Attribution when the buyer never clicks
A growing share of B2B research now happens inside an AI assistant. The buyer asks a question, reads an answer that mentions you, and then types your brand into Google or goes straight to your homepage. Last-click reporting files that as branded search or direct traffic, credits nothing, and starves the work that created the demand.
The click you are counting is not the click that mattered
In our own September 2026 audit we found ChatGPT sending clicks to a Growthym URL that returned 404. An assistant was citing a page that did not exist.
Two things follow. The assistant had formed a view of what we published and was recommending it to people asking real questions. And our analytics logged those arrivals as traffic to a broken page, so the work that earned the citation looked like nothing at all in every report we ran. That is the shape of the problem: demand created in a place you cannot see, landing in a bucket your reporting throws away.
Branded search volume is the earliest honest signal
If assistants and content are working, more people search your name. Track branded query volume in Search Console month over month, separated from non-branded, and treat the trend rather than any single month as the signal.
The limits are real. Branded search also rises after a conference, a podcast appearance, a funding announcement, or a competitor’s outage. You cannot isolate the assistant’s contribution from any of those. It is a directional measure of whether demand exists, not a channel report.
Segment direct and AI referrals, then accept what they hide
Create a segment for referrals from assistant hostnames and watch it as its own line. Watch direct traffic to deep pages too, because nobody types a long service URL from memory.
What this misses: many assistants and in-app browsers strip the referrer entirely, copy and paste carries no source, and a mobile answer read on one device often leads to a visit from another. Your assistant referral number is a floor, never a total. Treat growth in it as understated.
Ask the buyer directly on the form
Self-reported attribution is the cheapest instrument available. One open text field: how did you first hear about us. Open text rather than a dropdown, because a dropdown only returns the options you already believed in, and the answers you need are the ones you did not predict.
Its weakness is memory. People name the most recent touch, forget the podcast from four months ago, and sometimes write something unhelpfully short. Read it in aggregate, and place the field after the qualifying questions so it never blocks a submission. When several answers in a month mention an assistant by name, that is worth more than any dashboard segment.
Run a holdout when the volume allows it
The only method here that establishes causation is turning something off. Hold out a region or a set of accounts, keep everything else constant, and compare pipeline created over a period long enough to cover your sales cycle.
The honest caveat is that most B2B programs lack the volume for a clean read, and a long cycle means the answer arrives two quarters later. Contamination is also normal: buyers move between regions, and brand effects cross the line you drew. Run holdouts on the largest spend line you have, not on everything.
Building a view you can defend
None of these four measures is sufficient alone. Together they form a defensible picture: branded search for whether demand is growing, segmented direct and referral traffic for a floor on assistant-driven visits, self-reported answers for the narrative behind the numbers, and a holdout for the one channel where the spend justifies proving causation.
Growthym builds this as a single reporting layer rather than four disconnected habits, which matters because the disagreements between these measures are where the useful information lives. If you want a read on whether assistants currently cite you at all, and on which URLs they are pointed at, their free AI visibility report covers that ground before you commit to a measurement rebuild.
Start with the form. This week, add one open text field asking how the person first heard about you, make it optional, and open a spreadsheet of the answers. In a quarter you will have something no attribution model can give you: the buyer’s own account of what worked.
