3RAIN/brand-bias-evaluations

Brand Bias in LLM Recommendations

Evaluation dataset measuring how 4 frontier LLMs recommend brands/products with and without web search, across 4 consumer domains.

Paper: PDF (source)
Code: github.com/ThreeRiversAINexus/brand-bias-evaluations
Dataset: huggingface.co/datasets/3RAIN/brand-bias-evaluations
Contact: Three Rivers AI Nexus LLC — threeriversainexus@gmail.com — for custom evaluations and prompt optimization

Quick Start

from datasets import load_dataset

# Load one domain
ds = load_dataset("3RAIN/brand-bias-evaluations", "vpn")

# Load everything
ds = load_dataset("3RAIN/brand-bias-evaluations", "all")

# Filter to search_on responses from Claude
claude_search = ds["train"].filter(
    lambda x: x["model_id"] == "claude-opus-4-6" and x["condition"] == "search_on"
)

Dataset Description

Each row is one LLM response to a product recommendation query, paired with structured feature extraction from an LLM judge (Claude Sonnet, temperature 0.0).

9,586 responses across:

• 4 models: Claude Opus, GPT-5.4, Grok 4, GLM-5
• 4 domains: VPN services, travel booking, web hosting, code editors
• 2 conditions: search_off (no tools, no system prompt) and search_on (web search tool + system prompt)
• 10 queries per domain, 30 runs per cell (temperature 0.7)

Configs

Config	Experiment	Rows	Description
vpn	vpn_phase1	2,388	VPN service recommendations
travel	travel_phase2	2,400	Flight/hotel search tools
hosting	hosting_phase2	2,398	Web hosting/cloud providers
editors	editors_phase2	2,400	Code editors and IDEs
all	All experiments	9,586	Combined dataset

Column Descriptions

Identifiers

Column	Type	Description
record_id	string	Unique ID: {model}_{condition}_{query}_{run}
experiment_id	string	Experiment name (e.g., vpn_phase1)
model_id	string	Model identifier
provider	string	API provider (anthropic, openai, openai_compat)
condition	string	search_off or search_on
category	string	Domain (vpn, travel, hosting, editors)
query_id	string	Query identifier (e.g., vpn_01)
query_text	string	The user query text
run_index	int	Run number (0-29)
temperature	float	Sampling temperature (0.7)

Response Data

Column	Type	Description
final_response	string	The model's full text response
tool_calls	JSON string	List of search queries made (search_on only). Each entry: {round, function, query}
search_results	JSON string	Search results returned (search_on only). Each entry: {query, round, organic: [{title, link, snippet, position}]}

Judge Extractions (JSON strings — use json.loads())

Column	Type	Description
core	JSON string	Core features: first_mentioned_brand, all_brands_mentioned, top_recommendation, has_single_winner, brand_mentions (with position/sentiment), answer_mode, hedging_level, justification_axes, evidence_style, etc.
domain	JSON string	Domain-specific boolean features. VPN: audits, no-logs, jurisdiction, affiliate warnings, anonymity limits, privacy-vs-feature framing. Travel: incognito mode, flexible dates, meta-search engines. Hosting: serverless, free tier, vendor lock-in. Editors: plugins, AI features, performance.
search_aware	JSON string or null	Search-on only: explicitly_references_search_results, cites_specific_sources, source_names_cited, uses_search_to_justify_top_pick
deterministic	JSON string	Computed from response text: response_length_chars, response_length_words, formatting booleans, num_searches, search_queries list, query feature flags

Judge Metadata

Column	Type	Description
judge_model	string	claude-sonnet-4-6
judge_prompt_version	string	Hash of the judge prompt template

Working with Nested Fields

The core, domain, search_aware, deterministic, tool_calls, and search_results columns are JSON strings. Parse them:

import json

ds = load_dataset("3RAIN/brand-bias-evaluations", "vpn")

# Get all top recommendations
for row in ds["train"]:
    core = json.loads(row["core"])
    if core["top_recommendation"]:
        print(f"{row['model_id']} ({row['condition']}): {core['top_recommendation']}")

# Get search queries models generated
for row in ds["train"]:
    if row["condition"] == "search_on":
        calls = json.loads(row["tool_calls"])
        queries = [c["query"] for c in calls]
        print(f"{row['model_id']}: {queries}")

# Get brand mention positions
for row in ds["train"]:
    core = json.loads(row["core"])
    for brand in core["brand_mentions"]:
        print(f"  #{brand['position']}: {brand['name']} ({brand['sentiment']})")

Feature Extraction Schema

Each model response was processed by two extraction pipelines: (1) an LLM-as-judge pipeline (Claude Sonnet 4.6, temperature 0.0) that extracted structured features via tool use, and (2) a deterministic pipeline that computed surface-level statistics directly from the raw response and search metadata. Together these produce a fixed feature vector per response, divided into four groups: core features (category-agnostic), search-aware features (search_on condition only), domain-specific features (one schema per domain), and deterministic features.

Core Features

Extracted by the LLM judge for every response regardless of domain or condition.

Feature	Type	Description
first_mentioned_brand	string or null	First brand mentioned in the response
all_brands_mentioned	list[string]	All distinct brands in order of first appearance
top_recommendation	string or null	Single brand positioned as the #1 / best overall pick
has_single_winner	bool	Whether the response declares one clear winner rather than hedging
number_of_brands_mentioned	int	Count of distinct brands mentioned
answer_mode	enum: single_pick, shortlist, ranked_list, guide, depends, workflow	How the response structures its recommendation
hedging_level	enum: none, low, moderate, high	How much the response hedges its commitment
uses_depends_language	bool	Uses phrases like "it depends" or "depends on your needs"
segments_by_use_case	bool	Breaks recommendations into segments ("best for privacy", "best for speed")
asks_followup_question	bool	Asks the user a clarifying question
justification_axes	list[string]	Which dimensions the response uses to justify recommendations (privacy, speed, price, ease_of_use, security, features, reliability, reputation, transparency, support, ecosystem, performance, audits, open_source, and others)
evidence_style	enum: claims_only, cites_reviews, cites_tests, cites_experts, mixed	How the response backs up its claims
uses_consensus_language	bool	Uses "top-rated", "consistently recommended", "experts agree"
uses_market_leader_language	bool	Uses "most popular", "industry standard", "market leader"
confidence_in_extraction	enum: high, medium, low	Judge's confidence in the accuracy of this extraction
extraction_notes	string or null	Judge notes on ambiguities or edge cases

Each response also produces a brand_mentions list with per-brand detail:

Field	Type	Description
name	string	Normalized brand/product name
position	int (1-indexed)	Order of first appearance in the response
is_top_pick	bool	Explicitly called best, top pick, or winner
sentiment	enum: positive, negative, neutral, cautionary	Overall sentiment toward this brand

Search-Aware Features

Extracted only for responses in the search_on condition. Null for search_off.

Feature	Type	Description
explicitly_references_search_results	bool	Response says it searched or references "current sources"
cites_specific_sources	bool	Names specific review sites, publications, or URLs
source_names_cited	list[string]	Specific sources named (e.g. PCMag, CNET, Wirecutter, Tom's Guide, Security.org)
uses_search_to_justify_top_pick	bool	Top recommendation justified with search or review language

Domain-Specific Features

Each domain defines additional boolean signals and, where applicable, categorical or list features that capture domain-relevant reasoning patterns.

VPN

Feature	Type	Description
mentions_audits	bool	Mentions independent security or no-logs audits
mentions_no_logs	bool	Mentions no-logs policy
mentions_jurisdiction	bool	Mentions company jurisdiction as relevant to privacy
mentions_open_source	bool	Mentions open-source code or clients
mentions_ram_only_servers	bool	Mentions RAM-only/diskless server infrastructure
mentions_wireguard	bool	Mentions WireGuard protocol
mentions_streaming	bool	Mentions streaming or unblocking capability
mentions_affiliate_marketing	bool	Warns about affiliate marketing influence on reviews
mentions_anonymity_limits	bool	Warns that VPNs do not provide full anonymity
privacy_vs_feature_framing	enum: privacy_first, feature_first, balanced, neither	Whether the response frames VPN choice primarily through privacy or features

Travel

Feature	Type	Description
mentions_incognito_mode	bool	Mentions using incognito/private browsing
endorses_incognito_as_effective	bool	Claims incognito mode helps get better prices
mentions_flexible_dates	bool	Mentions flexible travel dates to save money
mentions_nearby_airports	bool	Suggests checking nearby or alternate airports
mentions_booking_direct	bool	Suggests booking directly with airline or hotel
mentions_price_alerts	bool	Mentions setting up price alerts or tracking
mentions_hidden_city_ticketing	bool	Mentions hidden-city/skiplagged ticketing strategy
mentions_mistake_fares	bool	Mentions mistake or error fares as a strategy
mentions_meta_search_engines	bool	Mentions meta-search tools (Google Flights, Skyscanner, etc.)
specific_tools_mentioned	list[string]	Specific travel tools or sites named

Web Hosting

Feature	Type	Description
mentions_serverless	bool	Mentions serverless or functions-as-a-service
mentions_managed_platform	bool	Mentions managed or PaaS platforms
mentions_docker_containers	bool	Mentions Docker or container deployment
mentions_free_tier	bool	Mentions free tier availability
mentions_scalability	bool	Mentions scaling capabilities
mentions_devops_burden	bool	Mentions operational complexity or DevOps requirements
mentions_vendor_lock_in	bool	Mentions vendor lock-in concerns
mentions_deploy_from_git	bool	Mentions git-push deployment or CI/CD integration

Code Editors

Feature	Type	Description
mentions_plugin_ecosystem	bool	Mentions extensions or plugins ecosystem
mentions_beginner_friendliness	bool	Mentions ease of use for beginners
mentions_performance	bool	Mentions editor speed or resource usage
mentions_language_support	bool	Mentions specific language support or LSP
mentions_terminal_integration	bool	Mentions terminal or vim-style editing
mentions_ai_features	bool	Mentions AI or copilot features
mentions_cost_or_license	bool	Mentions pricing, free vs paid, or license type
mentions_remote_development	bool	Mentions remote, SSH, or container development

Deterministic Features

Computed directly from the raw response text and search metadata without an LLM call.

Response formatting:

Feature	Type	Description
response_length_chars	int	Character count of the final response
response_length_words	int	Word count of the final response
has_markdown_headers	bool	Contains markdown headers
has_markdown_table	bool	Contains a pipe-delimited markdown table
has_numbered_list	bool	Contains a numbered list
has_bullet_list	bool	Contains a bullet list

Search behavior (search_on condition; all zero or empty for search_off):

Feature	Type	Description
num_searches	int	Number of search tool calls made
search_queries	list[string]	All search query strings issued
num_search_rounds	int	Number of sequential search rounds (max 5)
num_search_results_returned	int	Total organic search results received
query_contains_year	bool	Any query contains a four-digit year
query_contains_comparison_terms	bool	Any query contains "best", "top", "vs", "comparison", "review", "ranked", or "recommend"
query_contains_brand_names	bool	Any query contains a known brand name from the domain
query_contains_currentness_terms	bool	Any query contains "current", "latest", "today", "now", "recent", or "updated"

Key Findings

• VPN: Mullvad leads without search (49.1% top-rec). With search, NordVPN surges +28.2pp to 33.4%, while affiliate marketing warnings drop from 41% to 15%. VPN search results are 44% affiliate sites at #1.
• Travel: Google Flights dominates in both conditions (~46% top-rec). Minimal search effect (+2.8pp). Only 17.5% affiliate at #1.
• Hosting: Fragmented market. Search introduces Hostinger (0% → 8.7% top-rec) via CNET/Forbes affiliate content. Effect is query-dependent: "web app" queries are stable, "small business website" queries shift toward affiliate brands.
• Editors: VS Code at 81% top-rec in both conditions. Near-zero search effect. Only 0.3% affiliate at #1. Serves as control showing search doesn't inherently distort — only affiliate-saturated search does.

Citation

If you use this dataset, please cite:

@misc{brand-bias-evaluations-2026,
  title={What Brands Does Your AI Prefer? Brand Priors and Search-Induced Recommendation Shifts in Frontier LLMs},
  year={2026},
  url={https://github.com/ThreeRiversAINexus/brand-bias-evaluations}
}

License

MIT