How do I estimate the traffic AI is sending me that analytics never shows?
You infer it from the demand it creates rather than the sessions it delivers. Someone who reads an answer that names you often searches your brand, types your domain, or simply calls. So watch branded search impressions, direct sessions and inbound calls with no traceable source, and look for movement that is not explained by ad spend, seasonality or press. This is triangulation, and it should be labeled as an estimate every time it is reported.
The chain of events you are trying to catch
An assistant names your company in an answer. The person does not click the citation — many answers do not carry one for every claim. They remember the name, search it later, or dial the number the answer gave them.
Every step of that chain destroys the evidence of where it started. What survives is a branded search, a direct visit, or a call from someone who already knows who you are. Those are the fingerprints, and they are the only ones you get.
Note what this means for the shape of the effect. If AI is working for you, it shows up as demand that looks unusually well-informed and unusually cheap, distributed across channels you did not buy. That is easy to miss and easy to credit to the wrong campaign.
Three indicators, and what confounds each
- Branded search impressions. Confounded by broadcast advertising, vehicle wraps, yard signs, press coverage and word of mouth. Useful only when those are stable.
- Direct sessions. Confounded by bookmarks, email links, app referrers that strip the source, and mistagged campaigns. Watch new users specifically, not total direct.
- Untracked inbound calls. Confounded by referrals and repeat customers. The most valuable of the three for a service business, and the messiest.
The isolation you can actually perform
Hold paid spend flat for a defined window and watch whether the branded-to-non-branded ratio moves anyway. Compare year over year rather than month over month so seasonality cancels. Rule out press and promotions before crediting anything to AI.
The strongest evidence available is qualitative and sits in your phone system. When a caller says they asked an assistant, or repeats a phrase that exists only on your site, that is a traceable chain. Systematically capturing it is a byproduct of analyzing calls rather than a separate project, and the volume of such mentions over time is a more honest trend line than any modeled estimate.
Say the word estimate out loud, every time
The failure mode here is building a confident number nobody can audit. An estimated AI contribution figure that appears in a monthly report will, within two quarters, be treated as a measurement and used to justify a budget.
Report the indicators, report the confounders, and report the direction rather than a magnitude. That is more useful and considerably harder to argue with. It is also consistent with how any credible revenue analysis handles unmatched demand: name the unknown portion instead of quietly assigning it.
Topics: dark traffic · brand lift · estimation · attribution
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.