Three signals that are routinely counted as one
Most attribution confusion starts with a category error. Being named in an answer, having one of your pages cited as a source, and being put forward as the recommended option are three different outcomes, and only the second produces a clickable source. Collapsing them into a single 'AI traffic' number makes a weak signal look like a strong one, which then distorts the budget conversation.
Working definitions make the difference explicit. Mentioned: the brand name appears in the answer text, with no link to your site. Cited: the answer attributes a statement to a specific URL on your domain. Recommended: the answer positions the brand as the option to choose, which may or may not include a citation. Each one maps to a different business question, and each one needs a different marker in your analytics.
- Mentioned without a citation has no traffic to attribute, but it is a leading indicator of whether the entity is being retrieved at all. The brand check reports mentions separately from citations, which is the distinction this section depends on.
- Cited is the only one of the three that produces a referrer you can see in your logs. The citation report is where the cited URL set is produced.
- Recommended without a citation is the hardest to measure and the most commercially meaningful, which is why it needs its own survey or self-reported field
What is actually observable, and what is not
Attribution has to be built on signals that exist. Referrer headers and landing-page queries are the strongest. Self-reported 'how did you hear about us' fields capture recommended-but-uncited cases, imperfectly but usefully when the response mentions your brand name in a way a buyer would recognise. Everything else is inference and should be labelled as such.
- Track referrers from AI surfaces separately from search referrers, because the referrer string is often the assistant's domain rather than the page
- Use a landing page or query parameter that identifies the AI entry point, so the session is attributable without relying on referrer parsing alone
- Add a self-reported source option that names the assistant, and report it as a separate line from referrer data
- Treat 'dark social' style increments, where AI influence cannot be separated from other channels, as a directional signal rather than a countable lead
Label the lead with what the answer actually did
Attribution is not finished when a session is identified as AI-origin. The useful label is the one that tells the next person what to do. A lead that arrived because your page was cited is evidence that citation coverage is producing commercial value. A lead that arrived because you were recommended, with no citation, is evidence that entity recognition is working and that content coverage needs attention.
- Record the entry surface and the signal type separately: one answers 'where from', the other answers 'because of what'
- Keep the two in different columns in the CRM so the conversion rate of each can be computed without re-deriving anything
- Review the mix quarterly, because a rising mentioned-but-not-cited share is a signal that content structure is the bottleneck
- Do not merge AI-origin leads into a general 'digital' bucket; that is where the distinction stops being useful
What not to conclude from the numbers
Some readings of AI-origin lead data are tempting and wrong. A high citation rate does not mean AI is driving revenue, because citations are also produced in comparisons where your brand loses. A rising mention rate does not mean share of recommendation is rising. And a single month of change in AI-origin sessions is not attributable to a content change unless the question set, platforms and time window were fixed.
- Measure citation coverage and recommendation position separately; a comparison citation is not a win. The recommendation check reports position separately from citation coverage for exactly this reason.
- Report AI-origin lead conversion alongside total lead conversion, so the mix effect is visible
- Use the same fixed question set when comparing periods, or state plainly that the comparison is not controlled
- Keep the answer text for every recorded case, since it is the only way to check the label was correct
The point of separating these signals is not analytical neatness. It is that each one points at a different part of the work, and a blended number points at nothing. Sources reviewed on 2026-10-20. This article records a method with verifiable boundaries; it does not claim that any search or AI platform will index, cite or recommend anything.