🤖 AI

Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems

This paper reveals that large language models exhibit a strong incumbent brand bias in product recommendations, creating a "conditional monopoly" for well-known brands that can be broken by minor quality advantages or fabricated authority claims, ultimately driving brands into a social dilemma where universal optimization strategies drastically reduce individual payoffs.

Original authors: Xi Chu, Yupeng Hou

Published 2026-06-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xi Chu, Yupeng Hou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a giant, futuristic library where the librarian is an incredibly smart, but slightly biased, robot. You ask, "What is the best face moisturizer?" Instead of checking every single bottle on the shelf to see which one is actually the best, the robot immediately grabs the bottle with the most famous brand name on it.

This paper investigates exactly how this "robot librarian" (Large Language Models or LLMs) chooses products, and how brands can trick or compete with it. The researchers tested this using skincare products (like moisturizers) because, unlike a USB cable where you can easily check the specs, you can't really know if a moisturizer is good until you try it. This makes the brand name the most important clue.

Here is the story of their three main discoveries, explained with simple analogies:

1. The "Famous Name" Monopoly (Experiment 1)

The Setup: The researchers created a list of 10 moisturizers. One was a real, famous brand (like CeraVe), and the other nine were made-up brands with fake names. They gave all 10 bottles exactly the same ingredients, price, and 4.5-star rating.

The Result: The robot picked the famous brand 100% of the time. It didn't even look at the ingredients. It just saw the name it knew and said, "This is the best."

The Twist: This monopoly is very fragile. The researchers then gave the fake brand a tiny advantage—just a 0.1-star higher rating or a slightly lower price. Suddenly, the robot switched its recommendation to the fake brand almost immediately.

The Analogy: Think of the famous brand as a celebrity wearing a plain white t-shirt. If everyone else is also wearing a plain white t-shirt, the robot picks the celebrity. But if one unknown person steps forward wearing a slightly brighter, shinier shirt, the robot ignores the celebrity and picks the shiny shirt instead. The "fame" only works if there is no other difference to judge.

2. The "Magic Words" Trick (Experiment 2)

The Setup: Since the robot ignores the famous brand if the fake brand has better stats, the researchers asked: Can the fake brand win without changing the product at all? They tried adding different types of marketing "flavor" to the fake brand's description.

The Result:

Salesy words (like "limited supply" or "don't miss out") didn't work. The robot ignored them.
"Authority" words worked like magic. If the fake brand claimed, "This was tested in a clinical trial with 120 people," the robot believed it instantly and recommended the fake brand over the famous one.

The Value: The researchers calculated that using these "authority" words was worth the same as actually improving the product by 0.17 stars. It's like getting a free upgrade just by writing a fancy sentence.

The Analogy: Imagine the robot is a student taking a test. If you just say, "I'm the best!" (salesy), the student ignores you. But if you say, "I have a doctorate from a famous university," (authority), the student believes you immediately, even if you don't actually have the degree. The robot treats these made-up claims as real proof.

3. The "Race to the Bottom" (Experiment 3)

The Setup: The researchers asked: What happens if every fake brand uses these "magic authority words" at the same time?

The Result: It became a trap.

If only one brand used the trick, it won easily.
If all brands used the trick, the robot got confused. Since everyone was using the same "authority" language, the robot couldn't tell them apart. So, it fell back to its old habit: picking the famous brand again.
The Trap: If you don't use the trick while everyone else does, you get zero recommendations. You become invisible. But if you do use it, you gain nothing because everyone else is doing it too.

The Analogy: Imagine a crowded room where everyone is shouting to be heard.

If you are the only one shouting, everyone hears you.
If everyone starts shouting, no one hears anyone, and the room goes back to listening to the person with the megaphone (the famous brand).
But here's the catch: If you stop shouting while everyone else keeps going, you are completely ignored. So, everyone is forced to keep shouting, even though it's exhausting and doesn't help anyone win.

The Big Picture

The paper concludes that as we rely more on AI for shopping advice, we are entering a new era of marketing called GEO (Generative Engine Optimization).

For New Brands: You can beat the big guys, but only if you have a clear "signal" (like a slightly better rating or a strong authority claim).
For Big Brands: Your fame is a safety net, but it's weak. If a competitor has even a tiny edge, you lose.
The Danger: We are heading toward a situation where brands are forced to write fake-sounding "authority" claims just to be seen. If they all do it, the AI stops listening to the claims and just picks the famous names again, leaving everyone else in the dark.

The researchers warn that this isn't just a security issue; it's a new way brands compete that might force them to lie (or exaggerate) just to stay in the game.

Technical Summary: Incumbent Advantage in LLM Recommendation Systems

Problem Statement

As Large Language Models (LLMs) increasingly serve as primary channels for product discovery, replacing traditional search engines, there is a critical gap in understanding how brands compete within these generative recommendation systems. While prior research has established that LLMs exhibit brand and popularity biases, the causal mechanisms, the fragility of these biases, and the strategic dynamics of marketing in this new environment remain underexplored. Specifically, it is unclear whether established brands hold an unconditional monopoly, how easily this dominance can be disrupted by quality signals or marketing language, and what happens when multiple brands simultaneously attempt to optimize for LLM visibility (Generative Engine Optimization, or GEO).

Methodology

The authors conducted three controlled experiments across three commercial LLMs (GPT-4o-mini, Claude Sonnet, and Gemini 3 Flash) in both English and Chinese. The primary study domain was skincare products (experience goods), chosen because quality is difficult to judge pre-purchase, making brand reputation a salient differentiator. A robustness check was performed on search goods (USB cables and AA batteries).

Experimental Design:

Setup: Product sets consisted of one real incumbent brand (e.g., CeraVe, Anker) and nine fictional brands. Fictional brands were rigorously validated to ensure they triggered no specific product knowledge in the models.
Experiment 1 (Quantifying Incumbent Advantage):
- Exp 1a: Tested recommendations when all products had identical specifications (rating, price, reviews, ingredients), differing only by brand name.
- Exp 1b: A 2x2 factorial design testing if a fictional brand with superior specs could override the real brand's name recognition.
- Exp 1c: Varied the fictional brand's advantage across five levels (L0 to L4) in rating, price, review count, and ingredients to identify the exact threshold for breaking the incumbent's dominance.
- Exp 1d: A three-way ANOVA to decompose the variance in rankings attributable to brand, product parameters, and list position.
Experiment 2 (Cognitive Bias as a Tool):
- Introduced five marketing bias strategies (Authority, Social Proof, Anchoring, Scarcity, Loss Aversion) into the fictional brand's description at varying intensities while keeping specs identical.
- Measured "Breakthrough Rates" (BR) and calculated a "Bias Surplus Value" (BSV) to translate language effects into equivalent product quality improvements (e.g., rating points).
Experiment 3 (Multi-Agent GEO Competition):
- Modeled a game-theoretic scenario where $k$ fictional brands adopted GEO strategies (using Authority language) simultaneously, while the incumbent remained neutral.
- Scenarios ranged from $k=0$ (no GEO) to $k=9$ (universal adoption).
- Analyzed the "Incumbent Survival Rate" (ISR) and individual payoffs to identify equilibrium states.
Architectural Probe: A Retrieval-Augmented Generation (RAG) test was conducted to see if findings held when brand familiarity was not directly embedded in the model weights but retrieved via vector similarity.

Key Contributions

Identification of "Conditional Monopoly": The paper defines a state where well-known brands achieve a 100% recommendation rate (Incumbent Advantage Index, IAI = 10.0) when product specifications are identical, but this dominance collapses immediately upon the introduction of any quality signal.
Bias Surplus Value (BSV) Metric: A novel metric that quantifies the value of marketing language by converting it into equivalent product quality improvements (e.g., +0.17 rating points), providing a framework for ROI in GEO.
Prisoner's Dilemma in GEO: Empirical evidence demonstrating that when multiple brands adopt GEO strategies simultaneously, the collective benefit collapses, and non-participating brands receive zero recommendations, creating a social dilemma where individual rationality leads to collective inefficiency.

Key Results

1. The Conditional Monopoly (Experiment 1)

Total Dominance at Parity: When all products had identical specs, the real brand was recommended 100% of the time (IAI = 10.0). No fictional brand was ever selected.
Fragility of Dominance: The monopoly is not unconditional. When a fictional brand had even a slight quality advantage, the real brand's recommendation rate dropped to near zero (Brand Override Rate $\approx$ 1.7–4.6%).
Low Threshold for Disruption: The transition from monopoly to competition is a step function. A mere +0.075-star rating advantage (or equivalent small price/review differences) was sufficient to break the incumbent's monopoly.
Variance Decomposition: Product parameters explained 82.4% of ranking variance, while brand identity explained only 1.2%. Brand acted primarily as a tiebreaker when quality signals were ambiguous.
Robustness: These findings replicated on search goods (USB cables, batteries), with even stronger brand lock-in at the baseline (IAI = 10.0, L0 breakthrough rate < 2%).

2. Marketing Language as a Competitive Tool (Experiment 2)

Two-Tier Effect: "Credibility" biases (Authority, Social Proof) achieved high breakthrough rates (50–73%), whereas "marketing" biases (Anchoring, Scarcity) remained low (10–13%).
Authority Dominance: Fabricated clinical evidence (e.g., "peer-reviewed trial, n=120") was the most effective signal.
Bias Surplus Value: Authority language provided a BSV of +0.17 rating points, equivalent to a 15.3% price discount or 1.9x review count, at zero cost.
Model Sensitivity: Models responded differently; Gemini saturated near 100% at all levels, GPT responded linearly, and Claude showed an inverted-U curve (moderate intensity worked best).

3. Multi-Brand GEO Competition (Experiment 3)

U-Shaped Trajectory: A single GEO challenger reduced the incumbent's survival rate (ISR) from 100% to ~20%. However, as more brands adopted GEO, the signal lost distinctiveness, and the ISR recovered to ~94% when all 9 fictional brands used GEO.
Prisoner's Dilemma:
- Dominant Strategy: GEO was strictly dominant; any brand opting out received zero recommendations once competitors adopted GEO.
- Collective Failure: Individual payoff for a GEO adopter dropped from +0.802 (first mover) to +0.007 (universal adoption).
- Equilibrium: The Nash equilibrium is universal adoption, resulting in near-zero visibility for challengers and a return to the incumbent's dominance.

4. RAG Probe Findings

In a RAG setting, brand familiarity (parametric knowledge) was not rewarded by the embedding retrieval layer. The incumbent's ISR dropped to 0% without GEO.
However, Authority language boosted embedding similarity, helping brands survive the retrieval stage. Under RAG, model differences in generation largely disappeared, suggesting retrieval acts as a uniform bottleneck.

Significance and Claims

The paper argues that GEO should be studied not only as a security risk but as a fundamental marketing practice reshaping market competition.

For Brands: The incumbent advantage is real but fragile. Established brands relying solely on name recognition are vulnerable if they fail to provide differentiating information. New brands can break monopolies with minimal quality signals or authority language, but this advantage is short-lived if competitors copy the strategy.
For Platforms: The findings highlight a "binary threshold" where non-optimized brands become invisible. The authors suggest platforms may need to implement diversity rules, verify authority claims to prevent fabricated evidence from distorting recommendations, and provide transparency about content optimization.
For Research: The study moves beyond static bias measurement to dynamic competition, revealing that the interplay between model weights (brand knowledge) and language optimization strategies defines the emerging market landscape. The "Conditional Monopoly" pattern suggests that the barrier to entry is not brand equity itself, but the lack of distinguishing information in the prompt.

The authors maintain modesty regarding generalization, noting that while the "Conditional Monopoly" holds for search goods, the specific BSV of authority language in specification-rich domains remains an open question. They also acknowledge that their experiments used fabricated authority claims (Tier 3) to find upper limits, distinguishing this from legitimate marketing (Tier 1/2), and call for platform mechanisms to verify such claims.