The Dark AI Layer

A VP of operations at a 200 person company spends a week choosing a vendor in your category. Not on Google. In ChatGPT: “best tool for a team our size,” “how does A compare to B on security,” “is C worth the price.” By Friday she has a shortlist of three and a clear favourite. Three weeks later she types your name straight into her browser, lands on your pricing page and books a demo. Your CRM records the origin as “Direct.” The week that actually decided the deal appears nowhere in it.

Multiply that by every buyer in your pipeline and you have the measurement problem of the decade. It is tempting to treat it as a tracking bug, a tag you forgot to set, a channel you haven’t mapped yet. It is neither. It is structural, and understanding exactly why is the difference between an attribution project and an honest strategy.

First, dissolve the word “touch.”

Attribution - every flavour of it, from first-touch to last-touch to multi-touch, W-shaped, U-shaped, time-decay - is a theory of observable touches. Its entire intellectual scaffolding rests on one assumption: that the journey leaves logs. That a referrer header names the source. That a UTM string you appended survives the click. That a cookie stitches sessions into a person. That a session lands in GA4 at all. Those four instruments are the complete sensory apparatus of your marketing stack.

An AI answer trips none of them. There is no referrer, because the buyer never clicked a link off a results page, she read a synthesised paragraph. There is no UTM, because you cannot tag a conversation you were not part of. There is no cookie, because no page of yours ever loaded to set one. There is no session, because the session happened on openai.com, not on a property you own. The touch was not merely untracked. It was structurally untrackable, it occurred on a surface you have no presence on, inside an interface built to answer, not to refer.

Two ways the answer forms, both dark

A model names your brand through one of two routes, and neither leaves a trail. In the parametric route, the answer comes from frozen training-corpus consensus: the model has read your category across years of reviews, forums and documentation, and it recommends you from memory. Nothing was retrieved, so there is no source, no link, nothing to trace. In the grounded route, the engine retrieves live pages and may cite them and a citation is the closest thing to a trace that exists. But the buyer reads the synthesis and rarely clicks the footnote, so even a citation you earned usually produces no visit. Either way, the influence lands and the log stays empty.

The blind spot has a measurable size, and it predates AI

This is not a hypothetical. Even before AI, “dark social” already broke attribution, and one study put a number on it. Refine Labs ran a hybrid-attribution analysis across 620 declared-intent conversions and $21.5M in closed-won ARR over twelve months. Podcasts drove 53% of revenue ($11.4M) by what buyers said. Software-based attribution credited podcasts with 0%. Across the full study, the tools understated customer-reported revenue by roughly 90%.

That gap opened on human dark social, one buyer telling another in a Slack, a podcast, a private community. Generative AI does not add a new channel to that gap; it industrialises the gap’s mechanism. Where dark social was word of mouth at human scale, an LLM compresses an entire category’s consensus into a single confident paragraph and delivers it millions of times a day.

Why this is the decade’s problem, not last decade’s

The scale is no longer arguable. 94% of B2B buyers now use generative AI in their buying process, and twice as many name generative AI or conversational search as a more meaningful source of information than any other: ahead of vendor websites, product experts and sales (Forrester, Buyers’ Journey Survey 2025). Buyers are nearly 70% of the way through the journey - and have set 85% of their requirements - before they ever contact a seller (6sense, 2024). And 92% start with a vendor already in mind; 41% with a single preferred vendor before formal evaluation begins. Forrester’s phrase for what remains is exact: B2B buying is now “a process of confirmation, not selection.”

Stack those and the conclusion is blunt. The consideration set (the thing that decides the deal) is now formed, overwhelmingly, in the one place your analytics cannot see. The teams still arguing whether the demo request deserves 40% or 50% of the credit are optimising a layer the buyer had already left.

Attribution instruments the click. The decision now happens in a conversation with no click to instrument.

You can’t instrument it. You can sample it.

You will never get a referrer back from ChatGPT. So stop trying to reconstruct the touch, and start measuring the thing the touch was only ever a proxy for: whether you were in the answer. That is a different discipline (call it AI-discovery measurement) built on three questions attribution cannot ask. Visibility: do you appear in AI answers to the category, problem- and comparison-stage prompts your buyers actually type? Accuracy: when models describe you, are the price, the positioning and the differentiator right, or is a hallucination quietly shaping deals you will never see? Influence: how do you show up against competitors in AI-mediated comparisons: your share of the answer, the new share of voice?

None of these is a touch to attribute; each is a rate to sample. And because the same prompt yields different answers run to run (sampling temperature, non-deterministic retrieval, index churn) the only honest measurement is repeated, structured sampling: the same prompt set replayed across engines, surfaces and markets, again and again, reading share of voice out of the distribution rather than off a single screenshot.

The playbook

●      Stop instrumenting the click.  Referrers, UTMs and cookies measure a surface the decision has already left. No attribution model, however weighted, recovers a touch that was never logged. Retire the debate, not the last-touch column.

●      Measure presence in the answer.  Visibility, accuracy and influence in AI responses are the new top-of-funnel metrics. Unlike attribution, they look forward: they describe conversations happening now, not credit for deals already closed.

●      Sample, don’t screenshot.  Replay the same prompt sets continuously across ChatGPT, Gemini, AI Overviews, Perplexity and Copilot, across markets, and manage citation share as a trend you steer, the way you once managed rankings.

The dark AI layer is not a gap you close with better tags. It is a new surface to be present on and a new distribution to be measured. The brands that win the next funnel are the ones that stop asking “which touch gets the credit” and start asking the only question that now decides the deal: were we even in the conversation?

That sampling loop is exactly what we built AYAN to run: the same prompt sets replayed continuously across every major engine, surface and market, so citation share becomes a trend you manage rather than a screenshot you argue about, and so the two machines behind every AI answer, the frozen one and the live one, agree that your brand is the answer.

References

Forrester, Buyers’ Journey Survey (2025) / The State of Business Buying (2026) — 94% of B2B buyers use generative AI in the buying process; twice as many name generative AI or conversational search as a more meaningful source than any other.

6sense, 2024 Buyer Experience Report — B2B buyers are nearly 70% through the purchasing process, and have largely set 85% of their requirements, before engaging a seller.

Forrester, Buyers’ Journey Survey (2024) — 92% of buyers begin with a vendor in mind and 41% with a single preferred vendor; B2B buying is “a process of confirmation, not selection.”

Refine Labs, Hybrid Attribution study — 620 declared-intent conversions, $21.5M closed-won ARR, 12 months; podcasts drove 53% of revenue self-reported vs. 0% software-attributed; ~90% overall measurement gap.

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande (2024), “GEO: Generative Engine Optimization,” KDD ’24 — adding quotations, statistics and citations improved generative-answer visibility by up to ~40% across 10,000 queries.