How to Measure AI Citations

Almost every dashboard you already use was built for the click era. Search Console counts impressions and clicks. Analytics counts sessions. Neither has a row for the thing that increasingly decides whether anyone hears about you, which is how often a machine repeats your name.

The gap is not permanent, but right now it means most owners are flying blind on their most valuable distribution. It also means the few people who bother to look have an unfair advantage, because the data is free and sitting unopened.

There are exactly three views worth building a habit around. The first is Microsoft's AI Performance report, which exists because Copilot runs on Bing and ChatGPT search leans on the same index. The second is your access log, which shows the crawlers arriving and, more usefully, which pages they keep coming back to. The third is the least technical and the most revealing: open a fresh chat, ask about your own company, and read the answer as a stranger would.

Read together, those three answer different questions. The report tells you how often. The log tells you what they read. The chat tells you what they concluded. When all three agree, you know exactly which page is doing the work, and that is usually not the page you expected.

One more thing that same log will tell you, and it is uncomfortable. Most of what your dashboard calls traffic is not people at all. On one of my sites a seven day window held 26,919 requests, 58 percent of which announced themselves as machines, and the log turned out to be recording my CDN's addresses rather than anyone's, so the rest could not be identified at all. Splitting that number is the first honest step before you measure anything else.

One honest caveat before you start. Every number here is a floor rather than a ceiling. Bing sees Microsoft surfaces and whatever rides its index, so Claude and Perplexity quote you invisibly. Whatever you measure, the real figure is higher.

01The working notes