Playbook

Last month we pulled 8,307 contacts from a growth-stage B2B SaaS company's HubSpot and read the same funnel two ways.

The first read was the one their dashboard gives anyone who asks: HubSpot's Original Source field, built from referring URLs. It reported 11 leads from AI referrals across the window.

The second read was the intake form. On "How did you hear about us?", 809 buyers named ChatGPT, Claude, Perplexity, Gemini, or just "AI."

Here's the part most attribution content gets wrong, though. The instinct is to declare the form the truth and the dashboard a liar.

It's not that simple. The form has its own biases, and they run in the opposite direction. If you replace blind faith in the dashboard with blind faith in self-reports, you've traded one wrong number for another.

So this article is our full methodology for reading AI search and SEO in long B2B buying cycles. We call it the Two-Witness Rule, after the oldest standard of proof in law: nobody gets convicted on the testimony of a single witness.

Same principle here. No channel gets credit, and no channel gets cut, on one signal's word alone.

The method uses exactly two fields that already live in your HubSpot, treats both as witnesses rather than truth, and produces a number you can put in front of a CFO without getting picked apart.

In the audit above, buyers who explicitly named AI on the form had been machine-tagged as Direct, Organic, and even Paid Search. Zero landed in an AI bucket.

The contradiction hiding in your form data

There's a problem here that most attribution takes never confront.

Branded Google search is partly the downstream echo of an upstream AI conversation. A buyer hears your name from ChatGPT, then Googles the brand to verify.

That buyer writes "Google" on your form. They're reporting AI without knowing it.

You can't call the form the truth and simultaneously claim one of its top answers is secretly something else.

The fix is to stop treating either signal as truth and start treating both as witnesses, then read what their agreement and disagreement tells you. That's the whole model, and the rest of this article is the procedure for weighing the testimony.

Dreamdata's 2026 benchmarks put the average B2B journey at 272 days, 88 touchpoints, 4 channels, and 10 stakeholders before a deal closes.

When a single deal is touched dozens of times across four channels over nine months, "what was the source?" is the wrong question. There is no source. There's a set of channels that each played a part.

A deal can be touched by AI search, Google, and a peer referral. We record what the evidence supports.

Channel totals are allowed to sum past 100 percent on purpose, because that's what a multi-touch journey looks like.

This is a harder position to attack than a precise per-lead number, because it concedes the uncertainty up front and still produces a clear, trendable answer.

Every contact gets read against both witnesses and classified into one of five cases. Select a tier to see how it resolves.

The conflict rule

This is where the thinking lives. When the two witnesses disagree, we resolve by asking which one was in a position to see the event.

And when HubSpot says "Paid Search" but the buyer says "AI," AI gets the influence credit.

That paid tag can itself be a downstream echo: the buyer searched a brand they first learned from an AI answer, clicked a brand ad, and HubSpot logged the click. The paid click was the capture step, not the cause.

Why rules instead of an algorithm

One more design choice worth defending.

Some platforms distribute credit with data-driven models, Shapley values and Markov chains. We use rules instead, deliberately.

At the volumes most B2B SaaS companies run (tens of customers, modest lead counts), an algorithmic model overfits thin data and produces a number nobody can explain to a CFO.

Self-report is incomplete. Some share of your inbound never answers the question, and the honest way to handle that is a correction.

The naive version divides observed self-reports by the form completion rate and calls it a day.

We don't report that single point, because it rests on an assumption you can't verify: that the buyers who skipped the form have the same channel mix as the buyers who answered.

Instead, we report a band built from three explicit scenarios.

Then one formatting rule, enforced on every deliverable: every gap ships as a pair, raw and corrected, both labeled, every time.

The raw ratio in this funnel is 11 reported versus 809 self-reported. The corrected figure is a stated range with its base case, never a single false-precise point.

Nobody comparing two of our reports ever finds two different definitions of the same multiplier.

And the Dark bucket stays named. Reporting a bucket labelled "we cannot see this" is the single most credibility-building move in the whole report.

Classification is plumbing. The finding is what made this audit worth publishing.

Read through the model; this funnel's recovered organic pipeline rose 106 percent, comparing the back half of the window to the front half.

Recovered AI search pipeline rose 237 percent in the same comparison, the steepest lift of any channel. Daily lead pace ran at 2.3 times the prior period.

Direct stayed flat. That's the tell.

If the lift had been generic brand awareness, Direct would have moved with it. Instead the growth concentrated entirely in the two channels the dashboard underreports, and they rose together, not at each other's expense.

The defensible read: organic and AI are sequential surfaces in the same buyer journey.

The same content that earns rankings gets cited by assistants, and the same AI mentions come back as branded search. The dashboard files most of that story under Direct and calls it a mystery.

One quality note on those recovered buyers, from a parallel engagement running the same model: buyers who named AI search qualified to MQL or beyond at 89 percent, roughly double the inbound baseline.

Small channel, highest intent in the funnel. The buyer arrives having already received a recommendation.

They're not browsing. They're validating.

If you've read attribution content from paid media agencies, you've seen the triangulation model: CRM data, platform data, self-reports, and you act where all three agree.

It's good thinking, and it doesn't survive contact with organic unchanged, for two reasons.

The decision layer

The action framework that emerges on the other side has three moves and requires a signal to hold across two paired periods before acting.

1. Scale
   When both per-lead witnesses support it, and the visibility layer is rising.
2. Deprioritize
   When traffic and impressions are present but no pipeline recovery follows. Deprioritize, not delete, because organic compounds.
3. Test
   When the sources split: build out the cluster, watch recovered pipeline and visibility, then decide.

Stating the limits is part of why it holds up in the room.

1. It's not per-lead precision
   After referrer stripping, nobody can hand you a certain source for every lead, and anyone who says otherwise is selling you a dashboard. The method produces corroborated ranges and trends.
2. It's not causal proof
   It shows presence and influence, which are the right inputs for a budget decision. Proving causation takes incrementality testing, which needs volumes most B2B SaaS companies don't have.
3. Self-report stays biased even after correction
   The band adjusts for who answered, not for what they misremember. Open text dampens the familiarity bias a dropdown would create, but it doesn't eliminate recall error.
4. The Dark bucket is an estimate
   Reported as a named range, never blended into the confirmed counts.

Two witnesses already on record in your HubSpot. A five-tier matrix anyone in finance can audit. Conflict rules that respect which witness could see which channel.

Corrections shown as bands, not points. A Dark bucket you name instead of hide.

Rolled up to accounts, because deals close at accounts, and trended over paired periods so one hot month can't write the story.

That's the Two-Witness Rule, and it's a more truthful picture of an 88-touch, multi-stakeholder journey than either signal pretending to be the source.

And it's the difference between defending a channel with a number the CFO can dismantle in one question, and walking in with a picture they can't.

We run AI search and SEO programs exclusively for B2B SaaS companies. If you want a senior operator to audit your category's AI search landscape, map the prompts that matter for your ICP, and hand you a prioritized 90-day plan, let's talk