In 2026 the B2B buyer journey AI research phase happens inside an assistant, not on your website. Buyers ask an LLM to frame the problem, name the vendors worth considering and pressure test the claims, then arrive already shortlisted. Your analytics records that arrival as direct traffic. You cannot recover the missing touches, so you measure them with proxy metrics instead.
Most demand gen and RevOps leads we speak to describe the same symptom: pipeline is holding, and the report that is supposed to explain it has stopped making sense. Direct and unknown source keep expanding, paid and organic keep shrinking, and nobody can name the content that did the work. DevCommX builds the signal layer underneath GTM systems, so we rebuild this measurement often. If your lead definition is creaking under the same pressure, read why the MQL stopped predicting revenue first, then come back for the measurement layer.
The Touch You Cannot See: What Happens Before a Buyer Reaches Your Site
Independent research was always the bulk of the work. Gartner's research on the B2B buying journey found that buying groups spend only 17 percent of their time meeting with potential suppliers, and when several vendors are in contention that drops to roughly 5 or 6 percent each. The share of the decision made without you in the room is not the new part. Where that research happens, and whether it leaves a trace, is.
It used to leave a trace everywhere. A buyer typed a question into a search engine, landed on a blog post, read two comparison pages, opened a review site, then returned through a branded search a week later. Each step produced a session, a referrer and a timestamp. The trail over credited the last click, but it existed, and you could argue about it with data.
Now the same buyer opens an assistant and asks it to explain the category, name the four vendors worth looking at and summarise what customers complain about. That exchange can run twenty minutes and produce zero sessions on any vendor site. 6sense's 2025 Buyer Experience Report found 94 percent of buying groups had ranked a preferred vendor before first contact with a seller, and that they bought from that early favourite 77 percent of the time. The decision that matters is made in a room your analytics cannot enter.
Why Direct Traffic and Unknown Source Are Quietly Growing in Your Reports
Three leaks compound. A large share of assistant answers produce no click at all: the buyer reads the synthesis, notes two brand names and moves on. When a click does happen it often arrives through an in app browser or a rewritten link, so the referrer is missing and the visit is filed under direct. And most commonly the buyer searches your name days later, which converts an AI sourced touch into branded organic that looks like your SEO programme working.
The click loss is measurable across the whole web. SparkToro's 2026 zero click study, built on Similarweb clickstream data, put the share of United States Google searches ending without a click at 68.01 percent across the first four months of 2026, up from 60.45 percent in 2024. Pew Research Center, tracking real browsing across roughly 68,000 searches, found users clicked a result 8 percent of the time when an AI summary was present against 15 percent when it was not.
This is what zero-click B2B marketing actually costs: not the traffic, the visibility into it. The sessions you still get are better ones. Semrush found the average visitor arriving from an AI search source was roughly 4.4 times as valuable as the average organic visitor, because that person has already compared options before landing. Fewer visits, later in the process, worth more each. Good commercially, terrible for reporting built on session counts.
Before rebuilding anything, map which stage lost which signal.
Mapping the B2B Buyer Journey AI Stage by Stage, and Which Stages You Still Control
The AI marketing funnel is not a new funnel. It is the same five stages with the observation point moved to the bottom. Problem framing, vendor discovery, shortlist building and most validation happen in a context you do not host and cannot instrument. Contact is the first stage that produces a record, and by then the buyer has usually chosen who to talk to and what to ask.
What you still control is narrower than it looks and more valuable. You control the corpus: the pages, docs, comparison content and third party mentions an assistant retrieves for your category. You control the shortlist criteria, because the way you frame the trade offs in public becomes the frame the buyer arrives with. And you control the form, the one point where the invisible part of the journey can be made to declare itself.
Being retrievable is a separate discipline from measuring it. Our playbook on how to get cited in AI search covers the six signals that decide whether an engine quotes you. This post assumes that work is under way and asks the next question: how do you prove it moved anything. The two halves relate the way sending and measuring do in signal based selling versus intent data, where a signal is only useful once you can see what it caused.
Three Proxy Metrics for Invisible Research: Branded Search Lift, Self-Reported Attribution and Citation Share
1. Branded search lift. When someone learns your name inside an assistant and does not click, the usual next action is searching for you. That makes branded query volume the closest thing to a demand thermometer you have. Pull branded impressions and clicks from Google Search Console, hold the query set fixed, and track them monthly against unbranded impressions. You are looking for divergence: branded demand rising while unbranded is flat is the fingerprint of research you cannot see. Annotate paid brand campaigns and seasonality on the same chart before anyone reads a trend into it.
2. Self-reported attribution. The oldest measurement method in marketing is the only one that can see a conversation. Asking buyers how they heard about you recovers touches no pixel reaches: a podcast, a private community, an assistant answer, a colleague. HockeyStack Labs built its self-reported attribution benchmark from more than 8,000 responses across dozens of B2B brands, so the method scales past anecdote when the question is asked properly. Most teams ask it in a way that guarantees useless answers.
3. Citation share. Fix a set of thirty to fifty prompts a real buyer would type, covering the problem, the category, the alternatives and the objections. Run them monthly across the engines your buyers use, record which domains are cited and where, and track your share over time. This is the leading indicator upstream of the other two: citation share moves first, branded search follows, self-reported answers follow that.
Keep one direct measurement running alongside the proxies. Build an analytics segment matching assistant referrer hostnames so the clicks that do carry a referrer are counted separately instead of being buried in referral traffic. Treat that number as a floor, never the total, and keep it next to the self-reported count so the gap stays visible to whoever reads the report.
Fixing the Self-Reported Attribution Question, the Highest ROI Form Change You Can Make
Most self-reported attribution fields fail for three reasons. They are optional, so only the enthusiastic answer. They are a dropdown borrowed from ad platform taxonomy, so the buyer picks Google because Google was involved somewhere. And the answer lands in a free text note nobody parses. Each is fixable in an afternoon.
Ask one required open text question worded to request specifics: how did you first hear about us, and please be specific. Open text beats a picklist because the value sits in detail a picklist cannot hold, the name of the podcast, the community, the person or the assistant. Ask on the demo request form rather than a content download, because you want an answer from someone far enough along to remember the path. Asking again on the discovery call, in the rep's own words, catches anyone who typed something thin.
Then make the answer usable. Store it in a dedicated CRM field on the opportunity, not on the contact and not in a notes blob, so pipeline value can be summed by source. Normalise responses weekly into a fixed taxonomy with an explicit AI assistant category and a peer or community category, keeping the raw text alongside. Report pipeline created and closed won by self-reported source next to the same numbers from analytics. The two views will disagree, and that disagreement is the most useful thing on the page. This is ordinary revenue operations plumbing, and it beats another attribution tool.
A Worked Example: Reading One Quarter of B2B Buyer Journey AI Signals
The numbers below are illustrative, invented so you can follow the arithmetic and run it on your own quarter. They are not a DevCommX client result. Assume a company that took 240 demo requests in a quarter, made the self-reported attribution field required, and got usable answers from 146 of them, a 61 percent answer rate.
Normalising those 146 answers gives 38 naming an AI assistant explicitly, 44 saying Google or search, 26 naming a peer, community or podcast, 21 naming LinkedIn and 17 naming an event or something else. Meanwhile analytics, filtered to assistant referrer hostnames, counted 9 sessions from AI sources in the same quarter. Thirty eight self-reported against nine measured is a factor of roughly four. That ratio is the whole argument: the channel is not small, it is unmeasured, and the measured number is a floor.
Two readings finish the picture. Of the 44 who said Google, re asking a sample on discovery calls usually splits them into brand searchers and category searchers, and the first group is downstream of something that named you elsewhere. If branded impressions also rose while unbranded stayed flat, that divergence corroborates the self-reported story instead of repeating it. Three independent measures pointing the same way is as close to proof as this problem allows.
What to Stop Measuring, and What to Report Upward Instead
Stop reporting first touch channel at session level, because the first touch is no longer in your data and a model that assigns it anyway is inventing history. Stop reporting MQL volume as a headline: it measures how many late stage buyers found your form, not how much demand you created. Stop reporting traffic to a blog post as its contribution, when that post may be doing its best work as a source an assistant quotes to someone who never visits.
Report four things instead. Pipeline created by self-reported source, the only view that includes dark funnel B2B touches. Branded demand, trended against unbranded. Citation share on your fixed prompt set. And conversion quality by arrival type, since AI sourced arrivals are later stage and should show a different win rate and cycle length. For leading indicators that survive the shift, see our work on leading indicators and pipeline risk in HubSpot and on marketing to late stage deals.
Then set expectations honestly upward. Some share of pipeline is now unattributable, so say so, quantify the gap and show it shrinking as the answer rate climbs. Teams that pretend to full attribution optimise toward whatever the model can see, usually paid search, and quietly defund the content assistants are reading. That is why we treat measurement as part of the build, a theme running through our guide to agentic marketing and our view of a modern SaaS GTM strategy.
Build Your Dark Funnel Measurement Layer With DevCommX
DevCommX builds autonomous, signal-based AI SDR and GTM systems your team owns outright, and the measurement layer ships with the build: the self-reported attribution field, the CRM taxonomy, the branded demand trend and the citation prompt set, wired so the gap between reported and measured stays where leadership can see it. It is the same instrumentation under the outbound programmes where we have driven 40+ qualified demos in ~6 weeks, because you cannot tune a system whose inputs you cannot see. To get the dark funnel measurement checklist mapped to your stack, book a GTM strategy call.
References
- Gartner, research on the B2B buying journey, supports the 17 percent of buying group time spent with potential suppliers and the 5 to 6 percent per vendor figure.
- 6sense, 2025 B2B Buyer Experience Report, supports the 94 percent of buying groups ranking a preferred vendor before first contact and the 77 percent purchase rate.
- SparkToro, 2026 zero click search study, supports the 68.01 percent of United States Google searches ending without a click, against 60.45 percent in 2024.
- Pew Research Center, users click less when an AI summary appears, supports the 8 percent versus 15 percent click rates across roughly 68,000 searches.
- HockeyStack Labs, self-reported attribution report, supports the benchmark built from more than 8,000 responses across dozens of B2B brands.
- Semrush, study of the impact of AI search on SEO traffic, supports the finding that AI search visitors are around 4.4 times as valuable as organic visitors.
FAQ
How do you track leads from ChatGPT?
Track it two ways at once. Build an analytics segment matching assistant referrer hostnames to catch the clicks that do carry a referrer, then add a required open text self-reported attribution field to your demo form to catch the larger share that arrives with none. Compare the two counts every quarter and report the gap between them rather than either number alone.
What is the dark funnel?
The dark funnel is every buying touch that influences a decision but never reaches your analytics or CRM: an assistant conversation, a private Slack group, a podcast, a peer recommendation. It is not new, but AI research has made it much larger, because whole stages of evaluation now happen in a chat window that sends no referrer and creates no session.
Are MQLs still relevant?
The MQL is still a useful routing object and a poor headline metric. Scoring form fills made sense when a form fill marked the start of research. Buyers now arrive after research, so MQL volume measures how many late stage buyers found your form, not how much demand you created. Keep the routing logic and report pipeline by self-reported source upward instead.
What does the B2B buyer journey AI research stage look like in 2026?
A buyer opens an assistant, describes the problem in their own words, asks which vendors solve it, then probes pricing, integrations and common complaints. That single session can replace five or six website visits. By the time a form is filled the shortlist exists, the objections are formed, and the vendor being contacted is often already the preferred one.
What does zero-click mean for B2B marketing?
It means the answer is delivered without a visit, so traffic stops being a proxy for demand. SparkToro put the share of United States Google searches ending without a click at 68.01 percent across the first four months of 2026. For B2B the goal moves from earning the click to being the source the answer gets built from.
Can attribution software see AI-assisted research?
Only partially. Multi touch platforms can see an assistant referral when the click carries a referrer header, and can stitch it to a later opportunity. They cannot see a conversation that ended without a click, which is the majority case. Treat any tool claiming full visibility into assistant research with suspicion, and pair it with survey based measurement.
.webp)
