E-commerce
August 19, 2026
This page is not a ranking. It is a four-filter grid: price comparison (18 badges), bargain hunting (16), product search (19), checkout funnel (14). These totals are directory labels, not 67 distinct agents, not an addressable market, not a gross merchandise value (GMV). The same product often carries three badges. Reading four acquisition channels is already asking the wrong question. This article does not invent a better agent per task either.
Google launched the Universal Commerce Protocol (UCP) on January 11, 2026 as an open standard for the entire shopping journey: discovery, purchase, after-sales. This is not the directory's taxonomy. It is a named rail, initially on eligible US merchant listings in AI Mode and the Gemini app, with the brand remaining the merchant of record for the transaction. On the same day, Google announced Merchant Center attributes for conversational discovery, and Direct Offers, a Google Ads pilot labeled "Sponsored deal." Stripe documents the Agentic Commerce Protocol (ACP) as an open standard created with OpenAI and Meta: checkout funnel sessions, cart and catalog, delegated payment, delegated authentication, orders, and webhooks. This is a delegated purchasing protocol, not a coupon protocol or a price comparison tool. Lumping UCP, ACP, and the four filters into a single "agents per task" metric is budgeting for a rail you don't have, or a coupon you already do.
If the concept of a software buyer is new: what agentic commerce is. The 2026 timeline: what has been delivered. The grid by product, not by task: AI shopping agents. Official totals remain human shopping journeys: 2026 e-commerce statistics.
Summary
Four filters, four professions, one word: 'agent'
Four filters. Four merchant jobs. One word "agent".
Filter | Directory phrase | What it changes for you |
|---|---|---|
Product search | Summary of reviews, features, fit, for a short list | Attributes, structured reviews, Q&As. Not the JPEG header image |
Price comparison | Fetch and compare prices between brands | Price, currency, stock, identifier, freshness. Two lying CSVs are still two prices |
Bargain hunting | Monitor prices, coupons, flash sales | Promo codes as data. This is not Direct Offers (Google advertising) |
Checkout tunnel | Fill in address and payment, confirm on behalf of the human | Theme, web application firewall, 3-D Secure, or a UCP or ACP rail. Three mechanisms, one badge |
Separating the use cases is right. A buyer comparing a price is not a buyer confirming a mandate. Reading the counters 18 / 16 / 19 / 14 as four markets, no. These are labels. ChatGPT, Gemini, and Copilot appear in several buckets. Honey is in checkout tunnel, bargain hunting, and price comparison, not in product search. Rufus is in search and price, not in the checkout tunnel. Macaron is only in search. VendeeX (spend governance) is stuck to the checkout tunnel and search. Klarna (buy now pay later) is in three buckets. This is not a census. It is an editorial grid.
For you, the useful column is the third one. Not the logo. A "search" badge requires fields that a LLM can cite. A "price" badge requires a real offer price. A "bargain hunting" badge requires a code that exists in the promo engine, not in a banner. A "checkout tunnel" badge requires knowing if the agent clicks your theme or if it processes the payment in a chat. Four internal requests. A "we are in agents" slide opens none of them.
The child pages use the same cards, filtered: product search, price comparison, bargain hunting, checkout tunnel. Use them to browse, not to conclude a budget. None of these pages have accuracy rates, returns, or disputes. This one doesn't either.
A merchant who "chooses the checkout tunnel task" like a payment provider has read the title, not the badges. Operator, Honey, Gemini, and Klarna do not do the same job there. Treating them as a single ranking is like comparing a coupon, a buy now pay later option, a Chrome extension, and a Google rail.
Convert over 2,000 customers on average per month with Qstomy.
The world’s 1st Shopify AI dedicated to customer conversion
Empowering 200+ e-commerce merchants
67 pastilles, about twenty products
Add 18, 16, 19, and 14: you get 67 badges. The product directory lists about twenty. The difference is the overlap. A "all merchants" agent that searches, compares, and assists the payment funnel takes three labels. This is not three integrations. This is not three costs per acquisition.
Do not convert these totals into market shares. They shift when a listing is relabeled. They do not tell you how many of your customers use one of these products. They do not tell you which one converts. A Verified listing on this site does not buy a badge either: editorial and paid content remain separate.
The merchant test is not "how many buckets am I in". You are in no bucket. The agents are. You are in their index, their marketplace, their indexing, or not. The useful question: if a buyer arrives via search, price, coupon, or payment funnel, do my catalog and payment funnel hold up?
Vocabulary (discovery, mandate, merchant of record): glossary of agentic commerce. Here, "task" means human use in the directory. It does not mean a protocol brick.
Product search: features and reviews, not the header visual
Product search, in the directory phrase, synthesizes reviews, characteristics, and fit. This is the bucket closest to discovery that Google describes for the UCP: the journey begins before the purchase. On January 11, 2026, Google also announced dozens of Merchant Center attributes designed for conversational discovery (AI Mode, Gemini, Business Agent): answers to frequently asked questions, compatible accessories, substitutes. This is a feed orientation. This is not proof that your current CSV is sufficient. Nor is it a directory badge.
A search agent that only has your header visual has nothing to synthesize. Material, cut, return window, size, compatibility: it's either a field, or it doesn't exist for the machine. The job: what a machine needs to compare.
Rufus synthesizes in Amazon. Sparky in Walmart. Daydream in an affiliate fashion graph. Gemini can search on a Google surface, then, if you are eligible for the UCP, transition to a checkout funnel. Four "searches." Four catalogs. A non-Amazon merchant "optimizing for Rufus" is fixing the wrong feed.
Business Agent, announced on the same day as the UCP, is a brand chat in Search, using the retailer's voice, activated via Merchant Center in the US. This is not a fifth directory filter. It is just another Google surface. Do not merge it with "product search" as if Lowe's in the pilot were your integration.
After the purchase, the UCP also speaks of support. The directory has no customer service bucket. Chargeback reason codes do not say "the agent misread the size." The gap: where agentic commerce still breaks.
Price comparison: a true offer, or two prices
Comparing prices means reading a price, a currency, a stock status, a stable identifier. If the index says 49 and the product page says 59, the buyer clicks, then gets angry at you. The agent, however, has "compared". You will pay for the customer service.
Rufus compares within the Amazon ecosystem. Phia promises more than 40,000 sites: that's a pitch, not an audit. A "multi-merchant" LLM compares whatever it manages to extract: markup, feeds, or HTML. Three levels of data quality. A single word "price" in the filter.
In the post about the UCP, Google is not selling an independent comparison engine. It is selling an eligible listing, and then eventually a checkout funnel on that listing. The price that matters is the one from the Merchant Center and the offer. A promo code applied to the cart afterwards, invisible in the feed, will not travel. A Direct Offer, on the other hand, is an advertisement. Different bucket, further down.
Two diverging CSVs (Merchant Center, product page, and marketplace) mean three prices for three agents. Parity first. Comparison engine second. Platforms: what platforms must change.
Do not budget to "be the cheapest among the 18 price agents". You do not have their matching rules. You have an offer price, stock status, a GTIN. This is already the lever that survives the machine. The "-20%" badge in CSS, however, does not.
Bargain hunting: coupon, alert, or Google ad
Bargain hunting in the directory mixes three different jobs.
Coupons at checkout: Honey, Karma: an extension searches for a code and applies it. This is your promo engine plus JavaScript. It is not the UCP.
Alerts and fluctuating prices: a tracking tool that monitors. You often have neither the query nor the pixel. The buyer returns when the price drops.
Exclusive offer in an AI chat: Direct Offers (Google Ads, January 11, 2026): pilot, "Sponsored deal" label, e.g., 20% in the ticket. This is advertising. Not Honey.
Google says it is collaborating with Petco, e.l.f. Cosmetics, Samsonite, Rugs USA, and Shopify merchants to shape the pilot. Launch list. Not your addressable market. The useful mechanism: a discount that exists as campaign data can appear in AI Mode. A discount that only exists in a banner, cannot. Bundles and free shipping are announced as extensions, not as a module you configure this week.
Lumping Direct Offers and Honey together into an "AI bargains" metric is like grouping a Google bid with a coupon extension. One is paid for as advertising. The other is endured at checkout, sometimes cannibalizing a margin you had already planned for. Measure them separately. The directory puts them in the same bucket because a human says "I want a bargain". Your P&L, however, tells a different story.
Daydream in this bucket is affiliate fashion discovery, not a coupon engine. Sparky is Walmart. A Walmart "deal" is not a Madewell coupon. Do not build a single "agent bargains" connector.
Checkout funnel: three mechanisms under one badge
The directory phrase for the checkout tunnel: fill in address and payment, confirm on behalf of the user. Three actual mechanisms carry this label.
Checkout tunnel assistance: the agent clicks your theme (Claude, Comet, Operator, often Honey). Web application firewall, JavaScript, 3-D Secure. The human is still in the loop, or right next to it.
Automatic purchase on a designated rail: Gemini via UCP on eligible American product listings, Google Pay / Wallet; the merchant remains the seller of record, says Google. ChatGPT via ACP: sessions, delegated token, webhooks, says Stripe. Two integrations.
Other contract glued to the term checkout tunnel: Klarna (payment schedule). Shop (Shopify checkout tunnel). VendeeX (identity, limits). Amazon Buy for Me (Amazon payment page, potentially off-Amazon). This is not UCP.
ACP, as Stripe presents it, is used to complete purchases on behalf of buyers. The blocks are agent tunnel, cart and flow, delegated payment, OAuth, orders. This is the checkout tunnel bucket, not the deals bucket. Implementing ACP does not enroll you in Honey. It does not put you in Rufus. ChatGPT is the first agent platform mentioned on the ACP side. Each AI platform, in this ecosystem, manages its own integration.
UCP, on the other hand, wants a common language for the entire journey, not just the checkout tunnel. Yet, the first public deliverable is a checkout tunnel on AI Mode and Gemini, in the United States, on eligible product listings. Discovery and purchase can share a Google interface. They remain two different jobs: being findable, and then collecting payment. The specifications are compared here: how the protocols differ.
The Gemini directory entry says "Buy for me agentic checkout". Amazon has a separate product, Buy for Me, in the Shopping app. Two similar names. Two contracts. Do not merge them into a single "Buy for me" slide.
An agent that fills out your checkout tunnel and fails at 3-D Secure is not "worse" than Gemini via UCP. It does not have the same contract. The former needs a storefront that holds up: a Shopify storefront readable by an LLM. The latter needs a feed and eligibility. Correcting the wrong contract because the directory put both under checkout tunnel is the trap of this page.
One agent, three buckets, is not three channels
ChatGPT, Gemini, Copilot, Claude, Shop: several buckets. It is not four channels to turn on. It is a product that serves several human use cases, poorly labeled for a budget.
If you read | You risk | Instead, read |
|---|---|---|
Four filters = four projects | Opening a UPC request, a coupon request, a review request, a price request | A clean feed serves search and price. The coupon is from the existing tunnel. The rail can wait |
14 checkout tunnel agents = 14 providers | Comparing Honey and the APC | Mechanism: theme click, mandate, split payment, marketplace checkout tunnel |
Direct Offers = bargain hunting | Budgeting advertising like an extension | Sponsored deal = advertising. Honey = coupon on click |
A pitch like "we will put you in the checkout tunnel agents" can sell a more readable theme, an integration to the APC, a Shopify Agentic button, a directory listing, or Direct Offers. Five deliverables. One word. Ask which one. The directory you are reading is the listing, not the first four.
Protocols do not map one-to-one to filters. The UPC: discovery, purchase, after-sales. The APC: delegated purchase. Neither has a "coupon collection" building block. The bargain hunting in the directory is a buyer use case, not a 2026 specification.
Three actions, not 67 tablets
This week, by real task, not by directory counter.
Search and pricing: feed / product sheet / marketplace parity on price, stock, variant, identifier. Attributes that can be cited (material, fit, Q&A). Freshness aligned with actual variation.
Deals: separate coupon codes (promo engine, checkout JavaScript) and paid offers like Direct Offers. Do not merge into an "AI deals" cost per acquisition.
Checkout: name the mechanism: click on your checkout, UCP or ACP rail, marketplace, installment payment. Test a product sheet and a 3-D Secure. Bot policy aside.
The unique test: would this work still be useful if the directory relabeled everyone tomorrow? A GTIN and a real price, yes. An exclusive connector "for the checkout filter", no.
Name owners by actual task, not an "agent owner by task". Merchandising for attributes. Promo for codes. Payments for the mandate. Security for the headless browser. A single owner stops at the first web application firewall that blocks.
If you only have one hour: open five header references and ask if a human can describe them without the JPEG. Open the checkout and apply a real code. Ask the web application firewall what a headless Chrome receives. These three actions are worth more than reading the 67 badges.
You don't turn on a task like a channel
You don’t turn on "the search task" like a Shopify Agentic channel. Buyers open ChatGPT, Gemini, Rufus, Honey. You maintain a catalog, an entry point, a checkout funnel. Directory filters are not buttons.
If a vendor promises to place you "in price comparison agents," ask: Merchant Center feed, marketplace listing, on-page markup, indexing, or directory listing. Five deliverables. One word: "price."
UCP and ACP are chosen late, and only if the collection mechanism is truly a mandate. For a Honey coupon, the protocol is already your checkout funnel. For Rufus, it is the Amazon feed. Postponing the rail is not postponing the feed.
After the click or the mandate, it is your dispute, your return, your 3-D Secure. The directory is no longer in the funnel. Do not attribute a chargeback to it that it did not collect.
Three cards, only one that you hold
Usages are fragmenting. Google did not "replace" the four filters with the UCP. The UCP names a journey (discovery, purchase, after-sales) and an initial payment funnel in the United States. The ACP names a delegated purchase. The directory names four buyer habits. Three maps. Only one that you truly control: your data and your payment funnel.
Treat this page as an index: which habit is already hitting your stack (marketplace, coupon, LLM surface), which payment funnel mechanism, which field is missing. Three questions. Not a ranking by task.
You can ignore the bargain bucket and still maintain a true price. You can refuse the ACP and still survive Honey. You can be outside the UCP and still sell via Rufus. The directory does not decide. Your attributes, your web application firewall, and your eligibility, do.
Next, open a product page, not twenty filters. A surface with a public rail (ChatGPT or Gemini). A marketplace where you already sell. A payment funnel assistant that touches your theme. Each time, lay down the mechanism, not the badge.
Frequently Asked Questions
What is the best agent per task in 2026?
This page does not say. No public data on volume, accuracy, or disputes says so either. The counters 18 / 16 / 19 / 14 are labels. A ranking would be an invention.
Why is the same agent in three buckets?
Because the directory labels human use cases, not integrations. A LLM that searches, compares, and assists the checkout funnel gets three tags. This does not represent three commercial projects.
Which task do UCP and ACP correspond to?
ACP, as documented by Stripe, is a protocol for completing a purchase on behalf of the buyer: the checkout funnel bucket, in the mandate sense. UCP, as launched by Google, covers a journey (discovery, purchase, after-sales) with an initial checkout funnel on AI Mode and Gemini in the US. Neither is a coupon protocol.
Does Direct Offers fall under bargain hunting?
In buyer vocabulary, yes: a discount at the moment the query signals intent to buy. In the contract, no: it is a Google Ads pilot, labeled Sponsored deal. Honey is a coupon extension. Keep them separate.
Are Klarna or Honey checkout funnel agents?
In the directory, they carry the tag. A split payment and a coupon affect your existing checkout funnel. It is not a UCP or ACP mandate. Keep them separate in the P&L even if the grid mixes them together.
What to do this week?
Price and attribute parity. Tested promo codes. Web application firewall and 3-D Secure on a real product page. Postpone the protocol as long as the collection mechanism is not a mandate. Do not budget four channels across four filters.
Is an "agents by task" owner needed?
No. A feed owner, a promo owner, a payments owner, a bots owner. The directory filters are not an organizational chart. If a single person owns all four, work stops at the first silo.
Sources
Enzo
August 19, 2026
