E-GEO: Optimizing Product Descriptions for AI Search

TL;DR

E-GEO is a study of how to rewrite product descriptions so that AI assistants (Amazon's Rufus, ChatGPT with search) rank them higher. The gist: when a shopper asks "recommend some kitchen knives under $150", the AI looks at a pool of products and decides which ones to show first. The product description influences that choice.

The headline finding: 10 of 15 popular rewriting heuristics (FAQ, authoritative tone, technical jargon) don't work or actively hurt. Adding an FAQ gives +0.06 positions; a minimalist style costs 1.66 positions. Intuition lets you down: what seems "professional" or "persuasive" to you, an LLM judges differently.

The solution: The researchers took weak starting prompts and ran them through a self-improvement algorithm. After optimization, all 15 prompts converged on a common pattern, a universal strategy that works regardless of product type. Even the "write a fictional story" prompt (initially −4.03 positions) gives +1.22 after optimization.

Method Blueprint

ONE PROMPT → Rewrite the product description

Input: Original product description
Output: Description optimized for AI ranking

Key elements of an optimized description:
├── Focus on the shopper's intent (user intent)
├── Competitive positioning
├── References to reviews/ratings
├── Unique selling points (USP)
├── Easy-to-scan format (headings, bullets)
└── Factual accuracy preserved

Example in Practice

Task: An Amazon seller wants to improve the visibility of their kitchen knives in AI recommendations (Amazon's Rufus assistant is already live).

Prompt:

Rewrite the product description to rank as high as possible in AI search engines. 
Goal: when a shopper asks an AI "recommend some kitchen knives", my product 
should come out ahead of the competition.

Use this structure:
1. Start with the problem the product solves (anticipate the shopper's intent)
2. Show the competitive advantages — how it beats alternatives
3. Include social proof: reviews, ratings, awards
4. Highlight the unique features (USP)
5. Use headings and bullets for scannability
6. Add a call to action with a sense of urgency
7. Keep every fact — don't make anything up

Product description:
---
"Samurai" stainless steel kitchen knife set. Includes: 8-inch chef's knife, 
5-inch utility knife, 3.5-inch paring knife. Rosewood handles. 
Made in China. Steel hardness 58 HRC.
---

Result: The model will produce a rewritten description with a clear structure: problem → solution → advantages over competitors → social proof → technical details → call to action. The format will be easy to scan, with headings and lists.

Why It Works

The LLM's weakness in ranking: the model compares descriptions head-on: which one best matches the user's query. If a description is written for humans (an ad slogan, emotion, creativity), the AI may miss the relevance. The model doesn't read context the way a person does; it looks for explicit match signals.

The LLM's strength: the model is excellent at matching structured information to a query. If the description says outright "solves problem X" and the user's query is about X, the match is obvious to the model.

How the method uses this: optimized descriptions make relevance explicit. Instead of "premium knives for connoisseurs", it's "solves the problem of dull knives you have to sharpen every week". Instead of "excellent quality", it's "rated 4.8 from 2,000+ reviews, 58 HRC hardness vs. 55 for competitors".

Control knobs:

• Competitive comparison section — drop it if you don't want to name competitors directly
• Urgency element — drop it for products with no stock limits
• Format — can be simplified to plain prose if the AI system doesn't support markup

Prompt Template

Rewrite the product description to rank as high as possible in AI search systems.

Your goal: when a shopper asks an AI assistant a question, your product should 
rank above the competition.

REQUIRED ELEMENTS:
- Start with the shopper's problem that the product solves — anticipate their intent
- Show competitive advantages: specifically, how it beats alternatives
- Include social proof: reviews, ratings, awards, certifications
- Highlight 2-3 unique selling points (USP) — things others don't have
- Use headings and bullets for easy scanning
- Add a call to action with a sense of urgency or scarcity
- IMPORTANT: keep every fact from the original, don't make anything up

Product description:
---
{product_description}
---

Placeholders:

• {product_description} — paste the current description from the marketplace or product page

🚀 Quick start — paste into the chat:

Here is a template for optimizing a product description for AI search. 
Adapt it to my product: [product name and category].
Ask me questions to fill in the description.

[paste the template above]

The LLM will ask about your current description, competitors, reviews, and unique features, because these elements are critical for ranking in AI systems.

Limitations

⚠️ Only works at the re-ranking stage: If the product doesn't make the initial cut on keywords or embeddings, rewriting the description won't help. It has to pass the basic relevance filter first.

⚠️ Tested on GPT-4o: Results may differ for other AI ranking systems. The researchers report "qualitatively similar" results for Claude, but without exact numbers.

⚠️ The effect is modest in absolute terms: The average gain is 1-1.5 positions out of 10. But within a top 10, that can be the difference between 5th and 3rd place, which is critical for conversion.

How They Tested It

The team collected 7,151 realistic queries from the BuyItForLife subreddit, where people ask things like "which kitchen knives will last 20 years" or "recommend a hiking backpack that'll last a lifetime". These aren't short keyword queries like "white leather chair" but detailed requests with context, budget, and constraints. For each query, they picked 10 relevant products from the Amazon Reviews dataset (48 million products).

Then they tested 15 rewriting heuristics: ad-copy style, authoritative tone, FAQ, minimalism, technical jargon, storytelling, and others. Most had zero or negative effect. After that they ran a meta-optimization algorithm: GPT-4o analyzes the results, critiques the prompt, and proposes improvements.

The unexpected discovery: all 15 optimized prompts converged on similar patterns, even though they started from radically different places. Even the "write a fictional story" prompt (initially −4.03) contained the same elements as the others after optimization. That points to a universal strategy rather than a bag of context-specific tricks.

Original Prompts from the Paper

Context: One of the 15 optimized prompts (competitive), which showed the best result (+1.61 positions):

Transform the following product description into a highly optimized version that 
maximizes its chances of ranking at the top in a generative AI ranking engine. 
The goal is to make this product the most relevant and appealing choice for user 
queries. Include elements like:

- A compelling, benefit-focused headline that captures attention
- Clear, direct language that highlights key features and user benefits
- Semantic richness with relevant keywords and phrases that match potential search 
  queries
- Structured formatting using bullet points and subheadings for easy scanning
- Unique selling points (USPs) emphasized prominently to stand out from competitors
- Social proof references (e.g., 'highly rated,' 'top choice among professionals')
- A sense of urgency or exclusivity (e.g., 'limited edition,' 'best-seller status')
- Comparison hooks (e.g., 'outperforms leading brands in durability and design')
- Natural language integration that reads well while also appearing relevant to 
  AI-driven content analysis

Maintain the core factual accuracy of the original description while enhancing 
its appeal. Avoid excessive promotional language that might seem insincere.

Description: {description}

Adaptations and Extrapolations

💡 Adaptation for service descriptions:

The same principle works for B2B services, where clients increasingly use AI to find vendors:

Rewrite the service description to rank as high as possible in AI search systems.

REQUIRED ELEMENTS:
- Start with the client pain point you solve
- Show results: specific numbers, case studies
- Competitive advantages: how you differ from the market
- Social proof: clients, testimonials, ratings on review platforms
- 2-3 unique selling points (USP)
- A call to action with a specific next step

Service description:
---
{service_description}
---

🔧 Technique: turning off the urgency element:

For products/services without scarcity, drop the urgency block; it can come across as manipulative:

- • Add a call to action with a sense of urgency or scarcity
+ • Add a clear call to action with a specific next step
```

💡 **Extrapolation: optimizing your LinkedIn profile for AI recruiting:**

Recruiters use AI to screen candidates. Same principles:
```
Rewrite my LinkedIn experience description so that AI recruiting systems 
rank me higher.

REQUIRED ELEMENTS:
- Start with the business problem I solved
- Show competitive advantages: how my approach was different
- Specific results with numbers
- Key skills — spelled out, not between the lines
- A structured format for scanning

Current experience description:
---
{position_description}
---

Resources

Paper: "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce" (November 2025)

Authors: Puneet S. Bagga (Columbia), Vivek F. Farias (MIT Sloan), Tamar Korkotashvili (MIT EECS), Tianyi Peng, Yuhang Wu (Columbia Business School)

Code and data: github.com/psbagga17/E-GEO

Key references:

• Aggarwal et al. (2024) — the first definition of GEO
• GEPA (Agrawal et al., 2025) — a reflective prompt optimization algorithm