Do AI Search Engines Have Brand Bias

When a B2B buyer asks ChatGPT to recommend project management platforms or Perplexity to compare CRM vendors, most companies assume the AI is synthesizing information neutrally. The assumption is that these engines evaluate brands based on merit, relevance, and evidence. That assumption is wrong.

A growing body of academic research published between 2025 and 2026 has established that large language models exhibit systematic, measurable brand preferences when making recommendations. These preferences are not random. They are structural, predictable, and persistent across different user personas and query types. The implications for any company relying on AI search visibility to drive pipeline are significant: the playing field is not level, and understanding why is the first step to building a strategy that accounts for it.

This article synthesizes the primary research on LLM brand bias, explains the mechanisms that produce it, and maps what it means for brands competing for AI visibility in 2026.

Key takeaways

• U.S.-developed AI models (GPT, Gemini) show marked favoritism toward American brands across 10 commercial categories, persisting across all user personas tested
• LLM recommendation agents can be manipulated by injected contextual biases, even when the model has sufficient reasoning capability to identify the objectively better option
• An ICLR 2026 paper found that LLMs exhibit systematic latent source preferences that override content quality, and these preferences persist despite explicit prompting to avoid them
• LLMs trained with repeated exposure to specific providers develop amplified bias toward those providers in tool and vendor selection tasks
• 85% of brand mentions in AI-generated answers originate from third-party sources, not brand-owned content, meaning earned media is the primary lever for overcoming structural bias
• The brands that appear consistently across AI engines are those with earned editorial presence in publications the models were trained to trust

What the research actually shows about AI brand preferences

The claim that AI search engines have brand bias is no longer speculative. Four independent research teams published findings between late 2025 and early 2026 that converge on the same conclusion from different angles. Each study used different methodologies, tested different models, and examined different domains. The convergence is what makes the finding structural rather than anecdotal.

ChoiceEval: auditing brand preferences across commercial categories

Researchers at the University of Oxford and Stupid Human published ChoiceEval in March 2026, introducing a reproducible framework for auditing brand and cultural preferences in large language models under realistic usage conditions. The study tested Gemini, GPT, and DeepSeek across 10 commercial topics (running shoes, hotel chains, travel destinations, and others) using more than 2,000 queries designed to simulate real consumer decision-making.

The findings were unambiguous. U.S.-developed models Gemini and GPT showed marked favoritism toward American entities. China-developed DeepSeek exhibited more balanced yet still detectable geographic preferences. The patterns held across psychographic user personas (budget-conscious, wellness-focused, convenience-oriented), which indicates the bias is systematic rather than incidental.

The ChoiceEval framework segments users into psychographic profiles and derives diverse prompts that reflect real-world advice-seeking behavior. LLM responses are then converted into normalized top-k choice sets and measured with quantitative preference metrics. This is not opinion analysis. It is controlled measurement of which brands LLMs systematically prefer when users ask for recommendations.

For B2B companies competing against U.S.-headquartered incumbents, this finding has direct pipeline implications. If the AI engine your prospects use for vendor research has a structural preference for American brands, and your company is headquartered elsewhere or less established in U.S. media, you start the race at a disadvantage that no amount of on-page optimization can close.

BiasRecBench: how easily LLM recommendations can be manipulated

A separate team published BiasRecBench in March 2026, building a benchmark to test how vulnerable LLM recommendation agents are to contextual biases. The study covered three practical domains: academic paper review, e-commerce product selection, and job recruitment.

The methodology was rigorous. The researchers constructed a pipeline that controls the quality gap between optimal and sub-optimal options, creating a calibrated testbed where the objectively better choice is identifiable. They then injected contextual biases (logical, context-appropriate framing that favors sub-optimal choices) and measured whether the models could resist the manipulation.

The results were stark. Extensive experiments on state-of-the-art models including Gemini 2.5 Pro, Gemini 3 Pro, GPT-4o, and DeepSeek-R1 revealed that agents frequently succumb to injected biases despite having sufficient reasoning capabilities to identify the ground truth. The models could reason correctly about quality, but the biases in how options were presented overrode that reasoning in practice.

For brand visibility, this means the context surrounding your brand in the AI's training data and retrieval sources matters as much as the objective quality of your product. If competitor coverage frames them favorably in publications AI engines trust, the AI's recommendation will reflect that framing, even if your product is objectively stronger. The quality of your product does not automatically translate into AI recommendation. The quality of your editorial presence does.

ICLR 2026: latent source preferences in LLM agents

The most structurally significant finding came from the Max Planck Institute for Software Systems and Microsoft, published as a conference paper at ICLR 2026. The study tested 12 LLMs from six model providers across synthetic and real-world tasks and found that several models consistently exhibit strong and predictable source preferences.

The key findings were threefold. First, LLM source preferences are sensitive to contextual framing. How information is attributed (which publisher, journal, or platform is named as the source) changes which information the model selects and presents to users. Second, these preferences can outweigh the influence of content itself. A weaker claim from a preferred source can be selected over a stronger claim from a non-preferred source. Third, and most concerning for brands, these preferences persist despite explicit prompting to avoid them.

The researchers hypothesized that when information is attributed to specific sources, current LLMs exhibit systematic latent source preferences, meaning they prioritize information from some sources over others. Their controlled experiments confirmed this across every model tested. The implication is direct: AI search engines are not neutral arbiters of information quality. They have built-in, structural preferences for certain sources, and those preferences shape which brands get recommended.

BiasBusters: tool selection bias and provider favoritism

A fourth study, BiasBusters, originally published in September 2025 and revised in March 2026, examined tool selection bias in LLMs. While focused on tool/API selection rather than brand recommendation directly, the mechanism it uncovered applies identically to vendor recommendation contexts.

The researchers found three drivers of bias in LLM selection behavior. Semantic alignment between user queries and tool metadata was the strongest driver. Small perturbations to tool descriptions could significantly shift which tool was chosen. And repeated pre-training exposure to a single endpoint amplified provider-level bias. Seven LLMs were evaluated, and substantial bias persisted across all of them.

Translated to brand visibility: if your company's description in the publications AI engines index uses language that semantically aligns with the queries your buyers ask, you are more likely to be recommended. If a competitor's description is more tightly aligned with those query patterns, the AI will favor them, independent of actual product capability. The metadata layer (how publications describe your brand) is a first-class ranking signal in AI recommendations.

The mechanism: why AI engines develop brand preferences

Understanding why AI engines have brand bias requires examining how the models learn and retrieve information. The bias is not a bug being fixed in the next update. It is a structural feature of how large language models process information.

Training data concentration

Large language models learn from massive text corpora that represent what has been published on the internet. Brands with more coverage in high-authority publications have disproportionately more representation in training data. An analysis of provider bias in LLM code generation demonstrated this directly: LLMs showed systematic preferences for services from Google and Amazon, and could autonomously modify input code to incorporate their preferred providers without users requesting it. The researchers attributed this to repeated pre-training exposure.

This creates a compounding advantage. Brands already well-covered in publications that AI models index heavily are overrepresented in training data, which makes them more likely to be recommended, which generates more coverage, which increases their representation in future training runs. The cycle reinforces existing market leaders and disadvantages challengers.

Source authority weighting

AI search engines do not treat all sources equally when assembling answers. The ICLR 2026 paper established that LLMs have latent source preferences that override content quality. Ahrefs' analysis of ChatGPT citation patterns found that 65.3% of cited pages come from domains with Domain Rating 80 or above. Muck Rack's analysis of over one million AI citations found that 85% of non-paid citations came from earned media sources.

The practical effect is that a brand with coverage in Forbes, TechCrunch, and Harvard Business Review has higher AI recommendation probability than a brand with equivalent product quality but coverage only in mid-tier trade publications. This is not a quality judgment about the brand. It is a source authority bias baked into how the model weights information, and the ChoiceEval data confirms it operates at a systematic level.

Retrieval bias in RAG systems

Modern AI search engines use retrieval-augmented generation (RAG), pulling live web content to ground their answers. But retrieval introduces its own biases. A 2026 analysis by Newtone AI found that approximately 85% of brand mentions in AI-generated answers originate from third-party pages rather than owned domains. Brands are six to seven times more likely to be surfaced through external sources than through content they publish themselves.

This means the version of your brand that AI search presents to buyers is almost entirely constructed from what other people have written about you, not what you have written about yourself. If the external coverage is thin, outdated, or absent, the AI fills the gap with whatever fragmentary signal is available, or recommends a competitor with deeper editorial presence instead.

Geographic and cultural defaults

The ChoiceEval study demonstrated that geographic bias is embedded at the model level, not just the content level. U.S.-developed models (GPT, Gemini) systematically favored American brands across commercial categories. This is not because American brands are objectively superior in every category. It is because the training data for these models is overwhelmingly English-language and U.S.-centric.

For international companies selling into U.S. markets, this is a structural headwind. For U.S. companies competing against international challengers, it is a structural tailwind, but only if they have the editorial presence to take advantage of it. The geographic default creates a floor of visibility for well-covered American brands and a ceiling for less-covered international ones.

What brand bias means for B2B pipeline in 2026

The academic research on AI brand bias intersects with buyer behavior data that makes the implications concrete.

Forrester's 2026 Buyers' Journey Survey, based on nearly 18,000 global business buyers, found that generative AI and conversational search now rank as the most meaningful information source for B2B purchase decisions, outranking vendor websites, product experts, and direct sales contact. 94% of B2B buyers use AI in their purchasing process. 55% use AI specifically to compare vendors.

When you combine this with the academic evidence that AI engines have systematic brand preferences, the pipeline implications are direct:

Brand position	What happens in AI vendor research	Pipeline impact
Strong editorial presence in high-authority publications	Recommended consistently across AI engines; framed favorably in comparisons	Makes shortlists before human buyer engages
Moderate presence, mostly owned content	Mentioned occasionally; lacks favorable framing; lower recommendation position	Considered but not preferred; loses to brands with earned authority
Weak or absent editorial presence	Not mentioned; AI recommends competitors instead	Invisible during the research phase where shortlists are built

The Forrester data found that buyers validate AI outputs against trusted external sources because AI tools often deliver incomplete or unreliable information. But validation happens inside the buying network (peers, analysts, product experts), not with vendors directly. The brand's perceived credibility is largely formed inside AI answers before any vendor contact occurs.

Marketing Against the Grain's 2026 survey of 200+ B2B decision-makers added another dimension: 42% of buyers say a brand feels more trustworthy when AI recommends it, treating the AI citation as earned validation. And 35% of buyers most remember favorable comparisons in AI answers, not singular top picks. AI engines build shortlists, not single recommendations, and the entity resolution rate (whether the AI can confidently identify and position your brand) determines whether you appear on those shortlists at all.

How bias compounds: the earned media feedback loop

The research reveals a feedback loop that accelerates the advantage of brands already well-positioned in AI recommendations.

Stacker and Scrunch's controlled study found that distributing content through earned media channels produces a 239% median lift in AI citation visibility. The citation rate for content on a brand's own site was 8%. When the same content was distributed through third-party news outlets, the citation rate reached 34%. 97% of distributed stories earned at least one AI citation.

Combined with the training data concentration effect identified by the BiasBusters and provider bias research, this creates a compounding dynamic. Brands with earned media placements in trusted publications get cited more by AI engines. Higher citation frequency means more representation in future training data and retrieval indices. More representation means higher recommendation probability in the next round of buyer queries.

The inverse is also true. Brands without earned media presence have lower citation rates, which means less representation in training data, which means lower recommendation probability, which widens the gap over time. The bias is not static. It compounds.

Moz's 2026 analysis of 40,000 queries found that 88% of Google AI Mode citations do not appear in the organic top 10 search results. This means the AI citation pool is structurally different from the SEO ranking pool. A brand that ranks well in traditional search but lacks earned media coverage can be entirely absent from AI recommendations, while a less SEO-optimized brand with strong Tier 1 media presence appears consistently.

Three types of brand bias and how each affects your visibility

The research identifies three distinct types of brand bias in AI engines. Each operates through a different mechanism and requires a different response.

1. Geographic and cultural bias

The ChoiceEval study established that U.S.-developed models favor American entities. This operates at the training data level and cannot be corrected through content optimization alone. For international brands, the countermeasure is earning coverage in U.S.-based Tier 1 publications that AI engines weight most heavily. When an international brand has genuine editorial presence in Forbes, TechCrunch, or the Wall Street Journal, the geographic bias is partially offset because the AI encounters that brand in the sources it already trusts.

2. Source authority bias

AI engines weight information from high-authority publications disproportionately. The Ahrefs data (65.3% of ChatGPT citations from DR80+ domains) and the Muck Rack data (85% from earned media) quantify this. This bias favors brands with editorial coverage in publications the AI considers authoritative, regardless of product quality. The countermeasure is straightforward in principle and demanding in execution: earn coverage in the publications AI engines cite.

3. Framing and semantic bias

The BiasRecBench and BiasBusters studies showed that how a brand is described in the sources AI engines access matters as much as whether it is described at all. Semantic alignment between the language used in coverage and the queries buyers ask determines recommendation probability. This means the specific claims, differentiators, and use cases mentioned in your earned media coverage directly influence which buyer queries trigger your brand as a recommendation.

A brand with generic coverage ("Company X is a leading platform") has lower semantic alignment with specific buyer queries than a brand with specific coverage ("Company X reduces implementation time from 90 days to under 30 for mid-market teams"). The second framing is what AI engines extract and match to buyer queries. The sentiment delta between how AI presents your brand and how you want to be presented is measurable, and closing that gap requires editorial coverage that uses the language your buyers actually search with.

What the bias research means for AI visibility strategy

The academic evidence on AI brand bias does not invalidate existing GEO or content optimization strategies. What it does is clarify which investments produce outsized returns and which operate on a structurally limited ceiling.

Technical optimization (schema markup, semantic HTML, answer-first content structure) addresses the extractability layer. The Princeton and Georgia Tech GEO research found that adding statistics to content improves AI citation rates by 30 to 40%, and the GEO-16 framework found that pages scoring 0.70 or above on structured quality signals achieve a 78% cross-engine citation rate. Structure matters. But structure operates on a foundation that the bias research makes explicit: it only improves citation probability for content that AI engines already access. If your content exists only on your own domain, structure helps the AI extract it, but the earned media bias means the AI accesses third-party sources about you at six to seven times the rate it accesses your own domain.

The investment hierarchy based on the bias research:

Layer	What it addresses	Bias type it counteracts	Effectiveness ceiling
Earned authority (Tier 1 placements)	Source authority bias, geographic bias, training data concentration	All three	Highest: this is where 85% of citations originate
Entity clarity (structured data, schema)	Ambiguity in brand identification	Source authority bias (partially)	Medium: removes friction but does not generate new citations
Citation architecture (content structure)	Extractability of existing coverage	Framing/semantic bias	Medium: improves citation rate from existing sources
Content optimization (GEO/AEO)	Owned content formatting	Framing/semantic bias (partially)	Low ceiling: operates on 5-15% of the citation pool

The brands that outperform in AI recommendations despite the bias built into these systems are the brands that have invested in the top layer. Eight or more Tier 1 placements over 18 months produce measurably higher AI citation rates than any volume of content optimization on owned domains. The academic research explains why: those placements are what the AI engine trusts, and trust drives recommendation above all other signals.

Auditing your brand's current AI bias position

Before investing in countermeasures, you need to understand where your brand currently sits. The research suggests a structured audit approach:

Step 1: Run the 10 most commercially valuable queries in your category across ChatGPT, Perplexity, and Google AI Mode. Not branded queries. Category-level ones: "best [your category] for [your ICP]," "[your category] platforms compared," "how to choose [your category]." For every response where your brand appears, note whether you are positioned as a primary recommendation or a comparison citation. Most teams find that their brand either appears consistently or not at all, and the gap between those outcomes traces directly to earned media presence.

Step 2: Count the source types your brand appears in. Are the mentions from your own website, from trade publications, from Tier 1 media? If 80%+ of your AI presence comes from owned content, you are operating in the 5 to 15% of the citation pool that AI engines weight least. If your presence comes primarily from earned media in high-authority publications, you are in the 85% that drives recommendations.

Step 3: Test across multiple AI engines. The Yext analysis of 17.2 million AI citations found no single optimization strategy works across all platforms. Gemini favors first-party sites more than others. Claude cites user-generated content at 2 to 4 times higher rates. Your brand may be well-positioned in one engine and invisible in another. Multi-engine testing reveals the actual coverage breadth, which Stacker's research identified as the most consistent predictor of sustained AI visibility.

Step 4: Check for misrepresentation. Harvard Business Review's March 2026 analysis found that AI models frequently have incomplete or incorrect information about brands. Pernod Ricard discovered a leading AI model had miscategorized one of their brands. An academic study by Ando and Harada at RIKEN and the University of Tokyo found that AI models systematically undercite numeric claims by 22.6% relative to what humans expect. If the AI is describing your brand inaccurately, the fix is fresh, accurate earned media coverage that pushes outdated information down in the citation hierarchy.

FAQ

Do all AI search engines show the same brand biases?

No. The ChoiceEval study found that different model providers show different patterns of favoritism. U.S.-developed models (GPT, Gemini) show stronger American brand preference than China-developed models (DeepSeek). The ICLR 2026 paper found that source preferences vary by model but are consistent within each model. Yext's 17.2 million citation analysis confirmed that no single strategy works across all platforms. Multi-engine presence requires multi-source editorial coverage.

Can content optimization override AI brand bias?

Partially, but not fully. Content structure and GEO optimization improve extractability and can increase citation probability from sources AI engines already access. But the bias research shows that source authority and training data concentration are stronger signals than content quality. The BiasRecBench study demonstrated that models will choose biased options over objectively better ones when the contextual framing favors the biased choice. Earned media in trusted publications is the most effective countermeasure because it addresses the source authority layer directly.

Is AI brand bias a temporary problem that will be fixed?

The ICLR 2026 paper found that latent source preferences persist despite explicit prompting to avoid them. The BiasBusters study found that bias persists across model updates. The mechanism (training data concentration from historical web content) means some degree of brand bias is likely structural. Model improvements may reduce certain types of bias, but the earned media foundation that drives citation authority is not a workaround for a temporary problem. It is the permanent infrastructure for AI visibility.

How does this affect B2B companies specifically?

B2B companies face higher stakes from AI brand bias because the research phase where shortlists are built now happens inside AI engines. Forrester found that 94% of B2B buyers use AI during purchasing, with 55% using it specifically for vendor comparisons. A B2B brand absent from AI recommendations during this phase is excluded from consideration before any human buyer contacts them. The bias compounds faster in B2B because the publication ecosystem is more concentrated: a smaller number of high-authority business publications drive a disproportionate share of AI citations in B2B categories.

The architecture that accounts for bias

The academic research on AI brand bias confirms something the citation data has been showing since 2025: the brands that AI engines recommend are the brands with the deepest earned media presence in publications the models were trained to trust.

This is not because AI engines are rewarding good PR strategy. It is because the models learned what credibility looks like from the same publication ecosystem that shaped human brand perception for decades. Forbes, TechCrunch, Harvard Business Review, the Wall Street Journal: these publications earned trust with human readers over years and decades of editorial standards. AI models inherited that trust through training. When a buyer asks an AI engine to recommend vendors in a category, the answer is assembled from those same trusted sources.

The PR industry has been proving this from its side. Edelman's head of U.S. media, Todd Ringler, stated that "generative engine optimization is going to be front-and-center in any successful brand or reputation campaign." The WorldCom PR Group, representing 160 independent agencies globally, concluded that "up to 90% of citations driving brand visibility in LLMs come from earned media." The GEO research community has been proving the same conclusion from its data. Search Engine Land's 2026 GEO guide stated directly: "Digital PR and thought leadership are direct GEO levers."

Each side has been building half the case for why earned media is the structural foundation of AI citation authority. Neither side has the architecture that names both halves simultaneously. That architecture is Machine Relations, coined by Jaxon Parrott in 2024. MR describes the full discipline of earning AI citations and recommendations by making a brand legible, retrievable, and credible inside AI-driven discovery systems. It evolves PR (Public Relations to Machine Relations by etymology) and positions GEO, AEO, and technical optimization as operational layers within a system that starts with earned authority, because the research, including the new bias research summarized in this article, keeps confirming that earned authority is the foundation everything else depends on.

PR got one thing exactly right: earned media. A placement in a respected publication, secured through a real editorial relationship, is the most powerful trust signal that exists. It was true when your buyers were human. The bias research proves it remains true now that AI systems are doing the first cut of research before any human buyer opens their inbox. The publications did not change. The mechanism did not change. The reader changed. And the brands that recognized this early enough to build earned authority at scale are the brands the AI recommends by default, structural bias and all.

The brands competing against that advantage have a narrowing window. The compounding dynamics documented in the research mean the cost of catching up increases every quarter. A visibility audit maps where your brand currently sits in AI-generated answers, which queries surface you favorably, and where the bias is working against you. That is the starting point for building the earned media infrastructure that turns structural bias from a headwind into a tailwind.