How to prove AI Search drives pipeline (when your CRM says "Direct")

Key takeaways

• The path that breaks reporting: a buyer hears your brand in an AI answer, then reaches you through a branded search or a typed URL. Your analytics credits that last step, so the AI touch that drove the decision never gets counted.
• Last-click is structurally blind, not misconfigured. AI-created demand disappears through three doors: copy-paste becomes Direct, branded re-search becomes Organic, and a no-link mention becomes no event at all. No tag fixes this.
• Self-reported attribution is the one signal that survives the broken trail. Asking every new lead "How did you first hear about us?" captures the path no pixel followed, because it asks the only party present for the whole journey.
• You triangulate, you don't trust one number. Self-reported attribution, AI-referral sessions, branded-search lift, and AI Search visibility each prove one thing and miss another. Read together they form a causal chain you can defend.
• Visibility leads, pipeline lags. Being named in AI answers is the input that converts weeks later. Watch pipeline alone and you will conclude AI "isn't working" during the exact period it is building next quarter's demand.

How a lead travels from an AI answer to a wrong CRM label

Here is the literal sequence of events that breaks your reporting.

1. A buyer asks an AI engine a question. "Best [your category] tool for a mid-market B2B team." ChatGPT, Perplexity, Gemini, or Google's AI Overviews returns an answer that names a short list of vendors. You are on it.
2. The buyer forms a brand preference without clicking. Often there is no link to click, or they are on mobile and it is faster to just remember the name. The demand is created here, inside the AI answer, while nothing is tracked, because nothing has happened on your property yet.
3. The buyer goes to Google, or straight to your site, and searches your brand. They type your name. They click your homepage. Now a session starts on your site.
4. Your analytics credits the last touch. A branded Google search lands as Organic Search. A typed URL or a no-referrer context lands as Direct. The AI engine that drove the decision stays invisible.

An inside look into the messy attribution reality of AI Search. People tell the self-reported field "ChatGpt", "chat", "CHATGPT", in every spelling, while the CRM traffic-source columns next to it record Direct, Paid, or Organic. Self-reported and last-click contradict each other row by row, and that contradiction is the AI blind spot you can actually measure.

This is how the attribution model is built, so no tag fixes it. Google's own documentation confirms that last non-direct click is the model behind the standard reports: a conversion is credited to the most recent channel before it, and direct visits are skipped in favour of the last known source where one exists. A session with no detectable referrer gets bucketed as Direct, the channel that means traffic where no source was detected, including typed URLs and untagged links. AI engines sit outside GA4's default channel grouping, and many AI surfaces either strip the referrer or push users into a copy-paste or branded-search motion, so the AI touch never lands as its own channel. The demand-creating event and the credited event end up as two different things, separated by a tab switch your analytics cannot see.

If you want the upstream half of this, how to actually get named in those AI answers in the first place, start with our guide on tracking AI Search visibility with a prompt set.

Why last-click is structurally blind

This looks like a volume problem, the kind where you are under-counting AI by some percentage. It runs deeper than that. AI-driven demand disappears through three different doors, and only one of them leaves a trace.

• Copy-paste or typed URL becomes Direct. The buyer remembers your name and types it. No referrer. Credited to Direct, the channel that means "we don't know."
• Branded re-search becomes Organic Search. The buyer Googles your brand and clicks the organic result. Credited to SEO, work that did not create this demand.
• No-link mention becomes no event at all. The AI named you, the buyer did not act today, and they convert three weeks later on a different device. There is no session to connect to the original answer.

When the answer includes a link and the buyer clicks it, the visit arrives carrying utm_source=chatgpt and lands cleanly as AI-referral traffic in GA4. This is the visible case, door one, and it is your floor.

Here Claude sets a link, but clicking is optional. Many people ask a follow-up question first, or simply remember "Heyflow", leave Claude, and type the name into Google. The demand is created, the click is not, and last-click loses it through doors two and three.

Put the two together and the point lands: sometimes the click is tracked, often it is not, and that difference is exactly why the GA4 AI-referral number is a floor and not a total.

That is why "just look at the AI-referral traffic in GA4" is a trap. Referral sessions from chatgpt.com, perplexity.ai, and similar are real, and they are the floor, not the total, because they only capture door one's rare link-clicks. Most AI-created demand walks through doors that last-click either mislabels or never sees. You cannot subtract your way to the truth from a number that is structurally incomplete.

Self-reported attribution: the one signal that sees the whole path

When the digital trail is broken between the AI answer and the branded search, you need a signal that survives the tab switch. That signal is the buyer's own memory.

Self-reported attribution is the deliberately low-tech practice of asking every new lead, at signup, in the demo-booking form, on the first sales call, "How did you first hear about us?" It captures a path no pixel followed, because it asks the one party who was present for all four steps: the human. This sounds almost too simple, and there are companies selling attribution software that tell you to track every touchpoint and triangulate every signal. There is truth in that. The most honest thing you can still do is ask people where they heard about you, and most teams don't.

osapiens, a B2B SaaS CMMS platform for maintenance and operations teams and a Radyant client, asks "How did you hear about us?" as a required field on their demo-request form, early in the funnel and left open-ended so a buyer can name an AI assistant in their own words.

In our attribution framework this is the primary signal for AI-era demand, because AI Search is exactly the channel that defeats click-based tracking. Last-click tells you where the session came from. Self-reported attribution tells you where the decision came from. Those are increasingly different answers.

A well-designed self-reported field:

• Has an explicit "AI assistant (ChatGPT, Perplexity, Gemini, etc.)" option or stays open-ended. If your only options are Google, social, and referral, you have pre-decided that AI does not exist.
• Is captured as early as possible, on the first form rather than post-sale, while the memory is fresh.
• Is stored as a structured CRM field, auto-categorized, so you can actually report on it rather than burying it in a free-text note nobody queries.
• Is reconciled against your digital data. People misremember, so the power sits in the gap between what they say and what last-click logged.

That gap is the whole game. When 20% of your closed-won deals say "I found you through ChatGPT" and your CRM attributes almost none of them to anything AI-shaped, you have measured your last-click blind spot directly.

The data you can trust, and what each number actually proves

No single source is the truth, so you triangulate. Four signals, each with a clear job and a clear limit.

Read together, these four become one defensible story. You are increasingly named in AI answers, so visibility is rising. Branded search for you climbs, so demand is rising. New leads tell you AI is how they found you, so self-reported confirms it. A verifiable floor of them clicked straight through, so GA4 corroborates it. That is a causal chain you can put in front of a board, not a vanity metric.

What this means for your AI Search strategy

The strategic shift is to stop treating AI Search as a channel you attribute and start treating it as a system you instrument.

• Visibility is the leading indicator. Being named in AI answers is the input. Track it per engine, per prompt, as share of voice, rather than as a vague sense that you show up sometimes.
• Pipeline is the lagging indicator. It moves weeks later, through the broken trail above. Watch pipeline alone and you will conclude AI "isn't working" during the exact period it is building the demand that converts next quarter. This is the "traffic up, pipeline flat" trap, and we cover why traffic stopped being the right SEO metric in a separate guide.
• You win by instrumenting the path between them. Self-reported attribution is the bridge that lets you say "AI visibility went up, and here is the pipeline it produced," instead of hoping the two charts are related.

This is the backbone of how we run organic growth as a Growth Partnership: visibility as the input, self-reported attribution as the proof.

Common implementation mistakes

1. Trusting last-click as the source of truth. It was built for a clickstream world, and AI Search breaks the clickstream. Reading Direct and Organic as "where leads come from" will systematically under-credit the channel creating your demand.
2. Treating GA4 AI-referral sessions as the full number. They are the floor, door one only. Reporting them as "our AI traffic" under-counts reality and makes AI look negligible.
3. Running no self-reported field, or one with no "AI assistant" option, which guarantees you never measure what you never listed.
4. Optimizing for volume instead of citation. More AI traffic is not the goal. Being named as the recommendation is. A thousand sessions from a listicle scrape are worth less than being the one tool ChatGPT recommends by name.
5. Measuring AI as one undifferentiated channel. ChatGPT, Perplexity, Gemini, and AI Overviews behave and convert differently. Roll them into a single "AI" bucket and you lose the ability to act on any of them.

Step by step: set this up in a week

You do not need a new platform to start. You need four instruments wired into the tools you already have.

Day 1 to 2: stand up self-reported attribution.

• Add a "How did you first hear about us?" field to your primary conversion form: demo request, signup, or contact.
• Include an explicit "AI assistant (ChatGPT, Perplexity, Gemini, etc.)" option alongside your usual choices, plus an open-text fallback.
• Map it to a structured CRM field so it is queryable, and make it required where you can without hurting conversion.

Day 2 to 3: capture the AI-referral floor in GA4.

• Create a custom channel group, or a regex segment, that classifies sessions from known AI sources (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and similar) as "AI Search" instead of letting them fall into Referral or Direct.
• Label this internally as your floor metric. It will always under-count, and that is expected.

Day 3 to 4: baseline branded search in Search Console.

• Pull your branded-query clicks and impressions in GSC and set a baseline. Rising branded search is your demand proxy. When AI visibility climbs, this is where you see the echo before pipeline moves.

Day 4 to 5: track AI Search visibility (share of voice).

• Stand up tracking of whether you are being named in AI answers for your category's key prompts, across engines. This is the leading indicator the rest of the model hangs off. Our guide on tracking AI Search visibility with a prompt set walks through how to build one, and the principle matters more than the tool: measure citation, per engine, over time.

Day 5: build the one dashboard that ties them together.

• Put the four on one view: AI Search visibility (leading), branded-search lift (demand), self-reported "AI" leads (decision), GA4 AI-referral floor (verified clicks). When you can see all four move in sequence, an unprovable hunch becomes a defensible attribution story.

Proof it works: Heyflow

We run this exact model with Heyflow, a funnel and form builder for B2B growth teams, and it is the clearest example of why the measurement matters as much as the work.

On the growth side, the program is broad. We pair strategic content production and optimization on high-intent topics with scalable programmatic clusters and dedicated AI Search initiatives:

• High-intent content targeting the searches buyers actually run near a decision, on topics like "The best solution for collecting payments in your quiz funnel", "The 7 best form builders for Webflow 2026", and "How to automate and scale lead generation at your agency."
• Programmatic clusters: scalable families of pages built from one repeatable template against a structured set of queries, so a single proven page shape covers many variants at once. We point them at "jobs to be done" intents and patterns like "best funnel builder for ", which compounds visibility and lead growth across a whole category rather than one keyword at a time.
• AI Search initiatives: we turned Heyflow's existing YouTube library into AI-citable source material without producing new uploads, which produced 593 brand mentions in three months and drove tracked AI Search visibility from 7% to 52% across the prompt set. Read the whole case study.

Here is the part that matters. Alongside the AI Search visibility score, which climbed from 8% to over 50% after 6 months of collaboration, we always read first-click, last-click, and self-reported attribution together. From the start of the engagement we set up a single HubSpot dashboard that brings all organic leads into one view: leads where last-click attribution is organic search or an AI referral, and leads who named something like "Google", "search", or "ChatGPT" in the self-reported field. We also added a sales-reported attribution field, because the sales team often uncovers the real source by simply asking on a call, and that answer was not being captured anywhere before.

That setup is the point of the whole guide in practice. Because we watched the visibility score, last-click, self-reported, and sales-reported attribution side by side from day one, we could see how the share of inbound coming through AI tools changed over time. That is the only basis on which you can make strategically correct claims and adjust the strategy with confidence, rather than guessing which chart caused which.

With our client Heyflow we're taking it one step further:

Heyflow filters AI-sourced leads into a sheet and records which assistant each one named: ChatGPT, Claude, or Gemini. Asking which model, not just "an AI", lets you build a strategy per engine instead of one blurry "AI" bucket, because the engines cite and convert differently.

Conclusion: from "the CRM says Direct" to a story you can defend

A buyer heard your brand in an AI answer, searched it on Google, and booked a demo your CRM then filed under Direct.

The fix is not a better last-click model, because last-click was never built to see a tab switch. The fix is to instrument the path: ask every lead where they first heard about you, capture the verifiable floor of AI clicks in GA4, baseline your branded search, and track whether you are being named in AI answers at all. Read those four together and the AI touch that last-click hides becomes a causal chain you can put in front of a board.

Visibility leads, pipeline lags, and the work in between is measurement. Set it up once, the way we did with Heyflow, and "we think AI is working" becomes "here is the pipeline it produced, and here is how we know."

FAQ