Tracking the evidence that is not on your site.

How do you track whether AI systems cite a business?

An owned page saying a firm is the leading specialist in its field is a claim. A trade body listing it, a journalist quoting it, a supplier naming it on a partner page, a court record, a conference program: those are evidence. Retrieval systems weigh them differently for the same reason a person would.

Which means the majority of what matters here is off-site, and the tracking job is different from anything that can be done by crawling your own domain.

What is observable

• Whether a known source still says it. Pages get restructured, directories get pruned, an association renews its member list annually and drops anyone who lapsed. A citation that existed in March is not automatically there in September, and nothing tells you when it goes.
• Whether new independent references appeared, and on what kind of source. Ten low-quality repeats of a press release are not the same evidence as one substantive mention on a source that is itself cited.
• Whether the reference is correct. A mention carrying the old company name, a dead URL or the wrong location is a citation working against the business, because it adds a competing version of the entity.
• Whether a public answer quoting the category names the business at all, recorded as a dated observation rather than a position.

What is not observable, and why saying so matters

Nobody outside the platform can see which specific documents an assistant retrieved for a given answer. Any product claiming to show you "the sources ChatGPT used for your brand" across the board is inferring, and the inference may be reasonable, but it is not a measurement and should never be trended as one.

There is a second, more mundane limit worth stating plainly, because it changed what is possible for everyone: the Bing Search APIs were retired in August 2025, and the public search page returns deliberately misleading results to automated clients. We measured that directly: a headless browser asking for our own brand received ten real-looking organic results, all of them about time zones and weather in Japanese cities, with an HTTP 200 and a plausible result count. Nothing errored.

So a source we could once check cheaply became a source that lies confidently to a script. Any tracker still reporting Bing positions from automation is reporting that. We do not scrape it, and where we cannot measure something the row reads not observable rather than zero.

What to do with the finding

Corroboration is the slowest of the six signal families to move and the hardest to fake, which is precisely why it compounds. The work is unglamorous: be genuinely worth referencing, then make it easy for the people who might reference you to get the details right. Every correction to a wrong mention is worth more than a new weak one, because it removes a competing version of the entity rather than adding a vote.