Agentic commerce: what it is, what changed in 2026, and what a brand actually has to do

Agentic commerce is online purchasing in which an AI agent acts on the shopper's behalf, reading structured product data to discover, compare and in some cases buy a product without the shopper ever browsing a store. The shopper states an outcome, not a query, and the agent resolves it against whatever product records it can actually parse. That last clause is the whole commercial question, because an agent cannot shortlist a product whose record does not answer the constraint it was given.

Most writing on this subject is about what agents will do to retail. This page is about what a brand has to change, and by when.

What is agentic commerce, in plain terms?

Agentic commerce is the model where a shopper delegates the work of shopping to software. Instead of typing "black linen dress" into a search box and filtering the results, the shopper says something closer to find me an unlined black midi dress for a July wedding, under $300, that arrives by Friday, and an agent goes and resolves it.

The shift that matters is not conversational phrasing. It is that the constraints move from the filter sidebar into the request itself. "Unlined," "for a July wedding," "arrives by Friday" are all conditions the agent has to check against something. If your record carries fabric, lining, occasion and dispatch data, you are checkable. If it carries a style code and a house color name, you are not.

Agentic AI commerce and agentic commerce are used interchangeably. The distinction some writers draw between them is not load-bearing.

How is agentic commerce different from traditional e-commerce?

The difference is who reads your product record, and what they do when it is incomplete. A shopper who lands on a thin product page can still buy, because a photograph and a price are enough for a person. An agent working from the same record has nothing to evaluate and moves on silently.

The same catalog, read two ways. The failure mode in the right-hand column produces no error message.

That last row is the reason this is difficult to manage. Every other channel tells you when it is going badly. Agentic surfaces do not produce a disapproval, a warning, or a drop you can attribute. The product simply stops being an option, and the dashboard looks the same as it did last week.

What actually happened in agentic commerce in 2026?

The infrastructure arrived faster than the demand did, which is why the market feels simultaneously overhyped and urgent. Here is the record, with dates.

The agentic commerce timeline, September 2025 to August 2026. Kept current; see the changelog at the foot of this page.

Two things are worth reading out of that table rather than skimming past.

First, every one of those releases is a pipe, and not one of them supplies the quality of what flows through it. Google built the protocol, the cart, the payment rail and the reporting. It did not write your product descriptions. The input is the operator's job and it has stayed the operator's job through the entire build-out.

Second, May 2026 is the inflection, not January. January gave the market a protocol. May gave it a set of fields to fill in and a report that shows whether filling them in worked. The budget follows measurement, and measurement arrived in May.

What does an agent actually read when it picks a product?

An agent reads two layers: the feed you send out and the page it crawls. It resolves the shopper's constraints against structured fields first, because structured fields are cheap to check and prose is expensive to interpret.

In practice, that means the following, in roughly this order of influence:

1. 1.Title. The first thing scored against the request. Internal style (style numbers, house color names, brand-first construction) costs you here more than anywhere else.
2. 2.Category and product type. Where the item sits in the commerce graph. A top-level bucket gets you benchmarked against the wrong products.
3. 3.Filterable attributes. material, pattern, size_system, age_group, gender, color. These are the fields a constraint resolves against.
4. 4.The richer structured fields. product_highlight, product_detail, and now the conversational attributes. Most catalogs ship these blank.
5. 5.The crawlable page copy. Read for what the feed does not carry, and check for agreement with the feed.

The six conversational attributes Google added in May 2026 (question_and_answer, document_link, related_product, item_group_title, variant_option, popularity_rank) sit squarely in layer four. They are optional, they do not affect product approval, and most merchants have not filled them in.

What goes wrong when a product record is thin?

Two failure modes, and the second is worse than the first because it is loud.

The quiet failure is exclusion. The product is fine, the price is fine, and it is simply never a candidate, because nothing in the record answers the constraint the shopper stated. There is no signal. You find out at the end of the quarter, if at all.

The loud failure is misreading. An agent working from an ambiguous or incomplete record does not abstain, it infers. Where a person reading a confusing spec or a partial ingredient list asks a question or moves on, an agent produces an answer with confidence and acts on it. That inference reaches the shopper as a statement of fact about your product, and you did not write it. It is the worse of the two failures, because exclusion is silent and this one is wrong out loud.

Our position, stated plainly: every company selling services in this area is working it out as it goes, because the surfaces are new and nobody has the full picture of how citation and selection actually resolve. What is knowable is which inputs those systems read, and whether yours are complete. Start there rather than with a theory of the algorithm.

Which agentic commerce platforms and protocols matter?

Four protocols carry most of the traffic in this conversation, and they do different jobs. Confusing them is the most common error in vendor material.

The four protocols, what each one governs, and who backs it.

None of these is a strategy. They are plumbing, and they are converging: the ecosystem has moved to a published release cadence with fulfillment and payment work aimed at the holiday season. What a brand controls is not which protocol wins. It is whether the record travelling down all of them is complete.

What should a brand fix before Q4?

Work this in order. The first three cost nothing but time and gate everything after them.

1. 1.Audit titles against how people actually ask. Pull your top 200 SKUs by revenue and read the titles cold. If a title needs internal knowledge to decode, rewrite it so the most-searched descriptor comes first.
2. 2.Fill the filterable attributes. material, pattern, size_system, age_group, gender. These are the fields constraints resolve against and they are empty in most catalogs.
3. 3.Deepen google_product_category and product_type. A two-level internal path gives bidding nothing to act on and puts you in the wrong benchmark set.
4. 4.Populate the six [conversational attributes](/blog/merchant-center-conversational-attributes). Optional, zero risk to approval, and largely unclaimed. Mine your own support tickets and PDP Q&A for the question_and_answer values: the answers your team already types every week are the source material.
5. 5.Check the page and the feed. An agent reads both. Where they disagree, you look unreliable.
6. 6.Open AI Performance Insights in the Merchant Center and read your share of voice against similar brands before you change anything, so you have a baseline.
7. 7.Re-run all of it on a cadence. A catalog is a population, not a project. New SKUs arrive with the original defaults, and any pass you completed in June is already decaying.

Step seven is the one that turns a successful project into a flat year.

How do you prove agentic work actually worked?

Against a control, or not at all. This is the part of agentic commerce most likely to be sold on a score rather than a result, and a score is not an outcome.

The workable method is unglamorous: hold spend constant, change only the product-language record for a treatment group, and compare against an untouched holdout over a named window. In a matched-spend A/B test with a 28-day holdout, rewriting the product-language input produced a 28% increase in Google Shopping revenue against the untouched control. On Meta Advantage+, a holdout test cross-validated against Meta's own Conversion Lift study measured a 21.4% improvement in ROAS.

Those are paid-surface results, stated as such deliberately. On agentic and answer-engine surfaces the honest claim is narrower: you can improve readiness and input quality, and you can measure whether your share of voice moved. Placement, citation and final selection sit with the platform. Any vendor telling you otherwise is selling a guarantee they do not control.

Would it survive a finance review?

That is the question to carry into every agentic commerce conversation this year, and it disqualifies most of what is currently being sold.

A CFO asked to fund agentic readiness will ask four things, and they are the same four whether the surface is Google Shopping or an assistant nobody had heard of last spring. Are you fixing the input, or a layer above it. Does the fix run continuously, because the catalog churns. Does it cover the surfaces that pay this quarter, not only the interesting ones. And can you show lift against a control.

An agent cannot recommend a product it cannot parse, and it will not tell you when it skipped you. The work is to make the record answerable, then to prove the answer moved something. Start with your top 200 SKUs and the five empty attribute fields, and get a baseline out of AI Performance Insights before you touch anything.

Frequently asked questions

What is agentic commerce?

Agentic commerce is online purchasing where an AI agent acts on a shopper's behalf, reading structured product data to discover, compare and sometimes buy a product. The shopper states an outcome rather than a search term, and the agent resolves it against whatever product records it can parse.

What is the difference between agentic commerce and traditional e-commerce?

In traditional e-commerce a shopper browses and filters, so a thin product record creates friction the shopper can work around. In agentic commerce the agent checks stated constraints against structured fields, so a thin record results in silent exclusion from the shortlist with no error and no signal.

Is agentic commerce actually happening yet, or is it hype?

Both are true at once: the infrastructure shipped fast across 2025 and 2026, while live merchant adoption has lagged the announcements considerably. The practical position is that discovery through AI is real now, agent checkout is still early, and the product-data work required is identical either way.

What is agentic commerce optimization?

Agentic commerce optimization is the practice of preparing product data, feeds and page content so AI agents can understand, evaluate and recommend a catalog. It is an extension of the same product-language work that drives paid and onsite performance, not a separate discipline with separate inputs.

Which protocols matter in agentic commerce?

ACP governs discovery and merchant handoff in chat, UCP governs how agents interact with merchant catalogs and checkout, AP2 governs agent payment authorization, and MCP governs how models connect to external data. None of them determines the quality of the product record travelling through them, which stays the merchant's responsibility.

How do I get my catalog ready for shopping agents?

Start by auditing titles on your top revenue SKUs, filling the empty filterable attributes, and deepening your category and product type paths. Then populate the six Merchant Center conversational attributes, confirm your feed and product page agree, and take a share-of-voice baseline before making further changes.

What actually goes in Google Merchant Center conv…