When the Buyer Is a Machine: Rethinking Commerce Search for the Agentic Era

By Seth Earley, Founder & CEO, Earley Information Science

Published: October 5, 2026

Last Updated: October 12, 2026 | Version 1.0

────────────────────────────────────

Who This Is For: E-commerce leaders, VP/Directors of Digital Commerce, product data managers, merchandising directors, and enterprise architects responsible for commerce search and catalog strategy. Also valuable for B2B distribution leaders and CDOs managing product information.

Prerequisites: Basic familiarity with e-commerce search and product information management.

────────────────────────────────────

The most reliable signals in commerce search are disappearing. Click-through rate, dwell time, add-to-cart: every one of them assumes a person is browsing. AI purchasing agents do not browse. They query, filter, and decide — often dozens of times inside a single task — and they leave none of the behavioral traces that relevance models are trained on. The search program built for human shoppers degrades quietly as machine traffic grows, and the dashboards designed to measure it will be the last to know.

This is not a forecast. It is already in the data.

In March 2025, shoppers reaching retail sites through AI assistants converted 38% worse than shoppers arriving from paid search or email. By March 2026, according to Adobe Analytics, that same AI-referred traffic converted 42% better. The volume moved just as fast: Adobe, drawing on more than a trillion visits to US retail sites, reported AI-referred traffic up 393% year-over-year in the first quarter of 2026, following a 693% surge over the 2025 holiday season. In May 2026, traffic hit a new monthly high, up 138% over the prior year.

Those figures mostly describe humans arriving through AI assistants. The fully agentic transaction, where software completes the purchase end to end, now has dedicated checkout and payment protocols backed by the largest names in commerce. The structural shift is already underway.

From a Ranking Problem to an Eligibility Problem

Commerce search learns from shoppers: what was clicked, what was added to a cart, what converted. Industry-standard measures such as NDCG — normalized discounted cumulative gain, which scores how well the most useful results rank near the top — distill those behavioral signals into a score that guides merchandising decisions, personalization, and A/B testing. The system works because millions of small human choices continuously teach it what good looks like.

Put an AI purchasing agent in front of that system and two things break simultaneously.

First, the learning loop thins. Every agent-mediated transaction is a transaction the ranking model learns nothing from. As agent traffic grows, the behavioral data that remains describes a shrinking, increasingly unrepresentative slice of actual demand. The same relevance metrics computed on schedule will describe less and less of the business.

Second, the failure mode changes. A human buyer who encounters a weak result will scan pages, reword a query, and click through category pages. An agent applies the specification and moves on. If a product record is missing the operating temperature attribute, the product is not ranked lower in results. It never enters the candidate set at all. Human-era search failure was a ranking problem. Agent-era search failure is an eligibility problem, and eligibility failures are invisible in behavioral analytics because nothing happens to log. There is no abandoned session to review, no zero-result page to count, no click that shows a competitor won the sale. The demand arrives, evaluates the catalog in silence, and leaves.

B2B distribution has lived a version of this for years. In industrial distribution, the overwhelming majority of revenue in many companies flows through quoting rather than browsing. The quote desk matches specifications against product data, and every gap in that data becomes a delay, an error, or a lost line item. Contract catalog search — where group purchasing organizations and procurement platforms host catalogs that member buyers query rather than browse — has always been a retrieval problem with no behavioral signal to fall back on. A contracted SKU that does not surface on specification goes off-contract, and the revenue leakage does not register as a search failure anywhere in the analytics.

AI agents bring that same reality to every buyer type. Each agent works like a quote desk, matching specifications around the clock. Unlike a customer, it will not call a sales representative to ask about a missing attribute.

What the New Commerce Protocols Require

The infrastructure for machine buyers is being assembled in public. OpenAI and Stripe launched the Agentic Commerce Protocol (ACP) with in-chat checkout in September 2025. Google and Shopify followed with the Universal Commerce Protocol (UCP) in January 2026, co-developed with Walmart, Target, Etsy, and Wayfair. The standards contest will continue for years.

What matters for commerce operators is not which protocol wins. Every contender requires the same entry ticket from a merchant: a structured product feed with defined fields, machine-readable capability declarations, and accurate availability data. The catalog, not the website, is what qualifies a merchant to participate in agentic commerce.

What to Measure When Your Buyers Are Machines

When behavioral metrics stop describing the business, a four-layer measurement stack — each layer bounding the one above it — provides a more reliable picture.

The first layer is product data readiness: attribute completeness, classification consistency, and machine-readability of specifications, scored by category and weighted by revenue. This layer can be measured today, before a single agent appears in the traffic. Most organizations have never scored it, which means most cannot say which of their categories are eligible for agent-mediated demand and which are invisible to it.

The second layer is retrieval quality: for a structured query, did the right products enter the candidate set? This is scored against ground truth built from real specifications, quotes, and RFQs, with correct results defined in advance, not derived from behavioral data.

The third layer is task completion: did the agent accomplish the job it was given, and at what cost? Queries per task, tokens per task, and escalations to a human are the relevant measures. Cost matters here in a way it never did for human sessions, because an agent may run dozens of retrievals for a task a person handled with one search.

The fourth layer is business outcome: revenue, margin, and share of agent-mediated selections won. This layer must always be tracked alongside task completion, never independently. An agent can complete its task perfectly while the basket economics deteriorate.

That deterioration mechanism is worth understanding concretely. A human buyer arrives intending to purchase a drill. On the way to checkout, the retailer monetizes the gap between intent and outcome: the buyer selects a mid-value brand ranked first rather than the cheapest qualifying option, adds a bit set from a "frequently bought together" module, picks up the carry case, accepts a warranty prompt, and consolidates the order in one cart rather than splitting across vendors. Accessories and attach items often carry higher margins than the core product. Retail media adds another layer, monetizing the traffic itself through sponsored placements.

Run the same purchase through an agent. It was given a specification and optimizes strictly to that. Among functionally equivalent products it selects on price and availability. Brand affinity and position bias, which quietly subsidized margin for twenty years, do not operate on software. It purchases the drill without the bit set, because the accessories were not in the specification and it never encounters the module that would have suggested them. It splits an order line by line across vendors without friction, so the retailer wins the low-margin line and loses the profitable items that used to accompany it. Every buyer begins to shop like a procurement department cherry-picking a quote. Task completion reads as a complete success. The basket does not reflect it.

Measuring only from the top of this stack tells you something broke. It does not tell you why. Revenue drops the same way whether agents completed tasks at unfavorable margins, could not complete them, or could not find the products at all. Measure from the bottom instead: product data quality determines what retrieval can find; retrieval determines what an agent can complete; completion determines revenue. Each layer caps the one above it regardless of how well the top layer is tuned.

Merchandising for Buyers Who Never See the Page

For two decades, digital merchandising meant the page: photography, copy, placement, and promotion. Agents never see the page. They see the data. What photography and placement did for human buyers, attribute completeness and classification consistency now do for machine buyers, and the skills that produce those outcomes live in a different part of the organization than the merchandising budget does.

The gap is measurable. Adobe's machine-readability benchmark of US retail sites found that roughly one-third of homepage content is invisible to AI models, with product pages averaging 66% visibility. A substantial share of what merchandising teams built for human shoppers does not exist for an AI purchasing agent evaluating the same catalog.

Tom Wilkie of Grafana Labs described a parallel dynamic on the Earley AI Podcast that illustrates the mechanism. Grafana's observability software is open source, with 25 million users worldwide, and those users generate a vast public corpus of documentation and discussion. Foundation models trained on the open web absorbed all of it. The result: the models know how to use Grafana's software better than they know competing proprietary tools. Nobody planned that as a distribution channel. Openness produced it, and it now shapes which tool gets recommended before any salesperson is involved.

Commerce catalogs face the same selection pressure through two channels: training-time familiarity, where models absorb what exists in structured, consistently described public form; and inference-time retrievability, where at the moment of purchase an agent can only evaluate what it can fetch and parse. Both mechanisms reward the same investment. Catalogs that are structured, complete, and machine-readable get learned and get retrieved. Catalogs hidden behind thin data and rendering-dependent pages do not get out-ranked. They do not get considered. Structure becomes a distribution strategy, and it places revenue responsibility somewhere most organizational charts do not anticipate: the product data team.

What to Do While Traffic Is Still Human

Start with product data scoring. Audit attribute completeness and classification consistency by category, weighted by revenue. Overall averages mislead; an acceptable aggregate score routinely hides categories where the data cannot support a structured query. Revenue-weight the audit and the priorities set themselves. If a contract catalog is in play, score it first, since contracted categories concentrate both the revenue and the eligibility exposure.

Build specification-based ground truth. Take a set of real specifications from quotes, RFQs, and support interactions, define the correct retrieval result for each, and score retrieval against that set on a quarterly basis. This is inexpensive and unglamorous, and it is the metric that will still mean something when the behavioral data thins.

Segment every search metric by buyer type. Track agent share of traffic explicitly and report human and machine performance separately. The crossover point, when machine queries exceed human ones, will arrive at different times in different categories, and it is worth watching it happen rather than reconstructing it after the fact.

None of this requires predicting adoption curves. The argument holds at any agent share above zero, and catalog remediation runs in quarters, not sprints, because product data does not restructure on demand. Start with the layer that needs no agent traffic at all. Score the product data. Every other measure in the agentic commerce era is bounded by it.

Read the original version of this article on VKTR.