Measuring AI Impact: Why It’s Harder for Marketers

The Journey Your Analytics Will Never See

I look at acquisition reports every week. And lately, I keep seeing the same pattern.

A rise in brand search. « Direct » traffic swelling. No identifiable AI source. And a marketer asking me: « Where is this coming from? »

Neil Patel describes this scene. I see it with my clients. A user opens ChatGPT. They ask for a product recommendation. The AI cites your brand among its answers. They read it, then close the tab. They click on nothing.

A few days later, they type your name into Google. They click through to your site. They buy.

What does your analytics see? An organic brand search. A conversion attributed to Google. Period.

The recommendation that triggered everything exists nowhere in your data. No referrer. No UTM parameter. No cookie. Nothing.

The problem isn’t that AI doesn’t convert. It’s that AI converts somewhere other than where you’re looking.

Why does this difficulty hit marketers harder than other profiles? Because a marketer lives in a dashboard. Their budget is justified by lines: channel, cost, conversion. When part of the journey unfolds in a chat window nobody observes, the line doesn’t exist.

And what has no line gets no budget.

I see two reactions. The first: ignore AI because « it doesn’t show up in the numbers. » The second: panic and buy a tool that promises to track everything. Neither one solves anything.

My conviction is simple. SEO in 2026 is no longer played on the click alone. It’s played on the mention, then on the memory of that mention. And no pixel captures that directly.

In this article, I show you where the journey breaks, why classic models fail, and the three measurement layers I stack to get close to reality. Without promising a precision nobody has.

Why Last Click Always Gives Google the Credit

The mechanism is mechanical. Almost simple-minded.

Most analytics tools attribute the conversion to the last measurable touchpoint. If that touchpoint is a Google search on your brand name, Google wins. Every time.

Now, here’s what an AI journey produces. The recommendation creates the intent. But the intent is then expressed on the search engine. AI sows, Google harvests.

Here’s the diagnosis: my attribution model is blind to the step that matters.

Let’s take the three models marketers use most.

• Last click. It credits the brand search. AI stays invisible.
• First click. It credits the first measured contact. Again, the ChatGPT conversation left no trace, so it credits something else.
• Multi-touch models. They split the credit among known contacts. But an unknown contact receives no share. A precise split on incomplete data is still an error, just better presented.

That’s the trap of sophisticated dashboards. The more elaborate the model, the more confidence it inspires. And the more we forget that its raw material is full of holes.

I observe it with my clients: when we open the attribution discussion, the first question is about the tool. « Which software tracks AI? » The right question comes earlier. « What share of the journey happens outside my data? »

For a chat recommendation, the answer is: everything upstream. Discovery, comparison, sometimes the decision. All that remains in your data is the final formality.

A report that says « 100% of sales come from brand search » doesn’t describe the customer. It describes the place where the customer finally left a trace.

This finding has a direct consequence for your budget decisions. If you cut what doesn’t show up, you cut what feeds brand search. You’ll find out a few months later, when that search runs out of steam, without being able to explain why.

So before choosing a tool, let’s look at exactly where the journey breaks. There are three distinct leaks. Each one calls for a different answer.

The Three Leaks That Break the Journey

When I audit an AI-influenced journey, I look for three breaks. They’re independent. They add up.

Leak 1: the recommendation without a click

The AI answers. The user reads. They have no reason to click, since the answer is enough for them. Your brand was exposed, but your server recorded nothing.

No visit. No event. No impression counted in your tools.

This is the biggest leak. And the most overlooked, because you can’t measure what never hits your site.

Leak 2: the delay

The scénario Neil Patel describes involves several days between the recommendation and the Google search. That’s the heart of the matter.

Several days is more than enough to fall outside any reasonable attribution window. The longer the buying cycle, the wider the leak. An impulse purchase at $30 leaves little time for forgetting. A B2B contract or a piece of equipment costing several thousand dollars leaves a lot.

The marketer who compares sales against a short window concludes that AI weighs nothing. They simply shortened the window until they could no longer see it.

Leak 3: the channel switch

The recommendation is born in one tool. The conversion happens in another. Sometimes on another device. A journey started on a phone in a chat app, finished the next day on a desktop computer.

For your analytics, those are two different people. One doesn’t exist. The other arrives out of nowhere.

This breakdown changes the conversation with leadership. Instead of saying « we don’t know how to measure AI, » you say « we lose information at three specific points, and here’s how we compensate for each. » That’s a position of strength.

It’s far from trivial.

An « we don’t know » leads to « then we don’t fund it. » A « we know where we lose, and we estimate the loss another way » leads to a discussion about method.

Now let’s see why this challenge weighs more heavily on marketers than on other profiles.

Why Marketers Are More Exposed Than Others

You might think marketers, seasoned in measurement tools, are the best equipped. In my view, it’s the opposite.

Their strength is also their weakness. They’ve built entire practices around traceability. Every euro spent must be traceable to a line in a report. That discipline is valuable on paid channels. It becomes a brake when the channel leaves no trace.

It’s not the tooling. It’s the question of what counts.

I see three structural reasons.

• Budget follows data. A marketer has to defend their lines in front of a finance director. A channel that produces no attributable number is hard to defend, even if it’s genuinely working.
• Metrics are inherited. Reports were designed for journeys with a click at every step. Conversational AI doesn’t honor that contract.
• Time is short. One quarter to prove a result is comfortable for a paid campaign. It’s very short for a awareness effect that shows up through brand search, several days or weeks later.

An SEO doesn’t escape the problem, but approaches it differently. They think in visibility, queries, pages. They know part of their work acts upstream of conversions.

The marketer, on the other hand, answers for a result. And discovers that the result depends on a layer they can’t see.

That’s why the challenge is harder for them. Not for lack of skill, but because they’re judged on the part of the journey that AI empties of its traces.

I don’t have a perfect solution to offer you. Nobody does, whatever some tool vendors claim. What I have is a method to reduce the blindness. It rests on the idea that a single measurement is no longer enough. You need to overlay several, each imperfect, that cross-check each other.

That’s what I detail in the next section.

The Three Measurement Layers I Stack

No single layer tells the truth. Together, they draw a usable picture.

Layer 1: test queries by hand

Start at the source. List the questions your customers actually ask before buying. Then ask them yourself to the AI assistants. Regularly. Note: is your brand cited? In what position? With what description? Who appears alongside?

It’s artisanal. It works. It measures your presence in the recommendation, the one step your site doesn’t see.

I log the answers in a dated table. An assistant’s answers vary from one session to the next, so I never draw a conclusion from a single test. I look for trends across multiple readings.

Layer 2: track brand signals

If AI sows and Google harvests, brand search should move. So I track the volume of queries containing the brand name, direct access, and queries like « brand + reviews » or « brand + price. »

A movement in these curves after a change of visibility in assistants proves nothing on its own. But combined with layer 1, it becomes a serious clue. When both rise together, the hypothesis holds.

Layer 3: ask directly

The most underrated. A free-text field in the contact form or at order confirmation: « How did you hear about us? » Not a dropdown that forces a choice. A real open question.

What people write is often valuable. When some answer that they asked an assistant for advice, you get what your analytics will never give you: the admission of the journey.

It’s self-reported. So it’s imperfect. People forget, simplify, get it wrong. But that imperfection is different in nature from the one in your reports, which makes the cross-check useful.

Three layers. None is enough alone. All three together, yes.

Three Use Cases: E-commerce, B2B, Local

The same problem takes three faces depending on the business model.

E-commerce: the product recommendation

This is Neil Patel’s scénario. A user asks for a product, the AI suggests brands, they think it over, then search for the brand. The cycle is short, but not short enough to stay within a tight attribution window.

I test queries like « best [product] for [use] » across several assistants. I track brand search and associated queries. I add an open field at order confirmation.

Your product page is no longer the only storefront. The storefront is also the way AI describes you.

B2B and services: the long cycle

Here, the delay leak is at its maximum. A recommendation can precede a first sales conversation by several weeks. When the prospect arrives, nobody remembers what brought them.

I build the origin question into the first call, systematically. Not in a form nobody reads. In the conversation. The salesperson notes the answer word for word in the CRM.

After a quarter, those notes draw a picture that acquisition reports would never have produced. This is what I observe with my clients who apply this discipline: prospects’ wording reveals journeys that analytics had filed under « direct. »

Local: the proximity recommendation

A user asks for « a good plumber near me » or « the best restaurant for a birthday. » The AI answers with a few names. They call, or look up the listing on a map.

The trace is even thinner. A phone call has no UTM and no referrer.

I ask teams to put the question to every new customer, and to note the answer. That’s all. Simple, free, remarkably effective.

In all three cases, I proceed the same way: I don’t look for the exact number. I look for the direction and the consistency between layers.

L’article propose une méthodologie éprouvée en 4 semaines. Commencez par écouter vos clients (semaine 1), puis instrumentez progressivement vos outils (semaines 2–3), avant d’analyser les résultats et d’ajuster.

Your Plan for the Next 30 Days

Enough diagnosis. In your place, I’d proceed like this, in order.

Week 1: map. List 15 to 20 questions your customers ask before buying. The ones you hear on the phone, in meetings, in your support. These are your test queries. They beat any tool-generated list, because they come from the field.

Week 2: record. Ask each question to several assistants. Log in a table: date, assistant, question, brand cited or not, competitors cited, description used. Don’t try to interpret yet. Collect.

Week 3: instrument. Add the open question « How did you hear about us? » to your capture points: form, order confirmation, first call. Track brand queries and direct access every month.

Week 4: cross-check. Put the three layers side by side. Where is your brand cited? What do customers say? What do the brand curves do? Note the convergences and, just as important, the contradictions.

Then repeat every month. The value isn’t in a single reading. It’s in the series.

Keep this principle in mind: you’re not trying to attribute every sale to AI. You’re trying to find out whether visibility in assistants repeatedly precedes movements in brand and demand. That’s a different question. And that one, you can answer.

One last point on internal presentation. Avoid saying « AI brought us this much. » You can’t assert it, and the first question from a finance director will bring the whole structure down. Say instead: « We observe certain elements across three independent sources, we don’t see the others, and we compensate this way. » That’s honest. It’s defensible. And it’s far more credible than a single number with no origin.

I’m not selling you the method. I’m showing you: your place, your absence, your competitors’ presence.

So, in your latest acquisition report, what share of your « brand search » actually comes from a conversation you’ve never seen?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn