---
title: "Experiments & Case Studies for AEO: 20 Tactics"
description: "Experimental AEO tactics come from tests and case studies run by practitioners. Results vary by site."
url: https://aeo.so/categories/experiments
author: "Paul Andre de Vera"
updated: 2026-11-15
---

# Experiments & Case Studies

Experimental AEO tactics come from tests and case studies run by practitioners. Results vary by site. Treat these as hypotheses to measure rather than guaranteed wins.

This is step 7 of [how to do answer engine optimization](https://aeo.so/learn): test, measure and repeat.

Read the guides:

- [What is generative engine optimization (GEO)?](https://aeo.so/guides/generative-engine-optimization)
- [Is SEO dead?](https://aeo.so/guides/is-seo-dead)
- [What is query fan-out?](https://aeo.so/guides/query-fan-out)
- [AEO examples: pages that get cited, and why](https://aeo.so/guides/aeo-examples)
- [How to measure AEO: tracking AI citations and traffic](https://aeo.so/guides/measure-ai-visibility)
- [AEO tools: free and paid tools for AI search visibility](https://aeo.so/guides/aeo-tools)
- [Free AEO course: learn answer engine optimization in 10 lessons](https://aeo.so/guides/aeo-course)
- [State of AEO 2027: which sources AI engines cite](https://aeo.so/research/state-of-aeo)

## Create content specifically for AI training data

Creating content for AI training data means publishing accurate, in-depth, openly available material that future language models are likely to learn from. Pages that are clear, well-sourced and freely crawlable have a better chance of shaping what models know about your topic and your brand in later releases.

Read more: https://aeo.so/tactics/create-content-specifically-for-ai-training-data

## Experiment with prompt-optimized content

Prompt-optimized content is written to match the prompts people type into AI assistants, with direct answers to the task-style questions users ask. This experimental approach tests whether mirroring real prompts makes AI systems quote your page more often and more accurately in their responses.

Read more: https://aeo.so/tactics/experiment-with-prompt-optimized-content

## Experiment with AI-generated video content

Experimenting with AI-generated video means testing tools that turn scripts or articles into short videos for YouTube and social platforms. AI search increasingly cites video results, and low-cost generated video lets you test whether visual versions of your content earn more visibility than text alone.

Read more: https://aeo.so/tactics/experiment-with-ai-generated-video-content

## Monitor AI-driven search results

Monitoring AI-driven search results means regularly checking how ChatGPT, Perplexity, Gemini and Google AI Overviews answer the questions that matter to your business. Tracking whether you're mentioned, cited or misdescribed shows how AI systems interpret your content and where to improve your visibility.

Read more: https://aeo.so/tactics/monitor-ai-driven-search-results

## Understand and optimize for Retrieval Augmented Generation (RAG)

Retrieval augmented generation (RAG) is the process where an AI system fetches relevant documents and uses them to write its answer. Optimizing for RAG means writing clear, self-contained, factual passages that a retrieval step can find and a language model can quote with confidence.

Read more: https://aeo.so/tactics/understand-and-optimize-for-retrieval-augmented-generation-rag

## Study natural language processing fundamentals

Studying natural language processing fundamentals means learning how AI systems turn text into meaning, from TF-IDF to embeddings and transformers. Knowing how models represent and compare language helps you write content that matches how they interpret and rank information, which improves every other AEO tactic.

Read more: https://aeo.so/tactics/study-natural-language-processing-fundamentals

## Optimize for predictive summaries and query expansion

Predictive summaries and query expansion are techniques AI systems use to anticipate what a user needs beyond the literal question. Pages that cover the likely next questions, related subtopics and broader intent behind a query match more of these widened searches and appear in more AI answers.

Read more: https://aeo.so/tactics/optimize-for-predictive-summaries-and-query-expansion

## Test content formats using semantic similarity analysis

Semantic similarity analysis compares how closely different versions of your content match a target query in meaning, using embeddings rather than keywords. Testing formats this way shows which structures AI retrieval systems rank highest for a given question, giving you data instead of guesses.

Read more: https://aeo.so/tactics/test-content-formats-using-semantic-similarity-analysis

## Experiment with different chunking strategies for content optimization

Chunking is how AI systems split a page into passages before retrieving and ranking them. Experimenting with section length, heading placement and paragraph structure shows which chunking approach keeps your answers intact and retrievable, improving how often AI systems pick your passages.

Read more: https://aeo.so/tactics/experiment-with-different-chunking-strategies-for-content-optimization

## Implement content A/B testing for AI search performance

A/B testing content for AI search means publishing two versions of a page or section and measuring which one AI systems cite or reference more often. Systematic tests replace assumptions with evidence about which formats, lengths and structures work for your topics and queries.

Read more: https://aeo.so/tactics/implement-content-a-b-testing-for-ai-search-performance

## Analyze ChatGPT Deep Research activity to understand fan-out queries

ChatGPT Deep Research reveals the fan-out queries it runs while researching a question. Analyzing that activity shows the themes the system digs into behind a topic, and writing content toward those main themes works better than chasing every individual long-tail query one at a time.

Read more: https://aeo.so/tactics/analyze-chatgpt-deep-research-activity-to-understand-fan-out-queries

## Use Deep Research Refinement Synthesizer to identify content gaps

A Deep Research refinement synthesizer collects the qualifying criteria ChatGPT Deep Research applies when narrowing recommendations, such as budget, team size or industry. Mapping those criteria against your content shows the gaps to fill, letting you publish pages that answer the exact refinements AI systems use to choose what to recommend.

Read more: https://aeo.so/tactics/use-deep-research-refinement-synthesizer-to-identify-content-gaps

## Implement LLM Research Path Analyzer

An LLM research path analyzer traces the steps an AI system takes while researching a question: the searches it runs, the sources it opens and the facts it verifies. Studying those paths shows how models gather evidence and reason, which tells you what your content needs to include to be selected.

Read more: https://aeo.so/tactics/implement-llm-research-path-analyzer

## Inspect AI chat network requests to see the searches behind answers

Open your browser's developer tools while using ChatGPT, Gemini or another AI search tool, and watch the network requests. They often show the searches the model ran and the sources it got back. Test each model separately, as their data partners and cited sources differ, and note the domains each one cites for your queries.

Read more: https://aeo.so/tactics/inspect-ai-chat-network-requests-to-see-the-searches-behind-answers

## Report AI visibility next to traffic and revenue for stakeholders

Set expectations with leadership before they form a view from the news: organic clicks can drop while revenue, sales-cycle length and AI visibility tell a fuller story. Report several KPIs across Google, AI platforms and traffic segments, including branded search. Then no single weak number defines the channel.

Read more: https://aeo.so/tactics/report-ai-visibility-next-to-traffic-and-revenue-for-stakeholders

## Build AI tracking prompts from sampled query fan-outs

Build your AI visibility prompts from real fan-out data instead of guessing. Run a seed question about 20 times, collect the sub-queries the AI searches for, cluster them, and turn the most frequent clusters into tracking prompts. Test each model separately, as Gemini, ChatGPT and Claude search and cite differently.

Read more: https://aeo.so/tactics/build-ai-tracking-prompts-from-sampled-query-fan-outs

## Measure AI search with input, channel and performance metrics

Measure AI search in three tiers. Input metrics track how well your content covers the topics and entities that matter. Channel metrics track citations and share of voice in AI answers. Performance metrics track AI-referred visits and conversions. Treat AI search as its own channel, as pages that don't rank can still be cited.

Read more: https://aeo.so/tactics/measure-ai-search-with-input-channel-and-performance-metrics

## Check AI visibility tools against customer surveys

Treat AI visibility tools as directional estimates, not exact counts. Check their numbers against what customers tell you by adding a "How did you hear about us?" question to sign-up and sales forms, with ChatGPT, Perplexity and other AI tools as options. Rising AI answers there confirm what the trackers suggest.

Read more: https://aeo.so/tactics/check-ai-visibility-tools-against-customer-surveys

## Use Microsoft Clarity to see what AI-referred visitors do

Use Microsoft Clarity, a free behavior analytics tool, to see how visitors from AI answers move through your site, alongside Bing Webmaster Tools for how AI and search surface it. Microsoft Advertising's AI search guide pairs the two. Watch recordings and heatmaps for AI-referred sessions to find where they stall or convert.

Read more: https://aeo.so/tactics/use-microsoft-clarity-to-see-what-ai-referred-visitors-do

## Check LLM-based content scoring against real readers

If you use an LLM to rate tone, clarity or offensiveness in your content, check its scores against real readers before trusting them. A 2025 study of nine LLMs (arXiv 2311.09730) found their ratings sat closer to White annotators' than Black or Asian annotators', and adding the audience's identity to the prompt didn't reliably fix it.

Read more: https://aeo.so/tactics/check-llm-based-content-scoring-against-real-readers
