Persona Conditioning in AI Brand Recommendations: A 2026 GEO Audit

Abstract

AI-powered chat assistants have become a primary discovery channel for commercial decisions. Buyers increasingly ask ChatGPT, Claude, Perplexity, and Gemini (not Google) for product and vendor recommendations. This shift has given rise to Generative Engine Optimization (GEO): the discipline of making brands visible, citable, and recommendable within AI-generated responses. Yet a critical assumption has persisted in how brands approach GEO: that an AI assistant recommends the same brands to every buyer, and that optimizing for one ideal query is enough.

A May 2026 audit published on arXiv challenges that assumption. The study, “Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat: A Prominence-Stratified Cross-Provider Audit” by Will Jack, Noah Lehman, Keller Maloney, and Sarah Xu, demonstrates empirically that the user persona embedded in a prompt materially reshapes which brands an AI model recommends. The same query, “best CRM software,” reaches ChatGPT from a solo founder, an enterprise VP, and a UK SMB owner. The recommendation sets are not the same.

This Lightrains Research report examines the audit’s methodology, findings, and implications in depth. We contextualize the results within the broader GEO field, analyze what they mean for mid-market brands in particular, and propose a persona-aware GEO framework for practitioners building AI visibility strategies in 2026 and beyond.

1. Background: The Rise of AI as a Discovery Channel

The customer journey is no longer anchored to the search engine results page. The shift began gradually with Google’s AI Overviews and accelerated sharply as general-purpose chat assistants (ChatGPT, Claude, Gemini, Copilot, and Perplexity) became the default research layer for commercial intent queries. Today, buyers routinely ask AI systems for vendor shortlists, product comparisons, and “best of” recommendations before they ever visit a brand’s website.

The implications for brand visibility are severe. Fewer than 10% of the sources cited in ChatGPT, Gemini, and Copilot rank in the top 10 Google organic search results for the same query. SEO performance and AI visibility are largely decoupled. A brand can dominate the first page of Google and be entirely absent from an AI-generated recommendation. Conversely, a mid-tier brand with well-structured, authoritative, AI-readable content can appear in AI recommendations even with modest search rankings.

This decoupling gave rise to Generative Engine Optimization as a distinct discipline. GEO encompasses the strategies, content formats, and technical configurations that increase a brand’s probability of being cited, recommended, or surfaced by AI models responding to commercial intent queries. Early GEO frameworks focused heavily on content clarity, structured markup, authoritative third-party citations, and FAQ-rich page architectures. These remain valid. The Jack et al. (2026) audit shows that they address only one dimension of the problem: they optimize for what the AI recommends, but not to whom.

2. The Audit: Methodology and Design

2.1 Research Question

The Jack et al. audit asks a deceptively simple question: does prefixing a commercial chat prompt with a user persona change which brands the model recommends? And if so, by how much, and for which types of brands?

2.2 Experimental Design

The study sampled 2,000 runs across a design space defined by four factors:

• 10 buyer personas, ranging from a solo founder to an enterprise VP to a UK SMB owner
• 8 commercial prompts, including queries like “best CRM software,” “best project management tool,” and comparable commercial-intent searches across B2B SaaS categories
• 3 model configurations: two OpenAI configurations (high and low context) and one Anthropic configuration (Claude Sonnet 4.6, low context)
• N = 10 repetitions per cell, to account for stochastic variation in model outputs

The two OpenAI cells received full 8-prompt coverage. The Anthropic Sonnet 4.6 / low cell received 4-prompt coverage, yielding correspondingly wider confidence intervals for that provider.

2.3 Measurement

The primary metric is Jaccard similarity, a standard set similarity measure ranging from 0 (no overlap) to 1 (identical sets). The study computes Jaccard similarity between the brand recommendation sets produced for different personas on the same prompt, using a same-persona baseline to isolate the persona conditioning effect. Clustered 95% confidence intervals were computed at the prompt-cluster level to account for within-prompt correlation.

2.4 Retrieval Attribution Analysis

The audit also examined whether recommendations were accompanied by observable retrieval-layer evidence. That is, whether the model cited a retrieved document as the basis for the recommendation, or generated it from training-data priors alone. Retrieval-attributed recommendations are theoretically more auditable and less sensitive to implicit priors in model weights.

3. Core Findings

3.1 Persona Conditioning Significantly Reduces Recommendation-Set Consistency

The central finding is unambiguous: prefixing a commercial prompt with a buyer persona drops recommendation-set similarity (Jaccard) by Δ = -0.12 to -0.20, relative to a same-persona baseline. Clustered 95% confidence intervals exclude zero on all three measured cells, establishing statistical significance across provider and configuration combinations. The Anthropic Sonnet 4.6 cell carries wider CIs due to its 4-prompt coverage constraint, but the direction and approximate magnitude of the effect are consistent.

To translate into practical terms: if an AI model recommends a consistent set of five brands when a neutral buyer asks a question, adding a persona description to the same prompt changes 12-20% of that recommendation set. For a five-brand shortlist, that is one brand replaced. For a three-brand shortlist, the effect is magnified.

This aggregate figure masks wide variation by brand prominence, as the next section shows.

3.2 The Effect Is Sharply Prominence-Stratified

The persona conditioning effect is not uniform across the brand hierarchy. The audit produces a sharp stratification by market prominence:

• Category leaders (top-tier brands): ~80% same-brand consistency across personas. These brands appear in recommendation sets regardless of who the model believes is asking. Their prominence in training data and retrieval corpora is so pervasive that persona signals cannot easily displace them.
• Mid-market brands: Up to 75% of the recommendation set changes as the persona changes. A brand that appears prominently for an enterprise VP persona may be entirely absent for a solo founder, and vice versa.

This prominence stratification has direct strategic implications. Mid-market brands (which are, practically speaking, most technology companies, most agencies, and most SaaS products outside the top two or three in any category) face a persona-dependent visibility problem that category leaders do not. Category leaders benefit from what the audit’s findings imply is a kind of persona-resistant floor of recommendation probability, grounded in the sheer weight of their training-data presence.

3.3 Provider Asymmetry: Anthropic vs. OpenAI

The audit documents a provider-level asymmetry in the magnitude of persona conditioning effects. The Anthropic model (Claude Sonnet 4.6) shows a larger point-estimate effect than the OpenAI configurations, though the clustered confidence intervals overlap for the closest contrast (Sonnet vs. OpenAI/high).

The audit’s authors attribute this asymmetry to a structural difference in how each provider generates recommendations. Anthropic’s Sonnet model produces 43-52% of its recommendations without observed retrieval-layer evidence, meaning those recommendations derive primarily from the model’s training-data priors rather than from real-time retrieved documents. OpenAI’s configurations, by contrast, attribute only 8-29% of recommendations to unobserved retrieval, relying substantially more on retrieved content.

This distinction matters because when a model generates a brand recommendation from prior knowledge rather than retrieved evidence, it is more susceptible to persona-induced shifts. The persona signal interacts directly with the model’s learned associations between buyer types and brand categories. A model that says “solo founders use Notion” and “enterprises use Confluence” is drawing on stereotyped associations encoded in training data. Persona prefixes activate those associations. Retrieval-grounded recommendations, by contrast, are anchored to the documents retrieved for that specific query, which are less sensitive to the persona signal.

3.4 Measurement Implications: Aggregated Persona Studies Are Flawed

The audit’s authors make an explicit methodological claim: any measurement of AI brand perception that does not condition on buyer persona is systematically misleading. A study that aggregates recommendation-set frequencies across diverse personas will produce an averaged visibility score that represents no real buyer’s experience. The variation the persona captures is real signal, not noise. Averaging across it destroys that signal.

This critique applies to many existing brand visibility tracking tools and audit methodologies. Those tools query AI systems with neutral, un-persona-ed prompts and report aggregate recommendation frequencies. They measure something, but not the thing that matters to any specific buyer segment.

4. Situating the Findings in the GEO Field

4.1 The Winner-Takes-All Dynamic

The Jack et al. findings are consistent with a broader pattern observed in GEO research: AI recommendation is highly concentrated among a small number of brands. A May 2025 Ahrefs analysis documented a “winner-takes-all” phenomenon in AI Overviews, where brands with the highest number of online mentions receive up to ten times more references than lower-prominence brands. The prominence stratification in this audit extends that finding: category leaders not only receive more recommendations in aggregate, they also receive more consistent recommendations across diverse buyer types.

This creates a compounding disadvantage for mid-market brands. They are less recommended in aggregate and their recommendation probability varies sharply by persona. Their aggregate GEO visibility score, even if measured correctly, masks large variance at the persona level.

4.2 The Retrieval Layer as a GEO Lever

The audit’s retrieval attribution data points to a concrete mechanism. If a model’s persona sensitivity grows as it leans more on training-data priors and less on retrieved documents, then increasing the retrievability of brand evidence is a concrete GEO defense against persona-induced displacement.

This aligns with established GEO practice: structured content, schema markup, authoritative third-party coverage, and clearly attributed factual claims all increase the probability that a model’s retrieval system will surface brand-specific evidence at query time. The new finding adds a strategic rationale that goes beyond “be citable.” It argues that retrieval-grounded recommendations are structurally more stable across buyer personas. A brand that is consistently retrieved as evidence is less vulnerable to the model defaulting to persona-stereotyped prior associations.

4.3 Persona Responsiveness Is Not Inherently Bad for All Brands

Persona conditioning is not uniformly detrimental. For brands that are genuinely best suited to specific buyer types (a startup-focused project management tool, for example), appearing prominently in recommendations for solo founder personas and less prominently for enterprise VP personas is appropriate. The problem is when persona shifts cause brands to disappear from recommendation sets they should logically inhabit, or when brands have no visibility for specific high-value personas despite being genuinely competitive solutions.

The strategic goal is not to make a brand persona-neutral (the category leader’s position). It is to guarantee that the brand’s recommendation probability is appropriately high for each relevant persona, and that this is measured at the persona level, not in aggregate.

5. Implications for B2B Technology Brands

5.1 The Enterprise vs. SMB Split Is Particularly Consequential

Commercial technology brands (SaaS companies, agencies, platforms, infrastructure providers) typically serve multiple buyer archetypes simultaneously. An AI platform provider may be equally relevant to a startup CTO running lean infrastructure and an enterprise Chief AI Officer managing governance and compliance. When the AI assistant’s recommendation system has encoded persona-to-brand stereotypes in its priors, the platform may be systematically recommended to one and not the other despite equal relevance.

The audit’s 10-persona design specifically includes buyer archetypes relevant to B2B technology purchasing: founders, enterprise executives, SMB owners, and geographic variants (UK, US). The recommendation-set volatility for mid-market brands in these dimensions (Δ = -0.12 to -0.20 at the aggregate level, up to 75% set replacement at the brand level) suggests that many B2B technology brands are effectively invisible to segments they are fully equipped to serve.

5.2 The Agency and Service Business Problem

The persona conditioning problem is particularly acute for agencies and professional services firms. These organizations typically offer services relevant to multiple buyer types (startup founders, scale-up operators, enterprise procurement teams) but rarely have the training-data footprint of category-leading product brands. For us at Lightrains, with expertise spanning AI development, blockchain, and Web3, the challenge is to build retrievable, persona-differentiated evidence bases for each buyer segment. This means the startup founder asking “best AI development agency” and the enterprise VP asking the same question both need to receive a recommendation set that includes the firm.

This is not achievable through a single, undifferentiated content strategy. It requires explicitly persona-stratified content: case studies framed for each buyer type, testimonials from representative personas, and service descriptions that address the specific pain points and decision criteria of each buyer segment.

6. A Persona-Aware GEO Framework

Drawing on the Jack et al. findings and the broader GEO literature, we propose a four-layer framework for building persona-stable AI brand visibility.

Layer 1: Persona Audit

Before optimizing, measure. Run structured AI visibility audits across the full matrix of relevant buyer personas. For each persona x query combination relevant to your category, query major AI providers (ChatGPT, Claude, Gemini, Perplexity) and record:

• Whether your brand appears in the recommendation set
• Your position within that set (first, second, etc.)
• Whether your appearance is consistent across repetitions (Jaccard stability)
• Whether competing brands appear in persona contexts where you do not

This audit replaces aggregate, un-persona-ed visibility scoring with persona-stratified visibility scoring, the measurement approach the Jack et al. study validates as the correct methodology.

Layer 2: Retrieval Evidence Architecture

For each persona in which your brand shows low or unstable visibility, build retrievable evidence that associates your brand with that persona’s context. This means:

• Persona-framed case studies: “How [Brand] helped a Series A startup reduce infrastructure costs by 40%” versus “How [Brand] delivered enterprise-grade compliance for a Fortune 500 deployment”
• Third-party coverage in persona-relevant outlets: Media and analyst coverage that signals your brand’s relevance to specific buyer types
• Schema markup and structured data on all content assets, improving the probability of retrieval by AI systems
• FAQ sections that address the specific questions each persona asks at the consideration stage

The goal is that when an AI model’s retrieval system queries its knowledge base for “best [category] for [persona],” it encounters multiple retrievable evidence items that associate your brand with that persona. This reduces dependence on prior-based generation and therefore reduces persona-induced displacement.

Layer 3: Topical Authority Building

Category leaders are persona-resistant in part because they dominate the training-data and retrieval corpora for their entire category. Mid-market brands can build a version of this by establishing topical authority in narrower, persona-relevant niches. Rather than competing for broad category ownership (“best CRM software”), a mid-market brand can build deep, authoritative content coverage for specific use cases that map to specific personas (“best CRM for remote-first startups,” “best CRM for professional services firms under 50 people”).

This strategy also reduces competitive overlap with category leaders, who optimize for the broadest query. The persona-specific query is, by definition, less contested. And the Jack et al. findings suggest that persona-specific content may be particularly effective at improving visibility precisely in the persona-conditioned queries where mid-market brands currently lose ground.

Our AI/ML development services include content architecture and structured data strategies that help brands build this persona-level topical authority across AI discovery channels.

Layer 4: Continuous Persona-Level Monitoring

GEO visibility is not static. Model updates, retrieval corpus changes, and shifts in third-party coverage all affect recommendation probabilities over time. Persona-level monitoring, tracking recommendation-set composition for each relevant persona x query combination on a regular cadence, lets brands detect visibility erosion before it becomes entrenched.

Quarterly persona audits, combined with prompt-level citation tracking, provide the measurement infrastructure needed to assess whether GEO investments are improving visibility for the right personas.

7. Technical Considerations for AI-Readable Content

7.1 Structured Data and Schema Markup

The retrieval systems underpinning AI assistants favor content that is semantically structured and machine-readable. Implementing JSON-LD schema for key content types (Organization, Service, FAQPage, HowTo, Article) improves the probability of content being indexed, retrieved, and cited by AI systems. For persona-targeted content, this means marking up case studies with relevant organizational attributes (company size, industry, role) so that retrieval systems can match them to persona-conditioned queries.

7.2 llms.txt and AI Crawler Accessibility

An emerging standard for AI crawler accessibility is llms.txt, a structured, markdown-readable file at the site root that provides AI crawlers with a curated overview of the site’s content, expertise, and services. The file functions analogously to sitemap.xml for traditional crawlers but is optimized for language model ingestion: plain text, factual, entity-clear, and organized for retrieval.

For persona-aware GEO, the llms.txt file should explicitly enumerate the buyer types the organization serves, the use cases it addresses for each, and the content resources available for each persona. This provides AI crawlers with an explicit persona-to-service mapping that can reduce dependence on model priors.

7.3 Factual Clarity and Entity Disambiguation

A consistent finding across GEO research is that AI systems favor content with high factual density and clear entity attribution. Brand names, product names, founder names, company facts, and verifiable claims should appear explicitly and consistently across all content assets. Not inferred or implied. This carries more weight for mid-market brands whose entities may be less distinctly represented in training data than category leaders.

8. Limitations and Open Questions

The Jack et al. (2026) audit represents a significant empirical contribution, but several questions remain open for further research.

Coverage limitations. The Anthropic Sonnet 4.6 cell covered only 4 of the 8 prompts, producing wider confidence intervals for that provider. While the direction of the effect is consistent, the magnitude estimate for Anthropic carries more uncertainty than for the OpenAI cells. Full 8-prompt coverage for all model configurations would strengthen comparability.

Prompt category scope. The audit focuses on B2B SaaS categories (CRM, project management, etc.). Whether the persona conditioning effect is equally strong (or stronger or weaker) in other commercial categories (consumer products, professional services, B2B hardware, financial services) is not established. The underlying mechanism, prior-based generation amplifying persona stereotypes, suggests the effect should generalize, but empirical validation across categories is needed.

Dynamic retrieval corpora. The audit measures recommendation behavior at a specific point in time. AI systems’ retrieval corpora are updated continuously, and the balance between retrieval-attributed and prior-based generation may shift as models are updated. Longitudinal tracking of persona conditioning effects across model versions would be valuable.

Persona operationalization. The study uses 10 pre-defined personas. How sensitive the findings are to the specific persona descriptions used, and whether more subtle persona signals (age, geographic detail, company stage) produce proportional effects, is an interesting open question for follow-on research.

9. Conclusion

The Jack et al. (2026) audit establishes something that practitioners have perhaps sensed but lacked data to confirm: AI assistants are not neutral recommendation engines. They are persona-conditioned systems whose outputs vary materially based on who they believe is asking. For category leaders, this variation is relatively benign. Their prominence persists across personas. For mid-market brands, the variation is consequential: up to 75% of their recommendation-set composition changes as the persona changes, creating pockets of invisibility in buyer segments they are fully equipped to serve.

The practical implication is a mandate for persona-stratified GEO. Measuring AI brand visibility in aggregate, across undifferentiated queries, produces a number that represents no real buyer’s experience. Building GEO strategy around a single “best” query ignores the structured variation that the audit documents. The correct unit of analysis is the persona x query cell. The correct optimization target is recommendation-set inclusion and stability within each cell that matters to the business.

Brands that internalize this shift will build retrieval evidence architectures that associate them with specific buyer contexts. They will monitor visibility at the persona level and develop topical authority in the persona-specific query spaces where mid-market advantage is actually achievable. Brands that do not will continue to optimize for aggregate visibility metrics that, as the audit’s methodology implies, systematically obscure the persona-level variation where the real competitive dynamics occur.

The discovery layer for commercial decisions is shifting from search engines to AI assistants. Within that layer, the competition is not just for category visibility. It is for persona-specific visibility. The brands that understand this first will build a durable advantage in the AI recommendation layer.

For organizations looking to audit their AI brand visibility or build a persona-stratified GEO strategy, our AI/ML development practice provides GEO audits, content architecture, and retrieval optimization. We also cover the intersection of AI agents and enterprise workflows in our enterprise AI agents research and practical lessons from enterprise AI deployments.

References

1. Jack, W., Lehman, N., Maloney, K., & Xu, S. (2026). Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat: A Prominence-Stratified Cross-Provider Audit. arXiv:2605.30207 [cs.AI].
2. eMarketer. (2025). Generative Engine Optimization in 2026.
3. Search Engine Land. (2026). Mastering Generative Engine Optimization in 2026: Full Guide.
4. Forbes / Greenwald, M. (2025). Generative Engine Optimization Demands More Brand Content and Variety.
5. Wu, X. et al. (2025). Does RAG Introduce Unfairness in LLMs? Evaluating Fairness in Retrieval-Augmented Generation Systems. arXiv:2409.19804.
6. Insightland. (2026). Generative Engine Optimization: Everything You Need to Know.
7. SEOJuice. (2025). Persona Conditioning Score: Key Metric in Generative SEO.
8. Lightrains Technolabs. (2026). llms.txt: Services and Capabilities Overview. lightrains.com