Friction Was Doing a Job

For 20 years the growth playbook had 1 reliable move: remove a step. Fewer fields, fewer clicks, fewer pages. It worked so consistently that it stopped being a hypothesis and became an assumption. Agentic checkout was that assumption taken to its limit, and it is the most useful failed experiment the industry has run in years.

What the merchant page was actually doing

Treat a product page as a series of obstacles and removing it looks like pure gain. Treat it as a series of answers and the picture inverts. Every element on that page exists because some buyer needed it before they were willing to spend money: the photography, the reviews, the returns line, the delivery estimate, the fact that other people bought this and did not regret it.

None of that is friction. It is evidence. Strip it out and the buyer is left holding a recommendation from a system that will not be responsible for the outcome, which is exactly the moment people hesitate. I made the general form of this argument in The AI Trust Gap: the first session of an AI product is a trial, not a tour. Buying is the same trial with money attached.

The agent is a great qualifier and a poor closer

Here is the part that gets missed when this is written up as a failure. The discovery layer worked. Adobe's analytics reported that AI-referred visitors completed purchases at a higher rate than visitors arriving from traditional search. The assistant researches, compares, filters, and hands over a shopper who has already made most of the decision.

So the agent is doing the top of the funnel extremely well and the bottom extremely badly. That is not a contradiction. It reflects what each surface is good at. The assistant is good at narrowing a field of options. The merchant is good at closing a specific one, because closing depends on trust that a third party cannot lend you.

What this changes about your funnel

The first consequence is that you now optimize for a shortlist you cannot see. An agent evaluating your catalog reads structured attributes, live pricing, and real inventory. If your product data is stale or unreadable, you are not ranked badly, you are absent. That is a harsher failure than a low position on a results page, and it is the mechanic I described in When Your Next User Is an Agent, applied to a shelf instead of a signup.

The second is that AI-referred traffic is a segment, not a source. These visitors arrive pre-qualified, mid-decision, and carrying context you did not generate. Sending them to the same landing experience as a cold search visitor wastes the qualification. They need confirmation, not persuasion, and the page that serves them best is closer to a receipt than a pitch.

The third is measurement, and it is the ugly one. The comparison never happened on your property, the deliberation never happened on your property, and the click that used to record intent may never fire at all. That is the gap I wrote about in Attribution After the Click, and agentic commerce makes it a revenue problem rather than a reporting one.

The lesson is about assumptions, not about agents

The industry did not get this wrong because it was careless. It got it wrong because a rule that had held for 2 decades was applied to a case where the underlying reason had changed. Removing steps helps when the steps are overhead. It hurts when the steps are the argument.

Every growth team has a version of this in its own funnel. The onboarding screen someone wants to delete because it adds a click. The confirmation step that looks redundant. The pricing page detail that lengthens the read. Before removing any of them, the question worth asking is what that element is answering, and whether the buyer will still get that answer somewhere else.