A ChatGPT mention is not revenue.

How do you attribute AI search to revenue?

AI search attribution works by connecting identifiable AI-originating website sessions to on-site actions, lead records, opportunities and customer revenue, while separately measuring AI visibility and influence that cannot be directly tied to a click.

A practical attribution model should distinguish between three things:

Those three categories should not be reported as if they are the same thing.

The first measurable click may no longer be the first meaningful interaction.

Consider this journey: A potential customer asks ChatGPT: "Which firms can help us improve marketing attribution in South Africa?" They receive a shortlist. They ask: "Compare the first two." They read the answer. They close ChatGPT. Two days later they Google one of those companies by name. They visit the website. A week later they return directly and submit an enquiry.

What created the demand? Google? Direct? ChatGPT? All three touched the journey. But the analytics platform can only work with the signals available to it. That distinction is becoming increasingly important. The discovery interaction and the measurable acquisition interaction are no longer necessarily the same event.

Recommendation, Visit, Lead, Opportunity, Revenue

1. Recommendation

Before worrying about attribution, establish whether AI systems are actually presenting the brand to potential buyers. Measure: brand mentions; recommendation frequency; recommendation position; competitor presence; answer accuracy; citation sources. This is not revenue attribution. It is discovery measurement. It tells us whether an AI environment is creating an opportunity for acquisition to occur. For the detail on this stage, see How to Measure AEO Performance.

2. Visit

The next question is: Did somebody reach the website from an identifiable AI source? This is where normal acquisition measurement begins. Google Analytics can identify sources that refer visitors to a website through traffic-source dimensions such as source, medium and channel grouping. Google has also added an AI Assistants channel to its default classification for identifiable traffic from services including ChatGPT, Gemini, DeepSeek, Copilot and Grok. Google's AI Overviews and AI Mode are excluded from that AI Assistants channel definition. Useful measures include: sessions; users; landing pages; engagement; key events; conversion rate; source / medium; channel. This gives you observable AI referral traffic. It does not tell you every instance where AI influenced a buyer. That is a different question.

3. Lead

This is where many attribution systems break. Analytics knows where the session came from. The CRM receives: Aubrey Turner / [email protected] / Interested in AEO consulting. And somewhere between the website and CRM, the acquisition information disappears. Now sales has a lead. Marketing has a traffic source. Nobody has the connection between them. That needs to be fixed. Where practical, the lead record should carry fields such as: original acquisition source; original acquisition medium; landing page; campaign where relevant; first identifiable source; latest identifiable source; AI source where identifiable; timestamp; lead identifier. The exact schema depends on the stack. The principle does not. The acquisition information has to survive the form submission.

4. Opportunity

A lead is not yet commercial value. Once the lead enters the CRM, we need to know what happened next. Did sales qualify it? Was there a real requirement? Did it progress? Useful stages might include: Lead, then Marketing qualified, then Sales qualified, then Opportunity, then Proposal, then Won or Lost. Now we can start asking much better questions. Not: How much AI traffic did we get? But: What percentage of identifiable AI-originating leads became qualified opportunities? That is a far more interesting number.

5. Revenue

Eventually we reach the metric marketing cannot fake. Money. When the CRM and analytics architecture are sufficiently connected, we can measure things such as: customers acquired; first-year revenue; contract value; gross profit; recurring revenue; customer acquisition cost; payback; lifetime value. Then AI search can be evaluated alongside the rest of the acquisition mix. Not because every journey is perfectly attributable. It never will be. But because enough of the observable journey has been connected to make a useful commercial decision.

Not every AI-influenced customer belongs in the same bucket.

I would rather report a smaller number I trust than a bigger number I cannot defend.

A useful way to think about AI attribution is confidence.

High confidence: observable AI referral; acquisition source captured; lead identifier preserved; CRM record connected; revenue recorded.

Medium confidence: observable AI referral and conversion; incomplete CRM connection; commercial outcome partly known.

Low confidence: self-reported AI discovery; no digital referral path; or an inferred relationship based on branded search or timing.

All three can be useful. They should not be added together and presented as one perfectly attributed revenue number. That is how attribution turns into fiction.

Analytics is part of the answer, not the entire answer.

GA4 gives us useful information about acquisition. Its traffic-source dimensions identify where users and sessions came from, including source, medium and channel information. Traffic Acquisition is session-scoped, while User Acquisition focuses on how a user was first acquired. That distinction matters. A customer may first arrive via ChatGPT, then return through Google, then convert directly. Different reports can legitimately show different source perspectives. Neither is automatically wrong. They are answering different questions.

GA4 also cannot create source information it never received.

If referral or campaign information is missing, sessions can be classified as direct. Google specifically notes missing traffic-source information, redirects, URL handling and some browser or tracking conditions as reasons traffic may appear as (direct) / (none). So: AI traffic observed in GA4 is not the same as all demand influenced by AI. That equation is one of the most important things to get right.

If source data dies at the form, your attribution dies with it.

Your analytics platform knows about sessions. Your CRM knows about customers. The commercial measurement architecture needs to connect the two. At minimum, I want a lead record to preserve enough acquisition context that we can analyse outcomes later. For example:

Now we can query: How many AI-originating leads did we receive? Which AI sources produce the strongest leads? What pages do they land on? What percentage reaches opportunity stage? What revenue was won? That is a marketing dataset an executive can actually use.

This work sits inside a wider Revenue Measurement Architecture, and the Attribution Detective case study shows what happens when the handoff is fixed.

Do not force one attribution answer onto every business question.

Imagine: ChatGPT, then Organic Search, then Email, then Direct, then Sale. Which source deserves credit? The answer depends on the question.

Original source: useful for "Where did we first acquire this customer?"

Latest source: useful for "What brought them back immediately before the conversion?"

Multi-touch attribution: useful for "Which observable interactions contributed across the journey?"

Self-reported source: useful for "What does the customer remember creating awareness or preference?"

None of these needs to defeat the others. They answer different commercial questions.

Changing the attribution model cannot recover a touchpoint that was never observed.

GA4 supports attribution models for assigning credit to observable touchpoints, including data-driven attribution and last-click options for key-event reporting. That is useful. But attribution modelling happens after data collection. If the buyer interacted with an AI answer and that interaction never produced an identifiable signal in your measurement stack, switching attribution models cannot magically recreate it. This matters because AI search introduces more off-site consideration. A beautiful attribution model applied to incomplete data is still incomplete attribution.

Sometimes the simplest data point is still worth asking for.

For high-consideration businesses, I like supplementing digital attribution with a lightweight question: How did you first hear about us? Do not use this instead of digital measurement. Use it alongside it. Possible answers might include: Google; ChatGPT / AI assistant; LinkedIn; Referral; Event; Podcast; Existing customer; Other. Better still, allow free text when the sales volume makes that practical.

You will get imperfect answers. People forget. People simplify. But if CRM attribution says Google Organic while the lead says "ChatGPT recommended you", that is commercially interesting. It tells us the measurable click and remembered discovery event were different.

AI may create the search rather than receive credit for it.

One pattern I expect marketing teams to pay more attention to is: AI discovery, then branded search. The user learns about a company in an AI answer. Then they search the brand. Analytics sees Google Organic. Search Console sees the branded query. The AI interaction may never appear in either system.

That does not mean every increase in branded search should suddenly be credited to ChatGPT. It means branded-search trends become one useful supporting signal when viewed alongside: AI visibility changes; referral traffic; self-reported discovery; direct traffic; conversion trends; PR activity; other marketing campaigns. Attribution works better when multiple pieces of evidence agree.

Keep AI reporting connected to the funnel.

Do not create a completely separate universe of AI metrics. I would use four sections.

That lets an executive move from Are we appearing? to Is anybody responding? to Are they valuable? to Did we make money?

Imagine this happened over a quarter.

AI visibility testing shows the brand is recommended in 18% to 37% of tracked commercial prompts. Identifiable AI referral traffic grows 90 to 260 sessions. Those sessions generate 21 enquiries. The CRM connects 17 leads successfully to acquisition data. Sales qualifies 8 opportunities. Three close. Total first-year contracted revenue: R540,000.

Now we have something useful. We can say: We observed R540,000 in won revenue connected to identifiable AI-originating acquisition paths. We should NOT automatically say: AEO generated exactly R540,000. Why? Because some of those customers might have encountered other marketing. And some additional customers may have been influenced by AI but arrived through another route.

The useful commercial conclusion is not false precision. It is: AI search is producing observable qualified demand, and we can now evaluate whether further investment is justified.

Revenue without investment is not ROI.

If we eventually want to evaluate AI search ROI, we also need the cost side. That might include: consulting fees; internal marketing time; development work; content production; research; PR / authority activity; software; analytics implementation. Then AI Search ROI = (Attributed gross profit minus AI search investment) divided by AI search investment. I prefer gross profit where it is available because revenue can make low-margin acquisition look much healthier than it actually is. The exact business case will differ by company. But the principle is simple. Do not call something ROI when you have measured only traffic or revenue.

More on this in AI Search ROI: How to Measure Revenue From AEO & GEO.

Most attribution problems are plumbing problems.

Not modelling problems.

I repeatedly see the same breaks.

Neither extreme is useful.

Build the chain once.

A practical architecture may look like this:

Then layer AI visibility data and self-reported discovery over the top. Now you have three useful perspectives: What the AI engines are doing (Visibility). What identifiable visitors are doing (Acquisition). What customers are doing (Commercial outcome). That is the measurement architecture.

This is not really an AI-search problem.

It is a marketing measurement problem exposed by AI search. The same weaknesses that make AI attribution difficult usually affect: SEO; content; PR; organic social; referrals; partnerships; brand; dark social. AI search simply makes the flaw more obvious because so much consideration happens away from your website. If you fix the architecture properly, you do not only improve AEO reporting. You improve the way the business understands acquisition. That is much more valuable.

I do not want to prove AI search worked.

I want to find out whether it did.

That sounds like a small distinction. It is not. If the data shows ChatGPT referrals are tiny but exceptionally valuable, that matters. If visibility doubles and nothing commercially changes, that matters. If Perplexity sends fewer visitors but more qualified opportunities, that matters. If the company is getting recommended constantly but the answers contain the wrong product information, that matters. And if the evidence says the business would generate more return by fixing paid search, conversion rate or CRM leakage before spending another rand on GEO, that matters too. Measurement should help decide what to do next. Not manufacture a success story for the last invoice.

See also: AEO & GEO Consulting and the Making a National Brand the AI Answer case study.