Autogeo
[ICLR'26] AutoGEO: a Generative Engine Optimization framework to automatically learn generative engine preferences, and rewrite web contents for more traction.
Last verified:
What is Autogeo?
AutoGEO is an open-source framework for Automatic Generative Engine Optimization (GEO) that helps web content gain higher visibility in LLM-generated answers from generative search engines like Gemini, GPT, and Claude. The framework automatically discovers actionable preference rules from generative engines by analyzing which content gets cited, then uses these rules to rewrite documents for better traction without harming response quality.
The system has three core components: (1) Rule Extraction, which automatically mines content preferences from generative engines by prompting frontier LLMs to explain why certain content is preferred; (2) AutoGEO_API, a plug-and-play prompt-based GEO system that uses extracted rules as context engineering; and (3) AutoGEO_Mini, a cost-effective GEO model trained with reinforcement learning (GRPO) that reduces inference costs to approximately 0.7% of API-based methods. AutoGEO achieves up to 50.99% improvement over baselines while maintaining generative engine utility metrics like precision, recall, clarity, and insight.
AutoGEO is designed for researchers, AI visibility specialists, content optimization teams, SEO professionals working with AI search, and developers building generative engine optimization solutions. The framework supports three datasets (Researchy-GEO for academic content, E-commerce for commercial content, and GEO-Bench benchmark) and works across multiple generative engines. It is open-source under MIT license with code, checkpoints, and datasets freely available on GitHub, making it accessible for both research and practical deployment.
Autogeo pricing
Pricing model: Freemium
AutoGEO is completely free and open-source under MIT license. There are no paid plans or subscription tiers. The framework includes free access to all components: rule extraction code, AutoGEO_API implementation, AutoGEO_Mini training scripts, pre-trained checkpoints for E-commerce/GEO-Bench/Researchy-GEO datasets, and all three datasets. Users only incur costs for their own API calls to generative engines (Gemini, GPT, Claude) when using AutoGEO_API or running rule extraction. AutoGEO_Mini significantly reduces costs by cutting inference expenses to approximately 0.7% of API-based methods after the initial training investment.
Autogeo pros
- Automatically extracts preference rules from generative engines without manual prompt engineering
- Up to 50.99% improvement in visibility over strongest baselines
- Dual deployment options: plug-and-play API and cost-efficient fine-tuned model
- Maintains or slightly improves generative engine utility (precision, recall, clarity, insight)
- Cross-engine robustness works consistently across Gemini, GPT, and Claude
- Rule transferability with 78-84% overlap across different LLMs
- AutoGEO_Mini reduces inference costs to ~0.7% of API-based methods
- Open-source with MIT license, free to use and modify
- Pre-trained checkpoints available for E-commerce, GEO-Bench, and Researchy-GEO
- Substantially improves visibility for low-visibility documents
- Cooperative optimization approach without harming answer quality (unlike adversarial attacks)
- Three ready-to-use datasets included: Researchy-GEO, E-commerce, GEO-Bench
- Python-based with clear installation scripts and documentation
- Accepted at ICLR 2026, peer-reviewed research quality
- GRPO reinforcement learning training for efficient model optimization
- Well-documented code with 154 GitHub stars and active community
Autogeo cons
- AutoGEO_Mini requires 2 CUDA-compatible GPUs (A100 40GB+ recommended)
- Training AutoGEO_Mini takes ~4 hours for SFT and ~48 hours for GRPO on Researchy-GEO
- Rule extraction must be rerun when switching to different engines or domains
- AutoGEO_Mini model must be retrained when changing engines or datasets
- Requires API keys configuration for generative engine access
- May encounter GRPO dependency errors requiring open-r1 reinstallation
- Rule extraction tailored to specific engine and dataset combinations limits out-of-box flexibility
- Large GPU memory requirements may be prohibitive for individual researchers with limited hardware
Frequently asked questions about Autogeo
What is AutoGEO and what does it do?
AutoGEO is a Generative Engine Optimization framework that automatically learns generative engine preferences and rewrites web content for more traction. It extracts actionable preference rules from generative search engines like Gemini, GPT, and Claude by analyzing which content gets cited in their responses, then uses these rules to rewrite documents to improve visibility in AI-generated answers while preserving response quality and accuracy.
What is the difference between AutoGEO_API and AutoGEO_Mini?
AutoGEO_API is a plug-and-play prompt-based GEO system that uses extracted preference rules directly as context engineering for frontier LLM APIs like GPT-4 or Claude. It requires no fine-tuning and is easy to integrate. AutoGEO_Mini is a cost-effective GEO model trained with supervised fine-tuning and GRPO reinforcement learning that reduces inference costs to approximately 0.7% of API-based methods, but requires GPU hardware and ~52 hours of training time.
How much does AutoGEO cost?
AutoGEO is completely free and open-source under MIT license. All code, checkpoints, and datasets are freely available on GitHub. The only costs users incur are for their own API calls to generative engines (Gemini, GPT, Claude) when using AutoGEO_API or running rule extraction. AutoGEO_Mini significantly reduces ongoing costs after the initial training investment.
What datasets does AutoGEO support?
AutoGEO supports three datasets: Researchy-GEO (academic dataset for research content), E-commerce (commercial dataset for product/e-commerce content), and GEO-Bench (benchmark dataset from the GEO paper). Each dataset has corresponding pre-trained checkpoints available for Gemini generative engine, and all datasets are included free with the codebase.
Which generative engines does AutoGEO work with?
AutoGEO works with Gemini (e.g., gemini-2.5-flash-lite), GPT (e.g., gpt-4o-mini), and Claude (e.g., claude-3-5-sonnet-20241022). The framework demonstrates cross-engine robustness with consistent improvements across all three engines. Rules show 78-84% overlap across LLMs and transfer effectively between engines, though optimal results require running rule extraction for your specific target engine.
How do I install AutoGEO?
Clone the repository with git clone --recursive https://github.com/cxcscmu/AutoGEO, cd into the directory, run bash install.sh, activate the environment with conda activate autogeo, and configure API keys by editing keys.env. For AutoGEO_Mini training, also run bash install_mini.sh after completing Option 1. The installation includes all dependencies and sets up the Python environment.
What hardware do I need to use AutoGEO?
For AutoGEO_API and rule extraction, you only need a computer capable of running Python and API access to generative engines. For training AutoGEO_Mini, you need 2 CUDA-compatible GPUs with A100 40GB or more recommended. Training takes approximately 4 hours for supervised fine-tuning (cold start) and 48 hours for GRPO reinforcement learning on the Researchy-GEO dataset.
How does AutoGEO improve visibility without hurting quality?
AutoGEO uses a cooperative optimization approach rather than adversarial attacks. It extracts explicit preference rules from generative engines by having frontier LLMs explain why certain content is preferred over other content. These rules guide content rewriting to match engine preferences (like prioritizing quantitative data, maintaining authoritative tone) while preserving factual accuracy and key information, maintaining or slightly improving utility metrics like precision, recall, clarity, and insight.
What performance improvements can I expect from AutoGEO?
AutoGEO_API achieves up to 50.99% improvement over the strongest baseline, while AutoGEO_Mini achieves an average 20.99% gain over baselines. Both consistently outperform all baselines across all three datasets. The framework substantially improves visibility for low-visibility documents and maintains consistent improvements across Gemini, GPT, and Claude engines while preserving answer quality.
Do I need to retrain when switching to a different engine or domain?
Yes. Rule extraction is tailored to specific generative engines and datasets/domains. When switching to a different engine (e.g., from Gemini to GPT) or different dataset/domain, you must rerun rule extraction for AutoGEO_API and retrain AutoGEO_Mini accordingly. The framework provides scripts for rule extraction and training, and rules should be extracted fresh for each new engine-domain combination for optimal results.