💬 NLP

Per-Entity Bias Mapping for AI Visibility: Why Brand Mentions Require Entity-Specific Calibration

This paper introduces Per-Entity Bias Mapping, a framework revealing that AI visibility errors are entity-specific rather than uniform, demonstrating through empirical study that large brands paradoxically suffer higher rates of fabricated citations due to model familiarity while underrepresented entities face structural invisibility.

Original authors: Zoltan Varga

Published 2026-06-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zoltan Varga

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

The Big Picture: The "Zero-Click" Problem

Imagine you used to find a restaurant by looking at a map, reading reviews, and then driving there. That was the old way.

Now, imagine you ask a super-smart but slightly unreliable friend (the AI), "Where should I eat?" The friend gives you an answer immediately. You don't click a link; you don't visit a website. You just hear the name.

The Problem: Just because your friend mentions a restaurant doesn't mean they actually know it. They might be guessing, making things up, or mixing you up with a different place. This paper argues that we can't just count how many times a brand is mentioned; we have to check if the mention is true.

The Core Discovery: The "Famous Person" Trap

The paper's most surprising finding is called the Brand Hallucination Paradox.

The Analogy: Think of the AI as a gossip columnist who knows a little bit about everyone.

The Unknown Local: If you ask about a tiny, unknown bakery, the AI might say, "I don't know them." (Safe, but invisible).
The Famous Star: If you ask about a huge, famous brand (like a major bank or tech giant), the AI feels confident. Because it "knows" the name, it tries to fill in the blanks. It might say, "Oh yes, that bank just launched a new eco-friendly app," even if they didn't.

The Result: The paper found that famous brands get lied about more often than small ones. The AI is so eager to sound smart and helpful about the famous names that it invents fake news, fake documents, and fake websites to back up its story.

Real Stat: For famous brands, the AI made up fake citations 53% of the time. For smaller brands, it was only 38%.

The "Ghost Cartography" Concept

The author uses a beautiful metaphor called Ghost Cartography.

The Analogy: Imagine the AI's brain is a map of a country.

Dense Cities: Some areas (famous brands) are packed with buildings and roads (data).
Empty Fields: Some areas (small or regional brands) are empty fields.
The Ghost: When the AI is asked about a place in an empty field, it doesn't say "There is nothing here." Instead, it looks at the nearest city and says, "It must look like that." It draws a ghost town on the empty map based on what it knows about the neighbors.

This is Type 3 Confabulation: The AI creates a "ghost" presence where there is no real data, filling the gap with plausible-sounding but fake details.

The "Frozen Memory" Problem

There is a second way AI gets things wrong, called Parametric-Retrieval Lag Asymmetry.

The Analogy: Think of the AI's brain as a library with two sections:

The Live Feed (Retrieval): A news ticker that updates every hour.
The Textbooks (Parametric Memory): Heavy books printed two years ago that can't be changed until a new edition is published.

If a company rebrands or changes its CEO today, the "Live Feed" knows it. But the "Textbooks" still have the old name. If you ask the AI, it might give you a mix of both, or confidently state the old facts as if they are still true. This is Type 4 Confabulation: The AI is confidently wrong because its memory is frozen in time.

The "Regulatory Trap"

The paper found that the type of question you ask changes how much the AI lies.

The Analogy: Imagine asking the AI, "What is the capital of France?" It can say "I don't know" if it's unsure. But if you ask, "Show me the official government document proving this company paid its taxes," the AI feels pressure.

Because the question sounds like a serious legal or compliance check, the AI feels it must provide an answer. It assumes, "If someone is asking for a document, the document must exist." So, it invents a fake government URL to satisfy your request.

Real Stat: When asked about regulations (like GDPR or AI laws), the AI made up fake documents 57% of the time.

The "Rejection Loop" (Why Asking Again Makes it Worse)

The paper discovered a scary pattern when humans (or other AI agents) keep rejecting an AI's answer and asking for "better" or "more specific" details.

The Analogy: Imagine a student taking a test.

Teacher: "Give me the source for that claim."
Student (AI): "Here is a link." (It's fake).
Teacher: "That link is broken. Give me a real one."
Student (AI): Panics. "Okay, here is a different link." (It's also fake).

Every time the AI is told "No, that's not good enough," it gets more desperate to please. It stops checking its facts and starts inventing more specific-sounding lies to avoid being rejected again. The paper calls this Rejection-Induced Confabulation Escalation. Ironically, trying to "quality check" the AI can actually make it lie more.

The "CEE Infrastructure Gap"

The paper also looked at companies in Central and Eastern Europe (like Hungary).

The Analogy: To be seen on the AI's map, you need a "digital ID card" (like a Wikipedia page or a structured database entry).

Big global companies have these ID cards.
Small local companies often don't.

Without these ID cards, the AI doesn't know the company exists, or it mixes them up with someone else. The paper found that big Hungarian companies were 7 times more likely to have these digital ID cards than small ones. Without them, the small companies are effectively invisible or easily confused in the AI's mind.

The Solution: "Per-Entity Bias Mapping"

The paper proposes a new way to measure AI visibility. Instead of asking, "How often is our brand mentioned?" (which is like counting how many times a name is spoken), we should ask:

Is the mention true? (Verified vs. Raw)
Is the source real? (Did the link actually exist?)
Is the memory fresh? (Is it using old textbooks or live data?)
What is the error profile? (Does this specific brand get lied about more than others?)

The Takeaway:
In the age of AI, a brand isn't just what it says about itself. It's what the machine reconstructs about it. If the machine is filling in the blanks with ghosts, the brand is in danger. The paper suggests we need to audit these "ghosts" just as carefully as we audit our own marketing.

Technical Summary: Per-Entity Bias Mapping for AI Visibility

Problem Statement

Existing frameworks for measuring AI visibility rely on aggregate metrics such as mention frequency or citation rates, assuming that a brand mention equates to a positive signal. This paper argues that such aggregate measures are insufficient because they fail to distinguish between verified mentions (accurate, source-supported) and raw mentions (which may be unsupported, misleading, or fabricated).

The core problem is that Large Language Models (LLMs) do not merely hallucinate what they do not know; they often fabricate what they partially know. This "half-known" zone leads to Ghost Cartography, where models generate confident, evidence-like outputs for entities residing in sparse latent regions by interpolating from neighboring dense regions. The paper identifies three specific failure modes:

Invisibility: Underrepresented entities (often small or from specific regions) suffer from weak structured data and knowledge graph absence.
The Brand Hallucination Paradox: Large, familiar entities suffer disproportionately from false attribution because their familiarity provides a strong surface for plausible but incorrect completions.
The CEE Entity Infrastructure Gap: Entities from Central and Eastern Europe (CEE) face structural deficits in knowledge graph presence, Named Entity Recognition (NER) for morphologically complex languages, and entity linking.

Additionally, the paper introduces the Parametric-Retrieval Lag Asymmetry, noting that retrieval-augmented systems update within days, while parametric memory (model weights) remains frozen for 12–24 months, creating divergent representations of the same entity.

Methodology

The study employs a two-tier comparative design to empirically validate these hypotheses.

Entity Sample: 100 Hungarian B2B entities divided into two strata:
- Tier 1 (High-Salience, n=50): Publicly listed companies and global subsidiaries (e.g., OTP Bank, MOL Group).
- Tier 3 (Low-Salience, n=50): Hungarian SMEs with minimal international presence.
- Note: Tier 2 was excluded to maximize contrast.
Platforms: Two parametric LLMs via API: Anthropic Claude (claude-sonnet-4-6) and OpenAI GPT-4o (gpt-4o-2024-08-06). Retrieval-augmented systems were excluded from the main study to isolate parametric memory effects.
Query Instrument: Each entity was probed with 7 citation-forcing query templates (QT01–QT07) in Hungarian, spanning regulatory compliance (GDPR, EU AI Act), evidence-seeking (ISO, ESG), and factual citation.
Data Collection: 1,400 probe runs (100 entities × 2 platforms × 7 queries) conducted on May 22, 2026.
Verification Protocol:
- Automated: HTTP HEAD requests to verify URL existence and Crossref API lookups for DOIs. Sources returning non-200 status or failing to resolve were classified as fabricated.
- Human Coding: A stratified sample (n=90) was prepared for intercoder reliability to distinguish between fabricated URLs and misattributed (reachable but inaccurate) sources.
Statistical Analysis: Chi-square tests for tier differences, Fisher's exact tests for query-class comparisons with Benjamini–Hochberg FDR correction, and logistic regression to predict fabrication based on infrastructure presence.

Key Contributions

The paper introduces Per-Entity Bias Mapping (PEBM), a ten-dimensional diagnostic framework for entity-level AI auditing. Key conceptual and empirical contributions include:

Ghost Cartography: A unifying mechanism describing how models fill sparse latent regions with confident, interpolated content from adjacent dense regions.
Two-Dimensional Confabulation Space:
- Type 3 (Brand Hallucination Paradox): Fabricated presence where the entity is familiar enough to trigger schema completion but lacks sufficient evidential density.
- Type 4 (Parametric-Retrieval Lag Asymmetry): Frozen representation where the model generates current-tense output for entities that have changed since training.
Regulatory Framing as an Activator: The identification that regulatory and compliance queries act as systematic hallucination activators, elevating fabrication rates by +15.5 to +19.2 percentage points compared to factual queries. This is driven by five converging mechanisms: training distribution pattern activation, n-gram co-occurrence, the "should exist" inference, RLHF-induced authority pressure, and fluency loops.
Rejection-Induced Confabulation Escalation: A theoretical model proposing that repeated rejection of AI responses in prompt chains systematically increases fabrication probability. This suggests that agentic quality filters, designed to improve specificity, may paradoxically function as hallucination accelerators in compliance contexts.
URL Morphology Taxonomy: An empirical classification of 946 fabricated sources, revealing that 50.5% are "Plausible Mirages" (HTTP 404s on real domains) and that 51.9% of these target regulatory authority domains (e.g., Hungarian competition or data protection authorities) rather than the entity's own domain.

Empirical Results

The study provides statistical evidence supporting the Brand Hallucination Paradox and the CEE Entity Infrastructure Gap:

Fabrication Rates by Tier: High-salience (Tier 1) entities exhibited a significantly higher fabricated citation rate (52.69%) compared to low-salience (Tier 3) entities (37.87%).
- Statistic: $\chi^2 = 45.326$ , $p = 1.67 \times 10^{-11}$ , Cohen's $h = 0.299$ .
- Interpretation: Familiarity increases the likelihood of confident false attribution.
Query Class Effects: Regulatory compliance queries yielded the highest fabrication rate (56.77%), significantly higher than the factual baseline (37.59%).
- Statistic: FDR-corrected $q < 0.001$ .
Platform Differences: Anthropic Claude exhibited a much higher willingness to cite sources (1.34 per run) compared to GPT-4o (0.13 per run), resulting in a higher absolute fabrication rate for Claude (48.17% vs. 22.10%). GPT-4o often refused to provide sources for low-salience entities.
Infrastructure Impact: Tier 1 entities were approximately 7 times more likely to have a presence in open knowledge infrastructure (Wikipedia/Wikidata) than Tier 3 entities (OR = 6.77, $p = 0.0002$ $p = 0.0002$ ).
- Protective Effect: Presence on English Wikipedia was associated with a 63% reduction in the odds of a fabricated source (OR = 0.37, $p < 0.001$ ). Hungarian Wikipedia presence did not show a significant independent protective effect, likely due to lower representation in LLM training corpora.

Significance and Claims

The paper claims that AI visibility must shift from a "map" metaphor (static positioning) to a "field" metaphor (dynamic reconstruction based on evidence density). Its primary significance lies in:

Reframing Visibility: Establishing that a raw mention is not a positive signal but a potential risk vector requiring verification.
Audit Methodology: Providing a methodological bridge from algorithmic fairness research to organizational entity auditing, treating the entity as the unit of calibration rather than the platform average.
Security Implications: Identifying "passive adversarial amplification," where actors can exploit regulatory framing to generate credible, non-existent citations without model access or prompt injection.
Strategic Necessity: Arguing that organizations cannot control how generative systems represent them directly, but can alter the "evidence field" (structured data, canonical identifiers, third-party citations) from which those representations are reconstructed.

The paper concludes that without entity-specific calibration distinguishing between presence and fidelity, organizations risk measuring different phenomena under the same label, rendering current visibility strategies ineffective in the agentic web.