AutoGEO: How Generative Search Engines Pick Content — and How to Use It

The research system (for extracting rules):

STEP 1: Analyze document pairs → LLM explains why one document gets cited more
STEP 2: Extract insights → LLM distills patterns from the explanations  
STEP 3: Merge rules → hierarchical consolidation of insights into a single rule set
STEP 4: Filter → keep only relevant preference rules

Applying the rules (two models):

• AutoGEOAPI: rules are baked into a prompt for a powerful LLM → it rewrites the content
• AutoGEOMini: a compact model trained via RL with the rules as rewards → 140x cheaper

The researchers identified concrete patterns that generative engines favor when choosing content to cite:

Universal Rules (work across all engines)

1. Source Citation Tie every factual claim to a specific authoritative source with a clear reference.

2. Comprehensive Cover the topic thoroughly, touching every key aspect and subtopic.

3. Topic Focus Stick strictly to the topic and strip out everything else — navigation, ads, irrelevant info.

4. Neutral Tone Keep an objective tone; avoid promotional language, personal opinions, and bias.

5. Balanced View For contested topics, present several meaningful viewpoints or counterarguments.

6. Self-Contained The information should stand on its own, without needing external links to understand the point.

7. Actionable
For tasks and problem-solving, give clear, concrete, usable steps.

8. In-Depth Explain not just the “what” but the “how” and “why” — causes, mechanisms, context.

9. Conclusion First State the key takeaway at the very beginning of the document.

10. Specific Evidence Back claims with concrete, verifiable data, statistics, and named examples.

Domain-Specific Rules

For open-ended/research topics (Researchy-GEO):

• Writing Quality: polished text with no grammatical errors
• Up-to-date: current information, the latest state of knowledge

For commercial content (E-commerce):

• Step-by-Step Guide: step-by-step instructions or clear recommendations
• Production Details: concrete details — names, model numbers, technical specs
• Pros & Cons: justify recommendations with pros/cons and comparisons
• Modular: modular structure — each paragraph = one concept

For general topics (GEO-Bench):

• Accessibility: content is reachable without logins, subscriptions, or paywalls

Task: You're writing a business blog post about moving to remote work. You want ChatGPT or Claude to cite your piece when answering questions about remote work.

Optimizing with AutoGEO rules:

1. Conclusion first (Conclusion First):

Moving to remote work takes three stages: technical prep (2-3 weeks), 
team adaptation (1-2 months), settling into a rhythm (3-6 months). The key success factor is 
transparent communication and clear KPIs from day one.

2. Specifics instead of vague phrases (Specific Evidence):

• ❌ “Lots of companies have gone remote”
• ✅ “According to TechCrunch, 67% of US tech companies switched to a hybrid model in 2023”

3. Step-by-step instructions (Step-by-Step):

Week 1: Infrastructure audit
- Check employees' internet speed (at least 50 Mbps)
- Set up VPN and access to company systems
- Buy equipment: Logitech C920 webcams, noise-canceling headphones

Weeks 2-3: Trial mode
[...]

4. Balanced viewpoints (Balanced View):

Arguments for remote work:
- Office rent savings: $15,000-40,000/month for a 20-person team
- Access to talent outside major hubs

Arguments against:
- Risk of a 15-20% drop in engagement (Stanford study, 2023)
- Harder to control quality for new hires

5. Self-contained (Self-Contained): All the key information lives in the article — no “read more on another site.” Links serve only as data sources.

Result: Text written by these rules lands in generative search answers 36% more often. The AI sees the full picture, concrete data, practical steps, and balanced opinions — a valuable source worth citing.

The weakness of generative engines: When building an answer, the AI has to choose which of 5-10 documents to cite. The selection criteria are opaque — content owners have no idea why their material gets ignored.

The strength of LLMs: They can analyze patterns across huge volumes of data. The researchers took tens of thousands of real queries and answers, picked for each the document pair with the biggest gap in citation, and asked an LLM to explain the preference. Then they pulled common patterns out of those explanations.

The mechanics: The rules work because they reflect the engines' actual behavior, not theoretical guesses. It's reverse engineering of preferences:

1. We see: document A is cited at 40%, document B at 10%
2. We analyze: how A differs from B
3. We extract: “A has concrete statistics, B has vague phrases”
4. We state the rule: “Use concrete data”

Per-engine personalization: The rules differ. Gemini and GPT overlap by 79%, GPT and Claude by 84%. Each engine has unique preferences. For example, E-commerce cares about technical specs and step-by-step guides, while research topics care about depth and freshness.

Control knobs when applying the rules:

• Rule prioritization: Not all rules matter equally for your domain. Experiments show different rules dominate in different contexts (see Fig. 3 in the paper)
• Engine targeting: If you know the target engine (Google, ChatGPT, Claude), use its specific rules
• Rule combination: The full set beats individual rules — they reinforce each other

When rewriting text to boost visibility in generative search engines, follow these rules:

STRUCTURE:
1. Main takeaway goes in the first paragraph
2. Stay strictly on topic — cut everything extraneous
3. Modularity — each paragraph = one idea

EVIDENCE:
4. Every fact comes with specific sources and references
5. Specifics: statistics, names, numbers instead of “many,” “often”
6. For contested topics — several viewpoints

COMPLETENESS:
7. Cover every aspect of the topic
8. Explain not just “what” but “how” and “why”
9. Self-contained — no clicking through links required to understand

PRACTICALITY (for instructions/recommendations):
10. Clear, concrete steps
11. For products: technical details, specs
12. Justify via pros/cons

QUALITY:
13. Neutral, objective tone
14. No grammatical errors
15. Up-to-date information

---

MY CONTENT:
{paste your text}

MY GOAL:
{e.g.: a blog post about remote work / a product description / 
an expert piece on investing}

TARGET ENGINE (if you know it):
{Gemini / ChatGPT / Claude / not sure}

---

Rewrite the content, applying as many relevant rules as possible. Show exactly what you changed.

Notes on the template:

• You don't have to apply all 15 rules — pick the ones relevant to your content type
• For commercial content, emphasize details and practicality (rules 10-12)
• For expert/educational content — completeness and evidence (rules 4-9)
• Universal rules (1-3, 13-15) always apply

⚠️ Domain specificity: Rules for open-ended topics and E-commerce overlap by only 35-40%. If your content is specialized (legal, medical), effectiveness may drop.

⚠️ Over-optimization: Blindly following every rule can make text feel mechanical. Balancing optimization and natural flow is your call.

⚠️ Shifting preferences: Generative engines get updated. The rules reflect the time of the study (late 2024 — early 2025) and may change over time.

⚠️ Quality vs. visibility: The study optimizes visibility (how often content gets cited), not the final quality of the engine's answer. That said, experiments show quality holds up.

The team built three datasets: GEO-Bench (8000 open-ended queries), E-commerce (1667 commercial ones), Researchy-GEO (10000 research questions). Each query came with 5 candidate documents from ClueWeb22.

They ran generative engines on Gemini, GPT, and Claude. For each query they measured a document's visibility: how many of its words made it into the answer, at what position, and an overall score. They took the pairs with the largest visibility gap.

They asked an LLM to explain: why is document A cited at 40% and document B at only 10%? That yielded thousands of explanations. They extracted insights, merged similar ones through hierarchical consolidation, and filtered out the irrelevant.

Key finding: The rules work! Gemini-specific rules perform better on Gemini (+43% visibility) than universal ones (+35%). But universal rules still work on every engine, just less strongly.

The surprise: Rules for open-ended topics (Researchy, GEO-Bench) overlap by 88%, but for E-commerce only by 35-40%. Domain strongly shapes an engine's preferences. That explains why hand-made heuristics work poorly — they ignore context.

For practitioners: If your content lives in a narrow niche (say, fintech or healthcare), you can run a similar analysis on your own data and extract domain-specific rules.

1. Adaptation for Creating Expert Content

AutoGEO rules make a great pre-publishing checklist:

AI CONTENT OPTIMIZATION CHECKLIST

□ Takeaway/thesis in the first paragraph
□ Every fact has a source (name, year, link)
□ Concrete data instead of “many,” “usually,” “often”
□ On topic only — digressions, promo, navigation removed
□ “How” and “why” explained, not just “what”
□ For contested topics — at least 2 viewpoints
□ Self-contained — key information lives inside the text
□ Neutral tone, no value judgments
□ [For instructions] Step-by-step format with concrete actions
□ [For products/services] Technical details, specs, comparisons

Run your piece through this checklist before publishing — you'll boost your odds of getting cited by AI.

2. Technique: Audit with AI

Ask ChatGPT/Claude to score your content against the AutoGEO rules:

Analyze my text against the criteria generative search engines use 
to choose which sources to cite:

1. Is there a clear conclusion at the start?
2. Are all facts backed by specific sources?
3. Is concrete data used instead of vague phrases?
4. Does it stay on topic, or is there filler?
5. Are the “how” and “why” explained?
6. For contested points — are different viewpoints shown?
7. Is the text self-contained?
8. [If it's a how-to] Are there clear steps?

Score each item on a 10-point scale. Point out the weak spots.

MY TEXT:
[paste text]

3. Extrapolation: Applying It to Personal Notes

The rules work not only for web content but also for personal knowledge bases in assistants:

If you use ChatGPT/Claude to accumulate knowledge (Projects, memory), structure your notes by the same principles:

• Conclusion first: “Meeting outcome: decided to move to the new CRM by end of quarter”
• Specifics: not “discussed the budget” but “approved a $250,000 budget, split: 60% development, 40% integration”
• Self-contained: all key information in the note, not “see the email thread”

Later, when you ask the AI “what did we decide about the CRM?”, it will quote the structured note more accurately.

What Generative Search Engines Like and How to Optimize Web Content Cooperatively (2025) Code and datasets: https://github.com/cxcscmu/AutoGEO

Authors: Yujiang Wu, Shanshan Zhong, Yubin Kim, Chenyan Xiong Affiliations: Carnegie Mellon University, Vody