I am Stefan Finch, Founder of Graph Digital. I build go-to-market systems for industrial and financial services B2B firms across the UK and EMEA, and in that work I see the same pattern repeatedly: the accounts your lead score rates as cold are mid-RFP. The accounts it rates as warm are researchers, competitors, and curious contacts who will never buy.
Sales teams call this a "ghost deal." A high-intent account has been researching your category for months, building a shortlist, aligning internal stakeholders, stress-testing suppliers with their peers. They have generated no page views on your site, no form fills, no email opens, no webinar attendance. Your scoring model reads them as cold. Then a full RFP arrives. Your CRM has no record of them. The score never moved.
These are your invisible buyers: high-intent accounts that lead scoring cannot find, not because they are not buying, but because they are researching via channels your model cannot observe.
Content engagement correlates with buying intent: this is the proxy at the foundation of web-activity scoring. In 2026, it is no longer a reliable representation of how your best-fit buyers research decisions. That is the diagnosis this page sets out: why the proxy fails, and why it fails specifically in 2026 for complex B2B.
The cost — what ghost deals reveal about your pipeline
When a lead scoring model consistently misreads high-intent accounts as cold, the consequences accumulate in three directions.
The first is rep time. Sales effort is directed by score. When the highest-scoring accounts in your CRM are content engagers rather than genuine buyers: webinar attendees in the wrong geography, newsletter readers without purchase authority, contacts who open emails out of curiosity. Reps follow the signal. They work what is visible. The accounts that are actually buying, researching via peer referrals, AI assistants, and specialist networks, score zero. They do not get called. They arrive at RFP stage having formed a shortlist without your team's input, and often without your company on it.
The second is demand generation. The entire discipline of demand generation depends on a premise: buying intent is accumulating somewhere in the market, and if you can see where, you can direct investment toward it. When your scoring model cannot observe the research channels your best accounts use, that premise fails. You are not directing resources toward intent. You are directing them toward trackable activity, which is an increasingly small and unrepresentative share of the actual buying landscape.
The third is pipeline forecasting. A ghost deal, an RFP from an account with no prior engagement record, is definitionally unforecastable. Every quarter this pattern runs, the gap between what your CRM shows and what is actually happening in the market widens. Buying committees of six to ten people — Gartner's research on the B2B buying journey finds the typical complex-deal buying group involves six to ten decision-makers — conduct most of their decision-relevant research without touching your tracking layer at all.
Your scoring model is not measuring buying intent. It is measuring a channel your best buyers have already left.
Why lead scoring fails — the proxy problem
Lead scoring is built on a theory, not just a configuration. The theory runs like this: content engagement (email opens, page views, webinar attendance, form fills) correlates with buying intent. A contact who visits your pricing page three times looks like a buyer. One who downloads your case study and registers for a webinar looks more so. Assign weights, sum the score, rank the contacts, route the highest-scoring ones to sales.
The proxy was always approximate. Engagement measures content consumption, not commercial intent. A contact opening your emails may be a competitor running a channel audit. A contact who has never touched your content may be three weeks from issuing a contract. The model assumes that, in aggregate, engagement correlates with intent across a large enough contact base, and for certain buyer types, in certain periods of buying history, that assumption was defensible.
Two things were always wrong with it, before AI-era buying made the failure acute.
Engagement belongs to a person. In those same six-to-ten-person buying committees, a single contact's engagement tells you almost nothing about organisational buying readiness. The contact who fills in your webinar registration form may have zero influence on the contract decision. The director who signs the contract may never appear in your CRM at all.
Timing does not map. In long-cycle industrial and financial services B2B, the moment a contact engages with your content and the moment the account is actually evaluating suppliers are rarely the same. AI-era lead scoring is not the origin of this problem. The proxy was always measuring a proxy.
What changed in 2026 is the scale of the gap. Primary research in complex B2B does not start on your website any more. It starts with AI assistants, peer referrals, and specialist communities. And almost none of those generate a tracking-layer signal.
Lead scoring is not failing because the weights are wrong. It is failing because the inputs are absent.
The AI dark funnel: where research goes to be invisible
Dark social describes buyer research that happens through channels that produce no web analytics data: peer messaging, specialist Slack communities, industry networks, direct referrals. B2B marketers have understood dark social for years. It represents a real, if bounded, share of the buying journey. Marketing attribution models have always underweighted it.
The AI dark funnel is a specific and growing extension of this: buyer research conducted through AI assistants. When a commercial director at a UK industrial firm asks Perplexity which specialist suppliers are worth shortlisting for a long-term contract, that query generates no pixel. When a CFO at a financial services firm asks ChatGPT to compare advisory firms in their segment, no form is filled. When Gemini summarises the market landscape and names the providers worth considering, there is no content engagement event.
Graph Digital coins this construct because it is not yet defined or owned in B2B marketing. The AI dark funnel is not an extension of dark social in degree. It is different in kind. Dark social at least leaves traces: referral traffic, occasional UTM data, social mentions. AI-assistant research leaves nothing. The query does not appear in your analytics. The comparative evaluation does not surface in your attribution model. The intent signal does not fire.
In UK manufacturing and financial services, where buyers operate under GDPR posture and FCA regulatory awareness, this is particularly pronounced. Buyers in these sectors have practical reasons to prefer research channels that do not generate a digital trail, and AI assistants give them exactly that.
The result is a class of high-intent accounts that are invisible to your scoring model not because the model is poorly built, but because the research channel they use was never designed to produce tracking-layer data.
Why recalibrating the model doesn't fix it
The instinct when lead scoring is not working, when lead scoring problems show up as ghost deals repeating quarter on quarter, is to fix the model. Adjust the weights, add more signals, bring in intent data providers: platforms that aggregate third-party behavioural signals from across the web to supplement your first-party tracking data.
Intent data helps with a specific and bounded problem: buyers who are researching online through channels your first-party tracking cannot observe, but which third-party providers happen to capture. A contact visiting a category review site, a comparison directory, or an analyst platform may generate a third-party signal even if they never visit your website. For that problem, intent data adds something real.
It does not solve the AI dark funnel.
When a buyer researches your category through an AI assistant, there is no third-party signal to capture. The query happens inside a conversation interface that generates no standard web event. No aggregator captures it. No intent provider ingests it. The research bypasses your first-party tracking and the entire data layer that intent data depends on.
Adding sophistication to the model produces the same result: more precisely wrong scores. Machine learning models, predictive scoring, and AI-enhanced lead scoring all share the same input problem. The model's sophistication is irrelevant when the inputs that would tell it an account is buying are absent by design of the research channel.
This is why lead scoring problems in complex B2B are not fixable through recalibration. The model is measuring a research channel your best buyers have largely moved on from. The failure is theoretical.
The comparison between web-activity lead scoring and alternative measurement approaches (what changes conceptually, and why account-level measurement addresses the gap) belongs in account scoring vs lead scoring, not on this page.
The rational next step
If you are getting ghost deals, high-intent accounts arriving at RFP stage with no prior engagement record, the first useful question is not "how do we fix the lead score?" It is "what should we be measuring instead, and at what level?"
The prescriptive answer to that question is in how to build an account score reps actually use: the practical construction of an account-level model, the calibration against your actual ICP, and the wiring into rep workflow. That is the build.
Before the build, it is worth mapping what your current scoring model can and cannot see in your live pipeline, and which accounts in your named list are likely researching invisibly right now. That is a diagnosis question rather than a build question. It sits naturally within a go-to-market engineering engagement rather than a tool configuration.
The ghost deal pattern is not random. It is a predictable consequence of measuring buying intent via a proxy that AI-era buyers no longer trigger. Understanding which of your highest-value accounts are currently invisible to your scoring model is the first step before deciding what to build next.
Frequently asked questions
Why does lead scoring fail for complex B2B?
Lead scoring fails because it uses content engagement (email opens, page views, webinar attendance) as a proxy for buying intent. In complex B2B with long buying cycles and buying committees of six to ten people, this proxy was always approximate: engagement belongs to a contact, not an account, and the timing of content consumption rarely maps to the timing of purchase decisions. In 2026, it is also a measurement model that breaks by design. The highest-intent buyers research via AI assistants and peer networks that generate no tracking-layer signal, so they appear cold regardless of where they are in their evaluation.
How does AI-era buying behaviour break web-activity scoring specifically?
B2B buyers now conduct primary research via ChatGPT, Perplexity, Gemini, peer Slack communities, and referral networks. None of these generate a tracking pixel or content engagement event. A high-intent UK account can research your category for several months through AI assistants and peer referrals and produce no signal your scoring model can observe. The score is not miscalibrated. It is measuring a research channel your best buyers no longer use for primary evaluation.
What is the AI dark funnel?
The AI dark funnel is Graph Digital's coined term for the extension of dark social to AI-assistant research behaviour. Dark social has long described buyer research through peer networks and direct messaging, channels that bypass web analytics. The AI dark funnel extends this to AI assistants such as ChatGPT, Perplexity, and Gemini: tools that answer buyer research queries, compare suppliers, and shape purchase shortlists, generating no trackable signal for any first-party or third-party data provider. This construct is not yet defined or owned in B2B marketing. It names a class of research behaviour that existing dark funnel vocabulary does not fully capture.
Is the lead scoring problem fixable by recalibrating the model?
No. Recalibration (adjusting weights, adding intent data signals, or applying machine learning) assumes the tracking layer is capturing the right signal and needs tuning. When the highest-intent buyers research via AI assistants and peer networks that never touch the tracking layer, there is no input to tune. The proxy's inputs are absent by design of the research channel. Adding model sophistication to an absent input produces more precisely wrong scores, not better ones.
What is the commercial consequence of lead scoring failure in complex B2B?
Two consequences compound over time. First, misallocated rep time: reps work high-engagement, low-intent contacts while high-intent accounts research invisibly and arrive mid-RFP without your team's engagement. Second, an unforecastable pipeline: full RFPs from accounts with no prior engagement record are outside any model that observes only the tracking layer. The result is a pipeline with lower quality and reliability than the CRM data suggests, not because the data is inaccurate, but because it cannot see the research that preceded the commercial conversation.
Key takeaways
- Lead scoring uses content engagement as a proxy for buying intent, a proxy that was always approximate in complex B2B and that AI-era research behaviour has made unreliable by design.
- The AI dark funnel (buyer research through AI assistants such as ChatGPT, Perplexity, and Gemini) generates no tracking-layer signal, making high-intent accounts invisible to web-activity scoring models.
- Ghost deals, full RFPs from accounts with no prior engagement record, are not anomalies. They are the predictable output of a scoring model that cannot observe where primary research is happening.
- Intent data providers cannot close this gap: AI-assistant research produces no third-party signals for aggregation.
- Recalibrating a web-activity scoring model does not fix an input that is absent by design. This is a theoretical failure of the measurement model, not a configuration problem.
- The question this page cannot answer is what to measure instead. That belongs at the account level. See how to build an account score reps actually use.