Agentic commerce: what it is, and what brands should actually do about it

Agentic commerce is shopping where the software does the shopping. A person states an outcome, and an AI agent handles the search, the comparison, the variant choice and, increasingly, the payment.

The mental model that helps most: for twenty years the internet was built for a human with a cursor. Pages were designed to be looked at. Agentic commerce swaps the reader. The thing evaluating your product is now a program that never sees your hero image, never reads your brand story, and has no patience for a page that takes four seconds to load.

It reads fields. Title, category, size, color, price, availability, shipping window, return policy. Then it ranks.

That single change is the whole story. Every fix on this page follows from it.

Why it arrived now rather than in 2022

Three things had to land together, and in 2026 they did. Models got reliable enough at multi-step browsing to finish a task instead of abandoning it halfway. Payment rails built agent-specific permissions, so spending money on someone's behalf stopped being reckless. And the large platforms shipped protocols that let a merchant be read by an agent without building a bespoke integration for each one.

None of those created a channel on its own. Together they did.

The honest test of a new channel is whether transactions complete and whether the surface behaves differently from the one you already optimize. Both are now answered, and the second answer is the uncomfortable one.

Purchases complete end to end

Meta launched Muse on 8 September 2026, a personal agent running in a private cloud VM with a visible browser and a separate permissions system for sensitive actions, on free, twenty dollar and hundred dollar monthly tiers. Instinct, still invite only, pays through Stripe's Link wallet. Reviewers testing these agents through September completed real purchases on real merchant sites.

The AI shelf is a different shelf

This is the finding that should change your quarter. Productrise studied more than two million product listings across one hundred thousand result pages in the US and UK between 9 and 31 August 2026, running the same queries on both surfaces on the same day:

• Only 1.28% of products ranking in classic Google search also appeared in AI Mode for the same query.
• On products appearing in both, AI Mode prices ran 21.6% higher on average.
• The main seller differed on 49.6% of matched products.
• Across all listings, AI Mode median price was $149 against $100 in classic search.

The answer became an ad surface

OpenAI began testing ads in ChatGPT in the US on 9 February 2026 and now sells them directly through its own Ads Manager at ads.openai.com. On 10 September 2026 Amazon Ads opened a second door into the same surface, a pilot letting select US advertisers extend campaigns into ChatGPT as conversational ads bought through Amazon DSP, with Delta Vacations among the first testers. The unit OpenAI shows is a sponsored product card that appears below a response, clearly labeled and visually separated from the answer itself. Availability is still expanding market by market, and OpenAI maintains a published availability page rather than a fixed list.

Alongside that, the payment networks and model providers have been building merchant-side plumbing. Mastercard has moved on agent payments through Agent Pay and its wider Agent Suite, and Anthropic has published commerce agent blueprints that connect to catalogs, carts and checkout. Availability for all of this varies by market and by acquirer, so treat the specifics as something to confirm with your own payment provider rather than as settled infrastructure.

Shopify reports AI-referred orders growing nearly thirteen times year over year, and AI-referred visitors converting at nearly fifty percent higher rates than organic search. Those are Shopify's figures for their merchant base, not a promise about yours, but the direction is not ambiguous.

Here is the flow end to end. A shopper messages their agent: rain jacket, trip to Seattle next Thursday, under two hundred dollars, medium, nothing neon.

1. Interpret. The agent turns that sentence into structured constraints: category, price ceiling, size, color exclusion, delivery deadline.
2. Retrieve. It pulls candidates from merchant catalogs, protocol feeds or direct crawling.
3. Filter. It checks stock at medium. This is where most brands silently disappear.
4. Compare. It checks delivery dates against Thursday and discards anything arriving late.
5. Present. It returns two or three options with a one-line reason for each.
6. Confirm. The shopper picks one.
7. Transact. The agent completes checkout through a stored payment method.

Steps one through five happened with no human on your site. If you did not survive step three, you were never in the consideration set, and no media spend that week would have changed it.

The five drop-out points, in the order they cost you money

• Step 2, retrieval. Your product truth lives in theme templates and display logic, so there is nothing structured to retrieve.
• Step 3, filtering. Variants are split across separate product records, so the agent cannot confirm stock at medium and skips you rather than guessing.
• Step 3, again. Your category field says "outerwear" while the query needs "waterproof shell", and you never enter the comparison set.
• Step 4, comparison. Your shipping data is absent or stale, so the agent cannot prove you will arrive by Thursday and drops you for a merchant who can.
• Step 6, confirmation. You made the shortlist, and then lost to the other two options because your product image and copy were the weakest of the three at the exact moment a human was choosing.

Four of those five are data problems. The fifth is a creative problem. Fixing only one set leaves the other holding the door shut.

Run this against your own store before you plan anything. Each check is pass or fail, one point per pass. It takes about ninety minutes and needs no tools beyond a browser and an agent account.

Retrieval, four points

1. Structured feed exists. You have a product feed an external system can read without rendering JavaScript.
2. Feed is complete. Every sellable SKU appears in it, not only the ones in your Shopping campaigns.
3. Agents are not blocked. Your robots rules and bot protection allow the agent crawlers you want, checked deliberately rather than inherited from a scraper policy written years ago.
4. Product schema is present and valid. Product markup renders in the served HTML, not injected after load.

Interpretation, four points

1. Variants grouped. Size, color and pack count sit under one parent record, not as separate products.
2. Taxonomy is specific. Categories name the thing a shopper would ask for, not an internal merchandising bucket.
3. Descriptions are literal. The fields contain the words a shopper uses. Poetry lives on the page a human reads.
4. Attributes are complete. Material, fit, dimensions, compatibility and whatever else your category is actually filtered on.

Transaction and proof, four points

1. Price and stock are live at query time, not as of the last crawl.
2. Shipping promise is exposed as data an agent can compare against a date.
3. Returns and policy are machine readable, because agents increasingly filter on them.
4. Video and imagery are tagged and tied to the product record, so your richest assets are not invisible.

Reading your score

• 10 to 12. You are ahead of most of your category. Move to measurement and creative.
• 6 to 9. Normal for a competent brand. The gaps are usually variants and live inventory, and both are fixable inside a month.
• 3 to 5. You are being filtered out before comparison. Stop everything else and fix retrieval and interpretation first.
• 0 to 2. You are effectively absent from the channel. The upside is that the fixes are cheap and nobody in your category has done them either.

Score it honestly. The most common outcome we see is a brand that assumed it was an eight and audited at a four, almost always because of variants.

These are not creative mistakes. They are decisions made on the wrong evidence, and each one is invisible until you watch an agent fail on your own product.

Mistake one: treating variants as products

The most expensive error, and the most common. A jacket in three sizes and four colors becomes twelve product records. A human browsing your site never notices because your theme groups them visually. An agent sees twelve unrelated items, cannot reason that they are the same garment, and cannot answer the only question that matters, which is whether medium is in stock.

What good looks like: one parent record, twelve child variants, each with its own stock and price, all inheriting the parent's description and category.

Mistake two: writing product fields for humans only

"Ocean Breeze" is a name, not a description. If your texturizing sea salt spray is called Ocean Breeze and the description field is atmospheric rather than literal, then a shopper asking for texturizing sea salt spray gets someone else's product. The agent is not being obtuse. There is genuinely nothing there to match against.

What good looks like: the field contains the product type, the key attributes and the use case in plain words. The brand voice lives on the page a person reads, which is a different surface with a different job.

Mistake three: assuming your search ranking carries over

It does not, and the 1.28% overlap figure is the proof. If you are well ranked in classic search, the natural conclusion is that you are covered. Treating AI search as an extension of SEO means you never measure it, never staff it, and find out eighteen months late.

What good looks like: a separate weekly check on your ten highest-intent buying queries, run through the major agents, logged and owned by a named person. Our guide to AI visibility tracking covers how to build that loop.

Ordered by return on effort. All five are doable by a competent ecommerce or ops team inside four weeks, and none require a protocol decision.

1. Group your variants. The highest-leverage change available to most brands. On Shopify this means using the variant model properly rather than publishing separate products, and checking that your feed inherits the grouping. Our Shopify agentic commerce guide walks the specifics.
2. Rewrite your category fields. Replace merchandising buckets with the words a shopper would use. "Men's insulated winter boots" beats "footwear" and costs you an afternoon.
3. Make descriptions literal in the feed. You do not have to choose between brand voice and machine readability. Keep the poetry on the PDP and put the plain description in the feed field.
4. Fix your inventory sync frequency. If stock updates once a day, an agent will confidently sell something you do not have, which is worse than not being listed. Get it to query time.
5. Audit your bot rules. Check what your robots file and your bot protection are currently blocking, then make it a decision rather than an inheritance.

The harder work, and why it is worth it

• Attribute completeness at catalog scale. Filling material, fit, dimensions and compatibility across thousands of SKUs is genuine work. It is also what decides whether you appear in filtered queries, which are the ones with purchase intent.
• A real measurement loop. Building and maintaining a weekly agent check that someone actually reads is an organizational problem more than a technical one.
• Video and imagery tagging. Covered in the next section, because it is the gap almost nobody has closed.
• Protocol integration. Worth doing, worth doing last. See the protocol section below for why.

An agent cannot watch your video. It reads what is attached to it. Most brands attach nothing, which means the richest and most expensive asset in the catalog contributes nothing to how an agent understands the product.

This matters more every quarter, because video is the fastest-growing product asset and the least machine-readable one. A sixty second demo that answers the exact question a shopper asked is worth nothing if the only thing the agent can see is a file name.

What should be attached to every product video

• VideoObject schema with name, description, thumbnailUrl, uploadDate, duration and contentUrl.
• A transcript, because the spoken claims are product attributes the agent can otherwise never reach.
• Captions, which serve accessibility and machine readability at the same time.
• An explicit association to the product record, so the agent can connect the video to the SKU rather than treating it as unrelated media.
• Per-variant association where the video actually shows a specific variant.

The operational problem is scale. Tagging one video is trivial. Tagging video across a catalog of thousands of SKUs, and regenerating it when the catalog changes, is a production and metadata pipeline rather than a creative task. That is the gap we built for, and our guide to catalog video for AI search covers the full approach.

One honest caveat. Whether attaching this changes your placement depends on the surface, your category and your competitors. Treat it as a test on your own catalog and measure per surface rather than in aggregate.

Two open protocols matter, and you do not have to choose between them. Both standardize the same thing: how an agent discovers your products, builds a cart and completes payment without a custom integration per agent.

Agentic Commerce Protocol

Backed by OpenAI and Stripe, and the mechanism behind buying inside ChatGPT. Merchant documentation lives at agenticcommerce.dev. The practical ask is a product feed the agent can read and a checkout endpoint it can call.

Universal Commerce Protocol

Google's equivalent, extended at Google Marketing Live in May 2026 into AI Mode shopping, universal cart, YouTube shopping ads and Demand Gen, with buy-now-pay-later through Affirm and Klarna. Google's developer documentation is the primary source. Early partners include Nike, Sephora, Target, Walmart and Wayfair, alongside Shopify merchants.

The sequencing advice

Fix the catalog first. Both protocols read the same underlying product data, so clean data is the work that pays off under either standard and under whatever replaces them.

Choosing a protocol before your variant structure is correct is choosing a distribution channel for bad data. If you are on Shopify, much of this arrives through the platform anyway and your job is to enable it and check the output. Our protocol guide covers what each one asks of a merchant in detail.

The agent shortlists. A person still says yes. That division of labor is the thing to plan around, and it means creative matters more in this channel rather than less.

Your feed determines whether you make the shortlist at all, which is a data problem. What the shopper sees at the moment of approval determines whether they pick you over the other two options the agent surfaced, which is a creative problem. Winning one without the other gets you nothing.

Three concrete shifts

• Product imagery is doing a job it was not art directed for. When an agent returns three options side by side, your pack shot competes in a context you never designed. Assume the tight crop, the small render and the missing context.
• Ad surfaces moved into the answer. Placements now run inside AI responses, with their own constraints. Feed creative does not automatically survive there.
• More surfaces means more formats, which means more concepts tested per week. This is the argument we made before agents arrived: creative diversity multiplied by velocity is what produces smart volume.

What does not change

The fundamentals hold. A clear hook still beats a clever one. Proof still outperforms assertion. The difference is that you now need the same idea expressed for a feed, for an answer placement and for a shortlist thumbnail, which is a throughput question rather than a taste question.

Both are legitimate. The decision usually comes down to whether your constraint is knowledge or capacity, and most teams misdiagnose which one they have.

Do it in house when

• You have engineering time for catalog work and it is not competing with a replatform.
• Your SKU count is low enough that attribute completeness is a week of work, not a quarter.
• You already have someone who owns organic and can absorb agent measurement.
• Your creative volume needs are steady and your current team meets them.

Bring someone in when

• The catalog work keeps getting deprioritized, which is the single most common failure mode.
• You need per-SKU video across a large catalog, which is an operations problem disguised as a creative one.
• Your creative output is the constraint on how many surfaces you can test.
• You want the measurement loop running next month rather than next quarter.

What to ask any vendor

1. Can you show me an agent failing on my catalog today, specifically where and why?
2. Do you fix the data, make the creative, or only tell me what is wrong?
3. What is the throughput, in assets per month and days per turnaround?
4. Who owns the measurement loop after you leave?

That last question separates a project from a capability. Most engagements that fail, fail there.

Deliberately unambitious. Most brands lose a quarter to a strategy deck when four weeks of catalog work would have moved them further.

Weeks 1 to 2: find out where you stand

• Run the twelve point audit above and write down the score.
• Take your ten highest-intent buying queries. Run each through the major consumer agents. Screenshot what comes back.
• Ask an agent to buy one specific variant of one of your products. Note precisely where it fails.
• Do the same for your two closest competitors, so you know whether your gap is absolute or relative.

Weeks 3 to 6: fix retrieval and interpretation

• Group variants under parent records.
• Replace vague categories with specific ones.
• Make feed descriptions literal.
• Move inventory and price sync to query time.
• Decide your bot policy on purpose.

Weeks 7 to 10: close the proof gap

• Fill attributes on your top hundred SKUs by revenue, then work down.
• Expose shipping promise and returns as data.
• Tag existing product video with schema, transcripts and product association.
• Brief creative for the shortlist moment, not only the feed.

Weeks 11 to 12: make it a habit

• Make the agent check weekly, with a named owner and a place it gets read.
• Re-run the twelve point audit and compare to week one.
• Only now, evaluate protocol integration.

Ninety days of this puts you ahead of most of your category. That sounds like an overclaim until you try buying from your competitors through an agent and watch how many fail at step three.

Most tools in this market cover one link in the chain. Analytics products tell you what worked and produce nothing. Production shops produce without intelligence or strategy. Template tools generate variations with no read on whether any of them should exist.

We cover the whole chain: creative intelligence trained on more than 32 million ads, creative strategy, production, and campaign management. Across the business we have managed over $200 million in ad spend with a team of 70 plus across Bangalore, Noida, Singapore, Canada and the USA.

The three things we do for this specific problem

• Shopify catalog fixes. Variant grouping, taxonomy correction, literal product fields and live price and stock sync. The unglamorous work that decides whether you survive step three.
• Catalog video with correct tagging. Per-SKU product video generated at catalog scale with VideoObject schema, transcripts and product association attached, so your video is legible to agents rather than decorative.
• Performance creative at throughput. 100 or more creatives a month on a five to seven day turnaround, with roughly 3x creative hit-rate lift, at around half the cost of building the same capability in house.

Pricing, published

Software from $299 a month. Managed from $5,000 a month. A productized plan at 5% of ad spend with a $5,000 monthly minimum and a six month term. No percentage taken from creator payments.

Where we are not the answer

If your problem is purely engineering, a replatform or an ERP integration, we are not that. If you have a strong in-house creative team hitting its volume targets and your catalog is already clean, you do not need us and we will tell you so on the call. And if you want a forecast of what this will do to your return on ad spend, we will not give you one. We can tell you what we will ship and how fast. What it returns depends on your product, your price and your market, and anyone promising otherwise is selling something.

If you want a second pair of eyes on where you sit today, the free ad account audit is the fastest way in, and we work with D2C brands on exactly this groundwork.

Each of these goes one level below the pillar, with its own diagnostic and fixes.

Getting found

• LLM SEO, how brands get surfaced and cited inside AI answers.
• How to rank in ChatGPT, the three different things ranking now means there.
• AI visibility tracking, measuring a surface with no rank tracker.

Getting the data right

• Product feed optimization, for shopping channels and agents at once.
• Shopify agentic commerce, what a merchant actually changes in the admin.
• Catalog video for AI search, making video legible to machines.

Getting transacted

• AI shopping agents, a field guide to what exists and how each behaves.
• Sell on ChatGPT, instant checkout and what you give up.
• Agentic Commerce Protocol, ACP and UCP without the spec sheet.

If you want the broader creative groundwork, our marketing guides library covers the production side.