💬 NLP

Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines

This paper establishes a large-scale baseline for Generative Engine Optimization (GEO) by analyzing over 100,000 AI search responses to reveal that brand visibility is heavily stratified by maturity, favors corporate websites and "best-of" listicles as primary sources, and exhibits highly unstable sentiment, thereby providing actionable protocols for brands to improve their AI search presence.

Original authors: Pratyush Kumar (Ranqo)

Published 2026-06-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Pratyush Kumar (Ranqo)

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: From Blue Links to AI Conversations

Imagine the internet used to be a giant library with a card catalog. If you wanted to find a book, you looked up a keyword, got a list of "blue links" (traditional search results), and clicked the one that looked best. This was SEO (Search Engine Optimization).

Now, the library has a super-smart librarian (AI) who doesn't just give you a list of books. Instead, you ask a question like, "What's the best running shoe for flat feet?" and the librarian reads thousands of books, synthesizes the information, and gives you a direct answer with a few specific recommendations.

This new world is called GEO (Generative Engine Optimization). The paper asks: How do brands get the librarian to mention them in that answer?

The Main Discovery: The "Fame Ladder"

The researchers tracked over 100 brands across five different AI engines (like ChatGPT, Claude, and Gemini) to see who gets mentioned. They found a strict three-step ladder of fame that determines who gets heard:

The Household Names (Tier 1): Think Nike or Stripe. These are the celebrities of the internet. When you ask the AI a general question, these brands appear in the answer 73% of the time immediately. They are so famous the AI knows them by heart.
The Established Mid-Size Brands (Tier 2): Think Olipop or Klaviyo. These are well-known but not global giants. They get mentioned about 44% of the time.
The Niche & Small Brands (Tier 3): Think local startups or small creators. They get mentioned only 11% of the time.

The Analogy: Imagine a crowded party where everyone is shouting recommendations. The celebrities (Tier 1) are heard almost every time. The mid-size brands (Tier 2) get heard about half the time. The small brands (Tier 3) are like people whispering in the corner; the AI simply doesn't hear them unless you specifically ask for them by name.

Key Takeaway: If you are a small brand, the AI doesn't "forget" you; it just hasn't learned your name yet. The gap between the tiers is huge (about 30 percentage points per step down).

How the AI "Thinks" and "Cites"

The paper looked at where the AI gets its information. It's not just looking at a brand's own website.

The "Third-Party" Rule: When the AI gives an answer, 78% of the time it cites other companies' websites, not the brand's own site. It's like a food critic recommending a restaurant by quoting other food critics and review sites, rather than reading the restaurant's own menu.
The "Listicle" King: The most powerful type of content to get cited is a "Best Of" list (e.g., "Top 10 CRM Tools"). If a brand gets on one of these lists, that single page acts like a golden ticket, getting cited over and over again for many different questions.
Video is the New Giant: Surprisingly, YouTube is cited more often than Reddit, Wikipedia, or news articles. If a brand isn't on video, it's missing a huge chunk of the AI's "reading list."

The "Invisibility" Problem

The researchers found something strange about how often brands appear:

It's Binary: For most small brands, the AI either always mentions them or never mentions them. It's not a coin flip where they appear sometimes and disappear others. They are either "in the club" or "locked out."
The "Silent" Decline: If a brand does nothing to improve, their visibility doesn't crash dramatically. Instead, it slowly erodes (like a sandcastle losing a little bit of sand every day), especially on ChatGPT and Perplexity.
Sentiment is Chaotic: While whether a brand is mentioned is stable, how the AI talks about them (positive or negative) is wild. The AI flips its opinion on a brand 6.7 times more often than it decides to mention or ignore them. One day it sounds great; the next, it sounds neutral.

What This Means for Small Businesses

The paper offers a few practical rules for brands that aren't famous yet:

Know Your Tier: Don't try to optimize for "global fame" if you are a local shop. Focus on getting into the "Mid-Size" tier first by building a strong web presence (Wikipedia, press, YouTube).
Treat Engines Like Different Countries: What works for ChatGPT doesn't work for Perplexity. They use different sources and have different rules. You can't just copy-paste your strategy across all of them.
Put Facts Up Front: Don't hide your key facts in complex code or hidden text. Tell the AI clearly and early in your article: "We are the best X because of Y." The AI loves clear, quotable facts.
Check Often, Not Once: The AI changes its mind quickly. You need to check your visibility every few weeks, not every few months.

What the Paper Does NOT Say

It is important to stick to what the paper actually found:

No Magic Bullet: The paper does not prove that writing a specific article will guarantee you get mentioned. It only shows what currently gets cited.
No "Hacking": The authors explicitly warn against trying to "trick" the AI with hidden text or spam. That doesn't work and gets you banned.
No Causal Proof Yet: The paper says, "We see that big brands are mentioned more." It does not yet prove that if a small brand changes its website, it will instantly jump to the top. That is a future experiment they plan to run.

Summary Analogy

Think of the AI search engine as a very busy news anchor who reads from a stack of papers to answer your questions.

Big Brands have their own papers on the anchor's desk.
Small Brands are hoping the anchor picks up a paper written by someone else that mentions them.
The "Best Of" Lists are the most popular papers on the desk.
The Goal isn't to shout louder (SEO); it's to get your facts printed on the papers the anchor is already reading.

Technical Summary: Generative Engine Optimization at Scale

1. Problem Statement

The traditional paradigm of Search Engine Optimization (SEO), focused on ranking for "blue links" in classical search engines, is being superseded by Generative Engine Optimization (GEO). As users shift from querying traditional engines (Google, Bing) to AI assistants (ChatGPT, Claude, Gemini, Perplexity, Grok), the success metric for brands changes from "ranking for a keyword" to "being cited and recommended by a Large Language Model (LLM)."

While authoritative global brands often receive citations naturally due to their existing web footprint, a significant gap exists for Small and Medium Enterprises (SMEs), Direct-to-Consumer (D2C) brands, creators, and early-stage startups. These entities lack the established online presence required to be surfaced by AI models. The paper addresses the lack of large-scale, production-grade data regarding how AI search engines cite brands, what sources they prioritize, and how visibility can be measured and improved for non-dominant players.

2. Methodology

The study utilizes Ranqo (Ranqo AI), a production platform designed to measure and optimize brand visibility across AI search engines. The methodology involves a controlled, large-scale observational study conducted between March and May 2026.

Data Collection and Scale

Cohort: 102 brands tracked across five AI engines: ChatGPT (GPT-5), Gemini (Gemini-3), Perplexity (Sonar), Claude (Sonnet), and Grok (Grok-3).
Volume: 102,025 prompt responses, generating roughly 15,815 brand mentions and 149,912 source citations.
Prompt Strategy: The platform generates natural-language prompts across six categories: Discovery, Problem/Solution, Use Case, Expert, Comparison, and Brand Research. Crucially, the study focuses on unbranded prompts (e.g., "best running shoes") to measure genuine category visibility rather than self-referential mentions.
Configuration: Engines are queried via official APIs. Search-enabled variants are used; ChatGPT and Claude run on a weekly cooldown per brand to control costs (alternating between search-on and search-off states), while others are always grounded.

Measurement Framework

Ranqo extracts and scores four primary signals for every (prompt, platform, run) tuple:

Visibility (Mention Rate): The percentage of prompts where the brand is mentioned.
Average Position: The ordinal rank of the mention (1st, 2nd, etc.).
Sentiment: A continuous score ( $[-1, 1]$ ) and categorical label (positive/neutral/negative).
Share of Voice: The brand's mentions relative to qualified competitors.

Additionally, the platform performs a six-dimension page audit on tracked URLs, scoring: Crawlability, Content Quality, Page Speed, AI Readiness, Citation Potential, and Authority & Trust (E-E-A-T).

Statistical Approach

The paper employs non-parametric statistical methods (Kruskal-Wallis, Mann-Whitney U, bootstrapping with 10,000 resamples) due to the bounded and right-skewed nature of visibility distributions. It explicitly distinguishes between observational baselines and causal claims, reserving causal inference for future Randomized Controlled Trials (RCTs).

3. Key Contributions

The paper provides the first large-scale baseline for measuring GEO, offering:

A Production Measurement Framework: A reproducible method for tracking brand visibility, source attribution, and sentiment across multiple AI engines.
A Category Reference Study: Analysis of 2,500 responses across 10 CRM brands, establishing a baseline for category concentration.
Empirical Findings on Visibility Dynamics: Quantification of how brand stature, prompt type, and engine architecture influence citation behavior.
A Research Agenda: Seven proposed protocols (v1.1) to test specific interventions (e.g., content structure, schema markup, search-grounding) causally.

4. Key Results

A. The Brand-Stature Visibility Ladder

The most significant finding is a clear, three-tier hierarchy in unbranded visibility on the first tracking run:

Tier 1 (Global Household Names): 73% visibility (e.g., Stripe, Nike).
Tier 2 (Established Mid-Market/Regional): 44% visibility (e.g., Olipop, Klaviyo).
Tier 3 (Niche/Small Brands): 11% visibility.
Gap: Each step down the ladder results in a ~30 percentage point drop in visibility. This effect is statistically significant (Cohen's $d > 1.3$ ) and suggests that existing web prominence is the dominant determinant of AI visibility.

B. Day-1 Recognition and Persistence

Branded vs. Unbranded: When a prompt explicitly names a brand, recognition is near-universal (94–100%). However, unbranded recognition is significantly lower (12–52% depending on the engine).
Trajectories: In the absence of intervention, unbranded visibility is largely stable but shows a slow, statistically significant decline on ChatGPT (-1.34% per run) and Perplexity (-0.76% per run). Other engines remain flat.
Determinism: AI surfacing is near-binary. 77.5% of (brand, prompt, engine) cells are either "always mentioned" or "never mentioned." Only 6.8% of cells "flip" between runs, indicating that visibility is a stable property of the brand-engine pair rather than random noise.

C. Source Composition and Citation Patterns

Corporate Dominance: 78% of all citations point to corporate websites.
- Third-Party Corporate Pages: 75.2% (competitors, peers, vendors).
- Brand-Owned Pages: Only 2.9%.
- Implication: AI engines preferentially cite peer-brand product pages when constructing "alternatives" or "best-of" answers, creating a "Matthew Effect" where cited domains pull more citations.
Non-Corporate Sources: Among non-corporate sources, YouTube (4.2%) is the most cited, surpassing editorial media (3.8%), Reddit (3.3%), and Wikipedia (2.6%).
Content Format: Listicles (ranked "best-of" lists) are the highest-leverage format, accounting for 35.7% of all content citations. A single listicle can surface a brand across many different prompts.

D. Sentiment Volatility

Sentiment is significantly noisier than mention data.

Flipping Rate: 45.5% of sentiment patterns flip between positive and negative across runs, compared to only 6.8% for mentions.
Stability: Sentiment is 6.7 times noisier than mention. Consequently, sentiment-weighted scores require a much larger sample size (≥10 prompts) to stabilize.
Negativity: No tracked brand was consistently negative; negative framing was transient rather than systematic.

E. Engine Divergence

Different engines exhibit distinct behaviors:

Claude showed the highest unbranded recognition (51.5%) but operates on a smaller, higher-stature cohort.
Perplexity and ChatGPT showed downward visibility drifts.
Source Overlap: Cross-platform agreement on cited sources is low (Jaccard overlap ~0.12 in the CRM study), confirming that each engine operates as a separate market with unique source preferences.

5. Significance and Claims

The paper positions itself as a foundational, observational study rather than a definitive causal guide. Its significance lies in:

Quantifying the "Stature Gap": It provides empirical evidence that brand maturity is the primary driver of AI visibility, with a ~30-point penalty for each tier drop. This challenges the notion that content optimization alone can instantly elevate small brands to the visibility levels of global giants.
Debunking "Set-and-Forget": The data shows that without active intervention, visibility on major engines (ChatGPT, Perplexity) slowly erodes, necessitating a measurement cycle of weeks rather than quarters.
Shifting Optimization Focus: The findings suggest that for smaller brands, the most effective strategy is not just optimizing their own domain (which receives only 2.9% of citations) but targeting high-leverage external surfaces like listicles and video content (YouTube), and building "brand mass" (Wikipedia, press, sustained presence) to cross the Tier 2 threshold.
Methodological Rigor: The paper explicitly avoids overclaiming. It states that while it identifies strong correlations and baselines, it does not yet prove that specific GEO tactics cause visibility lifts. It proposes seven protocols (P1–P7) for future RCTs to establish causal links, particularly regarding the impact of content quality dimensions (Authority, AI Readiness, Citation Potential) versus technical SEO factors (Schema, Page Speed).

In summary, the paper establishes that AI brand visibility is measurable, highly dependent on brand stature, and distinct across platforms. It argues that the "hard problem" of GEO for small brands requires a shift from simple keyword optimization to building authoritative, citable content structures and securing placement in high-leverage external formats like listicles and video.