Dark Funnel Whitepaper

B2B Dark Funnel: Measure Signals Without Inventing Attribution

Combine buyer-reported influence, branded search and sampled AI visibility. Keep observed traffic separate from inference, and understand what these signals cannot prove.

v1.1 · October 2026 · Tenten GEO Research

CH.01

What Website Analytics Can and Cannot Observe

A buyer can learn about a company without visiting its site. Website analytics observes recorded visits and events; it does not record every private conversation, recommendation or offline interaction.

Conceptual illustration of human silhouettes inside a violet cloud, surrounded by lights shining toward it — A visual metaphor for incomplete observation, not a measurement of buyer behavior

Forrester’s January 2026 article summarizes a 2025 survey of nearly 18,000 business buyers: 94% reported using AI during purchasing, while also checking information with trusted people. That does not mean all research occurred in LLMs or that every resulting visit loses its referrer.

A private conversation that produces no visit creates no website session to classify. If a visit does occur, its recorded source depends on the available signals and attribution scope. GA4 can recognize AI Assistant referrals; Google AI Overviews and AI Mode traffic is included in Organic Search. Direct traffic has many possible origins.

A useful boundary

Keep unknown source as unknown. Do not relabel Direct, Unassigned or missing observations as AI influence without additional evidence.

CH.02

Define the Scope Before Choosing a Percentage

There is no universal percentage of a buying journey that your tools cannot see. Visibility differs by touchpoint, access, consent, instrument and the question being asked.

The former 70–80% dark-funnel chart did not have a verified matching method. 6sense instead reports that first seller contact occurred at about 61% of the journey in its 2025 B2B sample, versus about 69% in 2024. Contact timing is not a measure of time spent in AI or of how much activity analytics missed.

Buyer activity	Possible context	Available evidence	What remains unknown
Recognizing a need	Work, peers, media or AI conversation	Voluntary interview or observed visit	Unreported influences
Considering vendors	Prior experience, recommendations or research	Buyer-reported list and source records	Complete private research history
Comparing options	Vendor content, reviews, conversations or AI	Comparable questions and permitted records	The effect of each influence on the decision
Validating the choice	Trial, demo, pricing and procurement	Recorded events, CRM and interview evidence	Untracked interactions and counterfactual outcomes

Tenten’s illustrative evidence map; the rows do not assign journey shares or imply complete tracking

A useful boundary

State the unit before reporting a rate: respondents, opportunities, sessions or sampled answers. Do not add percentages from these different populations.

CH.03

Ask Buyers Without Leading Their Answers

Self-reported attribution adds the buyer’s account of an experience. It is useful context, with recall and selection limits, rather than a more accurate replacement for every technical signal.

Begin with a neutral question such as “What helped you find and evaluate us?” Let the buyer describe multiple influences. Follow up on AI only after the unprompted answer, and distinguish discovering the company from researching a company already known.

Step 1

Choose an appropriate collection point

Use an existing interview or optional field if appropriate. Consider effort, privacy and response bias before changing a form. Do not require private chat transcripts.

Step 2

Include different outcomes

Invite buyers from won, lost and undecided opportunities where practical. Use the same neutral wording and record nonresponse, rather than sampling only enthusiastic customers.

Step 3

Preserve the account and its provenance

With appropriate permission, retain the original response, collection date and normalized tags separately. Allow multiple sources and unknowns. Do not overwrite observed acquisition data with a self-reported label.

→Open text can reveal unanticipated influences; structured options can reduce effort. Choose for the task and retain an escape from forced answers.
→Timing can change recall. Record when the question was asked and avoid claiming one moment is always most accurate.
→Track the response rate and recruitment method alongside source shares.
→Self-report can omit, misremember or overstate influences. A trend is comparable only when collection and sample composition remain sufficiently consistent.

CH.04

Use Branded Search as Context

A branded query shows search behavior involving a brand name. It does not reveal what caused the search or which earlier conversation the person had.

Review Search Console clicks and impressions for a documented set of brand terms, keeping geography, device, search type and period comparable. These are impressions of your property, not total market search volume. Anonymized and truncated query data means the query table is not a complete activity log.

Observation	Possible explanations	What to examine
Brand impressions rise	Awareness, ranking, campaigns, seasonality or AI exposure	Comparable dates, position, campaign history and interviews
Brand + review queries rise	Evaluation needs or a reputation event	The query context and buyer accounts
Brand vs. competitor queries rise	Comparison interest from several possible sources	Product changes, campaigns and sampled answers
Branded clicks rise while other clicks stay flat	Different demand or search-result behavior by query group	Impressions, CTR and group definitions separately

Hypotheses for investigation, not rules that assign branded search to AI

A useful boundary

The absence of a known campaign does not establish AI as the remaining cause. Aligned search and visibility trends can support a question to investigate; they do not by themselves establish causation.

CH.05

Measure AI Visibility as Its Own Observation

Sampled AI answers can reveal what a product says in recorded conditions. Whether that observation predicts qualified pipeline needs evidence from your own market and measurement design.

A mention, a citation, a referral and a qualified opportunity are different events with different denominators. Do not assume a fixed sequence or delay between them. An increase in sampled citations may reflect a product update, changed prompts or measurement coverage as well as content changes.

→Choose a relevant prompt sample; 100–300 is a planning option, not a validated minimum.
→Record platform, interface, language, date and response status. Consumer UI and API outputs may differ.
→Separate brand mentions, linked citations and recommendations; compare competitors under the same conditions.
→Keep recognized referral sessions and CRM outcomes separate from sampled answer coverage.
→Investigate a repeatable, material change before reacting; one missing response is not a trend.

A useful boundary

Call visibility a candidate indicator until its relationship to the outcome is demonstrated. It is not proof that pipeline will rise next, or that a revenue change was caused by AI.

CH.06

Combine Evidence Without Double Counting

Put the three evidence types beside one another while preserving their definitions. Agreement is useful context; it does not remove their shared biases or reveal every unobserved interaction.

The same opportunity can have an AI referral, a branded search and a buyer-reported AI influence. Counting all three as separate opportunities inflates the result. Retain a permitted opportunity identifier for deduplication and report observed source, stated influence and analyst interpretation separately.

Evidence type	What it records	Main limitation	Appropriate use
Buyer self-report	An account of remembered influences	Recall, recruitment and nonresponse bias	Context for a permitted opportunity record
Branded search	Recorded property clicks and impressions	Multiple causes and incomplete query visibility	Demand and search-performance investigation
Sampled AI visibility	Mentions, citations and recommendations under set conditions	Limited prompts, changing products and unknown buyer exposure	Content accuracy and visibility investigation

Complementary observations with potentially overlapping biases; no automatic causal attribution

Weeks 1 to 2

Define the question and baseline

Choose the business question, existing evidence and a manageable prompt sample. Record collection rules, privacy requirements and missing data before adding instrumentation.

Weeks 3 to 6

Keep collection consistent

Use the agreed interview wording and cadence. Preserve source provenance, record platform changes and keep CRM outcomes deduplicated. Avoid collecting data with no decision attached.

Weeks 7 to 12

Review decisions and uncertainty

Compare like-for-like periods and document alternative explanations. Report which content questions improved and which outcome claims remain unproven. For incremental impact, use a suitable comparison or experiment with an explicit design.

The next decision

Use the existing evidence to decide what to improve next. The schedule above is a Tenten planning example, not a promise to recover all hidden attribution. Stop collecting when a signal cannot support a useful decision.

Sources & methodology

Updated October 5, 2026. Primary sources: Forrester’s January 21, 2026 summary of its 2025 buyer survey (https://www.forrester.com/blogs/state-of-business-buying-2026/); 6sense’s 2025 Buyer Experience Report (https://6sense.com/science-of-b2b/buyer-experience-report-2025/); Google Analytics direct-traffic and channel definitions (https://support.google.com/analytics/answer/15258820 ; https://support.google.com/analytics/answer/9756891); Search Console query limitations (https://support.google.com/webmasters/answer/17011259). The former Gartner-attributed 70–80% invisible-journey chart was removed because a matching primary method was not established. No universal AI-to-pipeline causal claim is retained. The interviews, evidence maps and suggested schedule are Tenten measurement recommendations. No client data, real conversion lift or reconstruction of private conversations is claimed.

Discuss your AI search priorities

Request a 30-minute benchmark interpretation

Send a booking request. We’ll email you to confirm a time within 24 hours.