You cannot fully know. A system that reports it precisely is reporting the last thing that happened, which in B2B is usually a branded search by somebody who decided weeks earlier somewhere you were not watching.
That is not an argument for giving up on measurement. It is an argument for replacing one unanswerable question with two answerable ones, which is most of what this post is about.
Across nearly 300 posts on this site, attribution has appeared in zero titles, which is a fair reflection of how much the topic gets discussed relative to how much budget it moves.
What the recorded source usually is
| Recorded as | What probably happened |
|---|---|
| Direct | A colleague sent them the link in Slack, or they typed the name in |
| Organic, branded search | They already knew the name. Something else supplied it |
| Organic, non-branded | The one case where the recorded source is probably right |
| Paid, last click | Possibly the cause. Possibly the final click of six months of research |
| Referral, none recorded | Dark social, a private community, a podcast, an assistant's answer |
Only the third row is reliable, and it is the row that has shrunk most: with 82% of B2B queries returning an AI Overview and zero-click at 75.1% of tracked searches in 2026, the non-branded discovery visit is exactly the thing that now happens without a visit.
The three invisible channels
Word of mouth. The single most common real answer in B2B, and structurally unrecordable. Somebody asked a peer, the peer named you, the peer's own discovery of you happened two years ago through something nobody tracked.
Dark social. A link pasted into Slack, WhatsApp or a private community carries no referrer. The visit arrives as direct, so the channel that caused it is credited with nothing and defunded at the next budget round on the evidence of its own invisibility.
AI assistants. A buyer asks for a shortlist, gets three names, and arrives already decided. First Page Sage's October 2026 report puts 12.1% of tracked sessions as arriving from AI platforms, and that is only the share that arrives with a traceable referrer. The ones who read the answer and then typed your name in appear as direct.
That third one is new enough that most attribution models have no row for it at all, which is covered in how buyers shortlist you in ChatGPT.
The damage, which is not just a wrong report
An attribution model that cannot see a channel reports that channel as converting at zero. The next planning cycle cuts it, the decision looks data driven, and it was based on a measurement gap.
This is the same mechanism described in what is an MQL worth if sales never calls it, where a channel whose leads nobody called shows a zero conversion rate and gets killed on the evidence. Here the cause is different and the outcome is identical: the thing you could not see became the thing you stopped funding.
The two questions that are actually answerable
Ask them. Put one open field on the form and in the first discovery call: how did you first hear about us. Record the answer in their words, not into a dropdown of your channels. It is self-reported, it is biased towards whatever they remember, and it is still the most accurate source of truth available, because the buyer is the only party who was present for the whole evaluation.
Run it for two quarters and the verbatim answers will name things no analytics tool reported. That gap is the finding.
Watch who arrives before the form. The research phase is where the decision gets made and it produces no record in most systems. Company level identification closes part of that gap: it names the organisation behind a visit, which is a signal about the evaluation rather than about the final click. That is the subject of a company read three pricing pages and nobody called them, and in Empiraa Signal the identified company is matched to the account you already hold and the owner is alerted, which is the part that makes it an action rather than a dashboard.
What to keep doing
Keep campaign parameters on paid activity. The link between a click and a spend is direct there, and it is the one place the measurement earns its confidence.
Keep a single source field on the record, and treat it as a label rather than a finding. It is useful for grouping and useless as proof.
Stop reporting a percentage split of pipeline by channel to two decimal places. The precision is fabricated and it is the format that gets a channel cut.
The honest position
Anybody selling you complete B2B attribution is selling something that cannot exist, because the decisive conversations happen in rooms and threads that no tag can reach. That is not a product gap, it is the shape of the problem.
What is defensible is a smaller claim: know what you spent, ask buyers what they remember, watch which organisations show up before they identify themselves, and judge the whole thing on pipeline created against the same period last year. That set of measures is less satisfying than a pie chart and considerably more likely to be true.
Related reading: we published forty posts and rank for none of them, and our organic traffic is down 30% and the pipeline is fine.
Signal includes company identification on every plan, with unlimited users so everyone who should see an alert can.
