How do I track AI search visibility?

AI search visibility tracking has to work differently from classic rank tracking because the thing being measured is structurally different. A keyword rank is a stable, checkable position: this term, on this search engine, on this day, holds position four. An AI assistant's response to a question is not a fixed position, it is a generated output that can vary between runs of the same query, drift as the underlying model or index updates, and differ meaningfully between ChatGPT, Perplexity, Google AI Overviews, and other systems that each have their own retrieval and generation logic. Treating any single check as a stable number is the first mistake to avoid.

The practical foundation is a fixed query set, built the same way good keyword research is built: real questions a real buyer would ask, phrased the way people naturally talk to an assistant rather than the fragment style typed into a search box. This should include direct brand questions, category questions where a business would reasonably expect to be mentioned as an option, and comparison questions against known competitors. The set needs to stay fixed over time so that changes in outcome reflect changes in visibility rather than changes in what was asked.

Running that query set means manually or semi-programmatically submitting each question to the assistants that matter for the business's audience, on a recurring cadence, weekly or monthly depending on how much change is expected, and logging a small set of consistent facts each time: was the brand mentioned at all, was it cited as a source with a link, was the information about the brand accurate, and how did it compare to what was said about named competitors. This is more labor-intensive than pulling a rank report, and there is no way around that yet, because no third-party tool has a fully reliable, standardized method for this across every major assistant.

Some third-party tools have started to offer AI visibility or citation tracking features, and they can help scale the process, but they should be treated as directional rather than authoritative. Because assistants can personalize responses, vary output between identical runs, and restrict or throttle automated querying in ways that differ from how a real user experiences the product, a tool's automated check does not always match what an actual person would see. Cross-checking a tool's findings with a handful of manual spot checks is worth the time before trusting a trend it reports.

Referral traffic is the other half of the picture, and it is more concrete than the query testing. Analytics platforms can identify sessions arriving from AI assistant domains and apps, such as chat.openai.com, perplexity.ai, or Google's AI Overview surfaces where referral data is exposed, and segmenting these out as their own channel, separate from generic search and direct traffic, shows real behavior rather than a simulated query. Watching this segment's volume and its downstream behavior, like conversion rate or pages viewed, over time is a genuine, if still typically small, measurement of AI-driven visibility translating into actual visits.

It helps to be honest about current limitations rather than presenting this as a mature, precise practice. Referral traffic significantly undercounts real AI influence, because a meaningful share of AI-assisted research ends in a decision without ever generating a click back to the source, similar to how a person might read several search results without clicking any of them before acting. That means AI search visibility work should be judged on trend and directional movement, brand mentioned more often, cited more accurately, or appearing alongside stronger competitors, rather than on a single clean number the way a monthly SEO report might present rankings.

The realistic cadence for most businesses is a monthly review: rerun the fixed query set, update the log, review the AI-referral segment in analytics, and note any meaningful shifts, especially after a significant content or schema change on the site. Expecting AI visibility to move at the same pace or with the same clarity as a keyword ranking is not realistic given how young and variable this measurement space still is, and setting that expectation upfront avoids reading too much into normal noise between checks.