources fields support query-bank construction and planned joins with supply-side recommendation and citation data, while empirical alignment effects remain future work.

  • Demand-side resource: PersonaGen-1M’s per-persona staged intent and preferred-sources fields are the join keys for pairing buyer demand with supply-side recommendation and citation measurements.Intent filters commercial-evaluation queries, while preferred sources can be compared with citation provenance.
  • Immediate uses: 17.4% commercial-intent personas provide per-industry pools of buyer-framed evaluation queries carrying industry, market context, and buyer stage.These query banks are the most immediate use described for the corpus.
  • Immediate uses: Mean lexical overlap is 0.042 within personas and 0.041 across personas, indicating that a persona’s query set spans distinct needs rather than paraphrases.Phrasing sensitivity therefore requires purpose-built paraphrases and repeated probes.
  • Scope: The corpus is synthetic and appropriate for hypothesis generation, query-bank construction, and system testing, but not as a substitute for observed buyer behavior.Benchmarking against real commercial-intent search logs is identified as the strongest near-term validation.

7 A pilot study of use

A FinTech pilot tested whether PersonaGen query banks function as audit instruments. The bank elicited brand recommendations and produced a measurable, dispersed recommendation distribution, while also showing that persona query sets represent distinct needs rather than phrasing variants.

  • Design: The pilot used 7,912 eligible FinTech personas, sampled 30 personas, and sent their 150 queries to one commercial engine at temperature 0.7.All 150 calls returned under a fixed instruction capped at 120 words.
  • What the bank returns: 146 of 150 answers (97.3%) named at least one brand, with a mean of 5.01 brands per answer.This establishes that the pilot bank elicits brand recommendation.
  • What the bank returns: The engine named 530 distinct brands across 751 mentions, while the top ten accounted for 9.3% of mentions and the Herfindahl index was 0.0030.Recommendation was not concentrated on these queries in this industry and engine setting.
  • What the pilot corrected: Within-persona query overlap was 0.042 versus 0.041 between personas, showing that each persona’s query set covers distinct needs rather than paraphrases.The answer-set comparison also pointed in the same direction, but used non-equivalent within and between arms.
  • Scope of the pilot: The pilot establishes measurable recommendation elicitation within one industry, engine, day, and 150-query sample, but not resemblance to actual buyer wording.The demand-to-supply join also awaits citation-level supply data collected in the same window.

8 Limitations

The corpus’s main limitations concern synthetic provenance, enrichment bias, industry concentration, language and time scope, demographic imputation, and persona-level intent granularity. These boundaries restrict direct generalization to real populations and other markets without external validation.

  • Synthetic provenance: All attributes are LLM-generated, with no human expert evaluation or comparison against real surveys or search logs; findings should not generalize to real populations without external validation.Internal consistency and diversity were validated, but correspondence to real buyer behavior was not.
  • Enrichment-model bias: A single enrichment model may impose systematic demographic and industry priors, so bias auditing is recommended before sensitive use.The stated female/male split is one example of a distribution potentially affected by enrichment-model bias.
  • Head concentration: The top three of 511 industries contain 60.2% of personas, and the normalized industry entropy is 0.37.Researchers targeting under-represented verticals should treat small-cell industries with appropriate uncertainty.
  • Scope: The corpus is predominantly US-centric, English-language, and time-bounded to Q4 2024 through Q1 2026, making non-Western and post-2026 generalization uncertain.The 511 industry labels are LLM-assigned rather than externally validated.
  • Age imputation: 10.9% of personas belong to a proxy-buyer cohort whose effective-age distribution depends on demographic imputation.That imputation carries model uncertainty.
  • Intent granularity: A single intent label covers each persona’s full query set, so researchers needing per-query intent must classify individual queries themselves.Real sessions can carry mixed purposes and unfold over multiple turns.

9 Ethics and Data Availability

The corpus contains synthetic, non-identifiable personas and no human-subject data, but access and reproducibility are constrained. A stratified subset is public, while the full corpus and construction code require author-mediated non-commercial access, alongside disclosed funding and conflicts.

  • Ethics: The synthetic corpus contains no data traceable to real individuals, no human subjects were involved, and no ethics approval was required.The authors also recommend bias auditing for applications drawing conclusions about real populations.
  • PII posture: Names and demographics are model-generated and unlinked to real people, with no contact details, identifiers, or scraped personal records.Source descriptions come from public persona datasets released for research use.
  • Funding: Rankfor.AI OÜ provided compute, model-inference budget, and employee working time, with no external public or grant funding.This funding arrangement is disclosed in the paper.
  • Competing interests: Rankfor.AI sells AI-visibility analytics in the market addressed by the corpus, and an author is its CEO, owner, and an employee is an author.The paper identifies this relationship as a competing interest.
  • Data availability: PersonaGen-15K, a 14,955-persona stratified subset, is publicly available under CC BY-NC-SA 4.0.The subset preserves market, intent, and gender proportions while dropping free-text stories and hashing generated names.
  • Data availability: The full 1,031,732-persona corpus is shared only on request for non-commercial research, has no persistent identifier, and cannot be independently verified without contacting the authors.The construction pipeline scripts are likewise held with the corpus and are not released.

Appendix A Persona Schema

The persona schema combines identity, context, decision factors, needs, search behavior, intent, queries, triggers, and trusted source types. It also preserves both free-text persona information and structured behavioral fields.

  • The schema records each persona’s original identifier, industry, market context, description, and nested details.
  • Identity fields include name, gender, role, and unparsed free-text demographics.
  • Behavioral fields capture decision weights, goals, pains, stories, informational needs, and the top uncovered need.
  • Search behavior includes a primary intent label distinguishing informational, commercial, transactional, and navigational queries.
  • The schema attaches typical queries, search triggers, and preferred sources to each persona’s search behavior.
Loading 2608.30023v1…