We started tracking citations across six AI engines for clients about twelve months ago. The original goal was simple: give brands a way to see whether and how often AI engines named them. Twelve months in, the data has reshaped how we think about marketing measurement.
The patterns we expected showed up. Engines reward repetition. Schema matters. Authorship attribution lifts rates. Off-site mentions compound. None of that surprised us, and the early playbooks worked.
The patterns we didn't expect mattered more.
The first surprise was the speed of recovery for previously-invisible brands. We assumed that brands starting from zero would need quarters of content investment to register. Many of them registered within six weeks. The lesson: AI engines are far more reactive to new high-quality material than Google. They notice freshness faster.
The second surprise was the volatility of citation rate week to week. Most brands have meaningful week-on-week swings in their AI citation rate — even brands with stable rankings. The drivers are not always identifiable. Engines update their models, change their retrieval strategies, surface different sources for the same question. Reading a single week's numbers in isolation is misleading; the trend over four-to-six weeks is the real signal.
The third surprise was the gap between engines. We assumed citation rates would correlate strongly across engines. In practice, they vary more than we expected. A brand can be cited consistently by Perplexity and Claude while being nearly invisible on ChatGPT, or vice versa. The reasons trace back to differences in training data, retrieval architecture, and content sources each engine prefers. Cross-engine optimisation is harder, and more important, than cross-search-engine optimisation ever was.
The fourth surprise was the framing effect. Whether you're cited matters; how you're described matters almost as much. Brands cited with positive, specific framing convert downstream traffic at meaningfully higher rates than brands cited with neutral, generic framing. We now treat framing as a primary GEO metric in its own right.
The data has changed how we work. The playbook from twelve months ago was a starting point. The one we use now is sharper, more engine-specific, and far more weighted towards framing and external authority. We expect to be saying the same thing about today's playbook twelve months from now.