In 2020, Airbnb found itself with no product to sell. Travel had stopped. The company cut approximately $800 million in performance marketing almost overnight. Survival, not strategy. When people started moving again, the spend did not fully return. What returned instead, according to the company's own reporting, was direct traffic. Customers they had been paying to reach via search had, it turned out, simply remembered where Airbnb was all along.

Brian Chesky made the obvious observation in a way that somehow still surprised the press: they had been paying to reach people who would have come anyway. The spend was not driving the relationship. It was obscuring the cost of not having one.

This is not a story about Airbnb. It is a story about what organisations measure, why they measure it, and what the consensus in a budget meeting is actually made of.

It is 2026. The story has not simplified.

I have recommended cutting direct-to-consumer performance spend in financial services twice. Once at a major Canadian bank, in mortgages: one of the highest-consideration, highest-cost product categories in retail banking. Once at a bank-owned insurance business, in a different competitive context but with the same structural logic. Both times, the room went quiet in that particular way rooms go quiet when someone has said something that is either obviously wrong or uncomfortably right.

Both times, volume held. In one case, it grew.

I am not making the argument that performance marketing is a waste. I am making the argument that the institutional case for performance marketing in large organisations is almost always built on metrics that are easy to defend in a budget meeting, not metrics that are most accurate.

In 2026, that problem has acquired a new dimension.

Attribution is not neutral

Every attribution model is a theory of how customers arrive.

Last-click attribution says the final touchpoint before purchase was the decisive one. Multi-touch attribution distributes credit across the journey. Media mix modelling attempts to measure long-run brand effects alongside short-run activation. Each model rewards different behaviour in the teams that use it.

Last-click attribution did not kill brand investment. It made brand investment invisible, which is considerably worse, because invisible things do not appear in the conversation about what is working.

When an organisation runs last-click attribution on a mortgage or a GIC, it will find that paid search performs exceptionally well. This is true. It will also fail to measure the eighteen months of email open rates, branch conversations, and late-night website visits that preceded the search query. This is also true. The attribution model is not lying. It is answering a narrower question than the one the organisation thinks it is asking.

The decision to cut performance spend was, in both cases I was involved in, an epistemological decision before it was a budget decision. We had to agree on which framework was telling us the truth before we could agree on what to do about it.

The structural bias

There is a reason institutional marketing consensus almost always lands on the side of measurable, short-term spend. It is not stupidity, and it is not laziness. It is the legitimate pressure of accountability in organisations where budget decisions are made in rooms with multiple stakeholders, quarterly targets, and a shared preference for things that can be explained simply.

Les Binet and Peter Field documented the mechanism. Brand investment works on a longer timescale, through channels that attribution models do not capture, and produces effects that are unmeasured rather than absent. This makes it systematically vulnerable to cuts in budget reviews because its effectiveness is harder to demonstrate at the pace of a fiscal quarter.

The controversial call is not a maverick call. It is what follows when someone in the room is reading a wider set of signals than the attribution dashboard. The argument exists. The data exists. What is usually missing is the institutional permission to say it plainly.

The measurement gap has narrowed. The imbalance has not.

Here, in fairness, is the strongest version of the counterargument.

AI-accelerated media mix modelling, tools like Google's Meridian and Meta's Robyn, now run in near-weekly cycles rather than annually. Incrementality testing with geographic holdouts has become considerably more accessible. The institutional excuse that brand effects cannot be measured is thinner in 2026 than it was in 2013. At the edges, the measurement problem is being solved.

This is worth acknowledging. It is not, however, the whole picture.

The cost of producing activation content has collapsed. AI tools have made performance creative, the ad variation, the landing page, the search copy, the email sequence, faster and cheaper to produce than at any point in the history of advertising.

Brand-building work has not gotten proportionally cheaper. Genuinely emotional, culturally resonant creative still requires the same judgment, craft, and institutional courage it always did.

The structural bias toward activation has amplified, not corrected.

The measurement gap has narrowed at the edges while the temptation to over-index on what is measurable has intensified. The mechanism is running faster now.

The new unmeasured channel

There is a new purchase influence channel that sits entirely outside every attribution model that currently exists. It does not appear in any dashboard. It cannot be bought in the conventional sense. It rewards brands that have built something over time, and it penalises, quietly, without error messages, brands that have not.

It is called a recommendation.

When someone opens an AI assistant in 2026 and asks which mortgage provider to consider, which wealth management platform to trust, which insurer to call, the assistant gives an answer. That answer is drawn from a model of the world built on reputation, cultural visibility, review patterns, and the accumulated weight of years of brand-building activity. The remarkable thing about AI recommendation engines is that they have absolutely no interest in which brand bought the most keywords last quarter. The model was trained before the campaign ran. The associations it holds were formed over years, not sprints.

You cannot buy your way to the top of that answer the way you bought a paid search result.

This is the new direct traffic.

Airbnb discovered that people remembered the brand without being prompted by paid media. In 2026, people are increasingly asking AI assistants instead of searching, and the AI recommends the brands that built something worth recommending. That entire channel is unmeasured by any attribution model. It produces effects that look like organic brand strength and are just as difficult to trace back to a single budget decision, because they cannot be. They are the residue of years.

The implication for financial services is particular. The category is one of the highest-trust, highest-consideration purchase environments in existence. An AI assistant asked about mortgages or retirement savings will surface the institutions associated with competence, stability, and trustworthiness at a cultural level. These associations do not come from a campaign. They come from twenty years of showing up.

The brands that have been quietly cutting brand investment in favour of measurable activation spend have been conducting, without realising it, a slow experiment in whether they can afford not to be recommended.

What to do about it

Decide which question you are actually asking before you build the measurement framework. "Where did this customer come from immediately before purchase?" and "what built the conditions under which this customer was already considering us?" remain different questions requiring different models. In 2026, add a third: "What built the conditions under which an AI assistant would name us unprompted to someone we have never reached?" That question does not yet have a clean measurement answer. That is not a reason to defer it.

Find the signals that cannot be bought and watch them carefully. Direct traffic is still the most honest metric available; the accumulated result of people actively deciding to seek you out. AI-sourced traffic and unprompted brand mentions in AI outputs are the new version of the same signal. You cannot optimise your way to either. You build toward them, over time, or you do not.

Build the internal case before the channel becomes legible. The time to argue that your measurement framework is incomplete is not the strategy review where you are defending a brand investment. It is now, before AI-influenced discovery appears on any dashboard. The brands that ask this question early will have the structural advantage by the time the data is clean enough to defend in a budget meeting.

Airbnb's direct traffic had always been there. They were simply paying not to know it.

In 2026, there is a new kind of presence accumulating outside every measurement system that exists. It arrives through an AI assistant's answer, and it is the product of having built a brand that means something at a cultural level. It will not appear in attribution data until someone builds the right question, which is to say, it will not appear until the organisations that asked the question early have already made it expensive for the ones that did not.

The most uncomfortable question in any marketing review is not "is this working?" It is "are we even looking at the right field?" In my experience, most organisations stop one question short.

They have been doing so since 2013. The cost of that habit is rising.

K Baksh