AEO A/B Testing: How to Measure Whether Your GEO Changes Are Actually Working
Unlike SEO, AEO improvements cannot be measured with simple before/after traffic analysis. A methodologically sound framework for testing structural changes, schema updates, and content rewrites against AI citation metrics.
SEO A/B testing has well-established methodologies: change the page, wait for re-crawl, measure position and traffic delta. AEO testing is more complex because AI citation is non-deterministic (the same query returns different results across runs), the measurement window is longer (AI index refresh cycles vary by platform), and there is no direct traffic metric for zero-click AI citations.
A rigorous AEO testing framework accounts for these characteristics. Without it, teams invest in changes that appear to work due to noise, and abandon changes that genuinely improve citation probability.
Why AEO testing is different from SEO testing
- Non-determinism: AI search results vary across queries, users, and time. The same query at different times or from different geographic locations may return different citations. You need multiple samples per query, not single measurements.
- Lag variability: Perplexity may reflect your changes within days. ChatGPT may take weeks or months. You need platform-specific measurement windows.
- No universal metric: Unlike organic traffic, there is no single citation metric accessible to everyone. Share of Model audits require manual query testing across platforms.
The citation testing framework
The most reliable AEO testing approach uses matched page pairs: two pages covering similar topics at similar authority levels, where one receives the optimization change (treatment) and one does not (control). Citation frequency for both pages is tracked before and after the change across the same query set.
Why you cannot just test one page
Setting up control and treatment
Select matched pages
Choose two pages with similar current citation rates, similar domain authority, and similar topical distance from your target queries. They do not need identical content — just similar citation baseline.
Establish baseline
Run 10–15 queries per page, across 3 platforms (Perplexity, ChatGPT, Gemini), 3 times each. Record citation presence and quality tier for each. Average these into a baseline citation rate.
Apply change to treatment only
Implement the optimization change (schema addition, structural rewrite, content density increase) to the treatment page only. Leave the control page unchanged.
Wait for index refresh
Allow 4–6 weeks minimum before re-measurement. Perplexity: 1–2 weeks. ChatGPT: 4–8 weeks. Gemini: 2–5 weeks. Measure each platform at its appropriate window.
Compare delta
Calculate the citation rate change for treatment vs control. If treatment improves and control holds steady, the change is likely effective. If both change together, external factors are confounding.
What to measure
| Metric | What it measures | Sensitivity |
|---|---|---|
| Citation presence rate | % of queries where page is cited at all | High — most sensitive to structural changes |
| Citation quality tier | Primary recommendation vs secondary mention | Medium — measures positioning improvement |
| Citation consistency | Variance across repeated queries | Medium — measures stability of citation |
| Platform-specific citation | Presence per engine separately | High — isolates platform-specific signal effects |
Sample size and timing
Minimum reliable test: 10 queries tested 3 times each = 30 data points per page, per platform. For statistically meaningful results at 95% confidence with a 10% baseline citation rate, you need at least 50 data points per condition. Run your queries across different times of day and week to account for LLM response variation.
The five most valuable AEO tests to run
- FAQ schema addition: Add FAQPage schema to a content page vs identical content without schema. Typically shows 15–40% citation rate improvement.
- Answer-first restructure: Move the direct answer from conclusion to introduction. Typically shows 10–25% citation rate improvement.
- Comparison table addition: Add a structured HTML comparison table vs keeping the same content as prose. Typically shows 20–35% improvement for comparison queries.
- Content density increase: Reduce word count by 30% while preserving factual claims. Typically shows 8–18% improvement on information-dense engines like Perplexity.
- Timestamp implementation: Add ISO 8601 dateModified to pages lacking it. Typically shows 10–30% improvement on time-sensitive queries, primarily on Perplexity.