AI-Driven Customer Journeys: 45% Can't Compare Them

The IAB raised its 2026 U.S. ad spend forecast to 12.3% growth on 10 September 2026, up from the 9.5% it projected in January. Strong first half, a Winter Olympics, a World Cup, and buyers who stopped worrying about the economy long enough to spend.

Buried further down the same release is the number an agency owner should find harder to sit with. 86% of buyers say they are already changing, or expect to change, how they measure media performance because of conversational AI tools and AI agents. And 45% name the difficulty of comparing AI-driven customer journeys to traditional ones as their single biggest measurement challenge.

Read that second number precisely, because it is easy to misread as “AI can’t be measured.” It does not say that. It says compared. Advertisers have measurements. What they have lost is the bridge between the old ones and the new ones, which is a different problem with a different fix.

The measurement industry is already building that bridge. In the six weeks either side of the forecast, it standardized its taxonomies, published guidance on modernizing marketing mix modeling, shipped a second edition of its AI disclosure framework, and turned open-source MMM into something a chatbot can drive. It is genuinely good work.

Then read the vocabulary. In the measurement framework IAB publishes on its own site, “decision” appears 13 times and the words “approver,” “actor,” “human” and “record” appear zero. That is the gap this post is about.

What the IAB actually reported

The primary source is worth reading in full, because the headlines compressed it into a growth story. Per IAB’s own release, the forecast rose 2.8 percentage points from January on the strength of the first half and a clearer economic read. Customer acquisition jumped nine points to 63% of buyers, brand equity rose six points to 43%, and repeat purchases held flat at 24%.

The measurement findings sit in a section IAB titled AI Is Changing the Consumer Journey Faster Than Marketers Can Measure It. That heading is the whole story. It fits in a press release because it is true, and it is not a compliment.

The three findings that matter for anybody running client accounts:

• 86% of buyers are changing, or expect to change, how they measure media performance because of conversational AI tools and AI agents within twelve months. That is not a fringe experiment. That is the market.
• The most common response, selected by 48%, is measuring brand visibility and citations directly inside AI tools. Buyers are also leaning on branded search and direct traffic as proxies, third-party AI discovery tools, referral monitoring, and more incrementality testing.
• 45% still cannot join those signals into one view, naming the comparison of AI-driven and traditional customer journeys as their top measurement challenge.

Two more numbers to hold next to those. Optimizing content for AI-generated answers is now the number one area of increased focus among buyers at 76%, and adapting to AI-driven search is the top media investment challenge at 44%. Meanwhile focus on using generative AI inside campaigns fell to 69% from 78% in January.

That last inversion is the tell. A year ago the industry was asking how to use AI. Now it is asking how to reach a consumer who uses AI — and quietly deprioritizing the AI it runs itself. As IAB’s Chris Bruderle put it, marketers are discovering that people are “evaluating brands inside these environments” and are now “figuring out how to connect those interactions to the signals and outcomes they have measured for years.”

The standards layer is arriving on schedule

Give the industry its due before criticizing anything. Much of the plumbing a finance agency now depends on is being written in public, by people who clearly understand the problem.

Project Eidos is IAB’s measurement harmonization initiative, run through its Measurement Center and the Measurement Advisory Committee. Its own framing is unusually plain about the stakes:

Marketing measurement is under strain. Signals are disappearing. Data is fragmented across the industry. Outcomes vary by definition. And as AI accelerates how measurement is built and scaled, confidence in the answers has not kept pace.

Eidos works on three fronts, and the third is the one this post cares about: unifying and harmonizing measurement; cross-channel outcomes, attribution and incrementality; and modernizing marketing mix modeling — described as “privacy-ready, comparable MMM approaches with standardized inputs and outputs.”

It is not a discussion group. Campaign Data Standards 1.0 was finalized on 10 August 2026 — a pair of complementary Ad Type and Ad Inventory taxonomies IAB describes as “a common language for cross-channel measurement.” And on 18 August the AI Transparency & Disclosure Framework V2 arrived, translating a materiality-driven approach into practical protocols for when AI involvement must be disclosed in consumer-facing content.

So in one month, the industry got a shared dictionary for what a thing is, a shared framework for how to measure whether it worked, and a shared rule for when to tell the consumer AI was involved.

Meanwhile the model layer is splitting in two

The measurement standard is converging. The tools underneath it are not. As AdExchanger reported on 11 September, Google’s Meridian added a chatbot interface that answers questions about how to build a model and audits your data quality, moved its GeoX geographic testing feature out of beta, and began accepting upper-funnel signals like Branded Google Query Volume alongside lower-funnel conversions. Google also wired Data Manager into Analytics and DV360 and adopted IAB Tech Lab’s ECAPI standard.

Two words in that sentence are doing category-confusion work. “Audits your data quality” means the chatbot checks your inputs. That is a useful feature and it is not a record of anything a human did.

At the same time as Google deepened its open-source MMM, AdExchanger reports that Meta has stepped back from aggressively pushing Robyn, its own open-source MMM tool. The open-source MMM ecosystem is not consolidating around a winner. It is bifurcating along the same lines as the walled gardens that built it.

Which makes the barrier to entry story more interesting than the tooling story. In an August interview with AdExchanger, Julian Runge — a Northwestern marketing professor who co-authored the first academic paper on open-source measurement — described what agentic AI does to MMM:

His own conclusion is the part nobody quotes: “You still need human intelligence that can understand what’s happening.” And on what goes wrong without it: “The risk, though, is overconfidence. If you dive into the advanced modules without being able to vet what’s happening, you can end up making bad decisions based on analyses you hallucinated together with the LLM.”

Hold that thought next to the IAB number. The barrier to running sophisticated measurement just dropped to a prompt, the guidance for measuring a cross-channel journey is finally being standardized, and the thing standing between a bad decision and a client’s budget is a human who had to exercise judgment.

Now read the framework’s own vocabulary

Here is the test. Take the measurement framework IAB publishes in its own research library — Building a Suite of Truth: A Hybrid Approach to Measuring Incrementality, published April 2026 and drawing on original research from Meta — and count the words it uses to describe what it asks a marketer to do.

It is 34,118 characters. Across all of them:

• “approval” — 0. “approver” — 0. Also zero for authorize, authorised and sign-off.
• “record” — 0. “trail” — 0. “timestamp” — 0. “provenance” — 0.
• “actor” — 0. “human” — 0. “responsibility” — 0. “accountability” — 0.
• “jurisdiction” — 0. “retention” — 0. “export” — 0. “client” — 0.
• “decision” — 13. The framework is full of decisions. It simply never asks who made one.

The word “audit” does appear — twice — and both uses describe inspecting your own measurement setup: “the initial audit of the current measurement system” and “auditing the status quo or source of truth.” In this framework, to audit means to review the model. Not to produce evidence of anything.

This is not a criticism of the document, which is careful and useful. The framework’s job is to make a number defensible, and it does that job with a three-rung ladder: rules-based attribution on the bottom rung, experiments on the top, MMM and advanced MTA in between. The third-party research it cites is honest about how much is at stake — Kantar finds advertisers carry 3.8 measurement solutions on average, and Meta’s own 307-study analysis found 31% of incremental conversions misallocated to other channels under non-incremental models.

But that is precisely why the silences are a specification rather than an oversight. A standard built to make the number comparable will spend every character on the number. Nothing in it is responsible for the other question, which is the one a client asks and a regulator repeats:

Two absences, and the layer nobody is filling

Notice what the August disclosure standard does cover, because it is the closest thing in this story to a record. AI Transparency & Disclosure Framework V2 sets out when and how AI involvement should be disclosed across consumer-facing advertising and marketing content. It governs what you tell the outside world. It says nothing about what you keep for yourself.

Put the three layers side by side and the shape of the gap is obvious.

None of the first three is doing anything wrong. It is not the standards body’s job to log your agency’s approvals, and it is not Google’s job to tell a regulator which of your account managers moved a budget. The measurement layer standardizes the number. The model layer estimates the number. The disclosure layer describes AI to the consumer. Nobody in that list was ever asked to produce the decision record.

That is why “we have a measurement framework” and “we can prove what we did” are not the same sentence — a distinction that keeps costing agencies money, most recently in a federal courtroom where the remedy was an export right for publishers and a prohibition for advertisers, and in the EU, where the deployer duty landed on the agency rather than the model provider.

The 45% is a comparison problem, and comparison has a floor

There is one more reading of the IAB number worth holding onto. Buyers did not say their AI-era data is missing. They said it cannot be compared to the data they already trust. Practitioners interviewed by AdExchanger have described models diverging by several orders of magnitude on the same spend, which makes calibration — not more dashboards — the scarce skill.

Every one of those calibrations is a judgment call somebody makes and somebody signs. When the standard arrives and the model agrees with the experiment, someone decided which experiment to trust and how much of the model to rewrite. That decision is as auditable as a budget change and, today, as likely to be recorded.

What a finance agency should actually do with this

You are not going to standardize your clients’ measurement architecture, and you do not need to. The useful moves are smaller and sit one layer under the standards.

• Separate the two questions in your own reporting. “Did it work?” belongs to the measurement stack. “Who decided, and under whose authority?” belongs to your operations record. If one document is trying to answer both, it is answering one of them badly.
• Record the calibration calls, not just the results. The model change, the test that justified it, the person who approved it, the date. That is the least standardised and most defensible artifact an agency can hold.
• Name AI as an actor. A record that logs the human whose session happened to be open while an agent moved a budget is not a record of the actor. Runge’s supervision requirement is the thing a regulator will look for, and it has to be attributable to a named person.
• File the change against the client. Ott’s agency hierarchy runs Client → Brand → Account, so a budget move is filed against the engagement it was made for — which is the axis a supervisor asks about, and the one the standards being written in 2026 do not address.
• Store it independently of the platform that accepted the change. A record living inside an ad account survives exactly as long as the account does. Ask what happens to your evidence when an account gets restricted.

The good news in the IAB study is that the industry is finally building the measurement vocabulary it has needed for a decade, and the forecast it raised is real money an agency can plan against. The uncomfortable news is that the 45% figure is not a technology gap waiting on a better model. It is a comparability problem, and comparability problems get solved when somebody can say, with evidence, what they did and why.

The standards will arrive. The dictionary will be shared. What will still be true on the day they do is that the number belongs to the platform, the model belongs to whoever built it, and the decision belongs to you. Which is the same reason the agentic ecosystem counts launches rather than records — and why governing the draft is only half the coin. Measure the journey. Record the decision.