Do consumer AI agents need another protocol?

Updated November 16, 2026

Key findings

2x fasterMuse answered the same shopping queries twice as fast on an agent storefront as on the original website
80%of queries were faster for Muse on the agent storefront. The rest were level, and none were slower
0%of agent sessions needed a new protocol. Every one got through by browsing or with a protocol that already exists
3 routesFive agents entered the same storefronts three different ways, and none of them ever switched
62%of sessions used the store's UCP endpoint. 19% used WebMCP inside a browser, and 19% used no protocol at all
71%of sessions read the store's pages at some point. Even among the sessions that used a protocol, 64% still opened product pages

Source: Aigency agent route study, November 2026

The short answer

tl;dr: No, not for the shopping itself.

Computer use models + the good ol' browser can get the trick done for most.

We sent five consumer AI agents shopping on our storefronts, and every session got through by browsing or by using a protocol that already exists.

The agents did not agree on a route: 62% of sessions used the store's UCP endpoint, 19% used WebMCP inside a browser, and 19% used no protocol at all.

The one real gap is identity and permissions, which is a different layer from shopping.

What changed the result most was the site: on a storefront built for agents, an agent answered the same queries about twice as fast, with no protocol involved.

Why ask whether agents need another protocol?

A store that wants to be ready for AI agents is now offered a long menu. There is MCP for exposing tools, Google and Shopify's Universal Commerce Protocol (UCP) for catalog and checkout, WebMCP for tools that live inside the page, the store's own API, and discovery files such as llms.txt that tell an agent where all of that lives. Underneath sits the oldest route of all: the agent opens the site in a browser and reads it the way a person would.

Meta and Sierra have now added the Personal Agent Protocol, known as Poppy, with partners that include Instinct, Shopify, Stripe and Walmart. Poppy is about how a personal agent identifies itself to a business and what that business allows it to do. It is described as complementary to commerce protocols such as UCP, and it was the subject of a recent interview in The Aisle.

The control problem Poppy goes after is real: a store cannot easily tell a person from an agent acting for one. It also raises the question a merchant actually has to answer: with this many routes already on the table, which ones do agents use when nobody tells them what to do? We did not have to guess. We run agent sites, so we pointed five consumer agents at two of them, gave them ordinary shopping tasks, and asked each one afterwards exactly how it got its answer.

What we tested

Our researchers ran shopping sessions with five consumer AI agents: Instinct, Grokbot, Hark, Muse (Meta's agent) and Dots (the ChatGPT agent). The agents were sent to two of our own e-commerce storefronts. Every route was open on both: rendered pages a browser can read, a UCP catalog endpoint served over MCP, WebMCP tools inside the page, and discovery files pointing to all of it.

The tasks were the top of the funnel, where agents spend their time today: find a product that fits a described need and a budget, compare a few options, and recommend one. This is the assisting role described in our guide to agentic commerce: the agent does the looking and the person decides. No task asked an agent to pay for anything. After each answer we asked the agent for an audit of every page it opened and every tool or interface it used. We then checked those audits against our own server logs, and they matched.

  • InstinctUCP catalog endpoint, plus product pages
  • GrokbotUCP catalog endpoint, plus product pages
  • HarkUCP catalog endpoint only
  • DotsChatGPTA browser, using the page's WebMCP tools
  • MuseMetaA browser, with no protocol
The five agents and the route each one took

Five agents, three routes, and nobody switched

The headline result is that no single route won. Three agents went to the UCP catalog endpoint. One opened a browser and used the tools the page offered through WebMCP. One opened a browser and used nothing but the page.

Agents split across three routes

  • UCP endpoint
  • WebMCP in a browser
  • No protocol
  • All sessions

    UCP endpoint 62%, WebMCP in a browser 19%, No protocol 19%

Share of all sessions by the route the agent took to read the store. Agents' own audits, checked against our server logs.

The more telling result is what happens when the sessions are split by agent. Every agent used its own route in 100% of its sessions. The task changed, the product changed, and for one agent the storefront changed, and the route did not.

Each agent stayed on one route in every session

  • UCP endpoint
  • WebMCP in a browser
  • No protocol
  • Instinct

    UCP endpoint 100%

  • Grokbot

    UCP endpoint 100%

  • Hark

    UCP endpoint 100%

  • Dots

    WebMCP in a browser 100%

  • Muse

    No protocol 100%

Share of each agent's own sessions by route. Instinct also made a successful WebMCP call alongside its usual route. Muse's figure covers the sessions with a complete audit.

This was a choice, not a limitation. Grokbot and Dots each run on a computer with a full browser, and they went opposite ways: Grokbot never opened one, and Dots did everything through one. Instinct said in its first audit that it went straight to the endpoint because that was what had worked before. The store offered every route on every visit and it made no difference to which one an agent took.

Two details are worth a second look. The agent that used no protocol at all is Muse, built by Meta, one of the two companies that wrote Poppy. And Instinct, a Poppy partner, went straight to a shopping protocol that predates it. The companies closest to the newest protocol are, today, shopping with a browser and with UCP.

Protocols did not replace the page

Counting which protocol an agent used undersells how much plain reading went on. Most of the sessions that used a protocol also went and opened product pages, because the endpoint gave them a shortlist and the page gave them the detail they needed to recommend something.

Fewer than a third of sessions relied on a protocol alone

  • Protocol only
  • Protocol plus product pages
  • Browsing only
  • All sessions

    Protocol only 29%, Protocol plus product pages 52%, Browsing only 19%

Share of all sessions by how the agent read the store.

Put the other way round, 71% of sessions read the store's pages at some point, and among the sessions that used a protocol, 64% still opened product pages. The protocols worked as a fast way to find candidates. They did not remove the need for a page that an agent can read.

What the sessions did, and what none of them needed

  • Used browsing or an existing protocol
    100%
  • Used a protocol that already exists
    81%
  • Opened a browser at some point
    50%
  • Went through a browser from start to finish
    38%
  • Read a discovery file first
    31%
  • Needed a protocol that does not exist yet
    0%
Share of all sessions. A session can appear in more than one row.

That last row is the answer to the question in the title, as far as shopping goes. Nothing an agent tried to do in these sessions was blocked for want of a new standard. Where an agent wanted structure, the structure was already there. Where it wanted to look for itself, it looked.

Browsing, checkout, identity: three layers, one real gap

"Do agents need another protocol" is really three questions, because an agent does three different jobs on a shopper's behalf. Lumping them together is how the debate gets muddled.

The three layers of an agent's shopping trip
LayerWhat the agent doesDoes it need a new protocol?
BrowsingFinds, compares and reads productsNo. Browsing and the existing protocols covered 100% of our sessions.
CheckoutCompletes the purchaseNot where the agent hands the shopper to the retailer's own checkout.
Identity and permissionsSays who it is and what it is allowed to doThis is the open layer, and it is the one Poppy addresses.

Our sessions measured the first layer only. For the second, the clearest public summary we have seen is from analyst Juozas Kaziukėnas, who tracks what each assistant can do. His November update describes the agent that used no protocol on our sites:

Muse has shopping carts but they hand off checkout to the retailer.

Juozas Kaziukėnas, commerce analyst · X

A handoff needs no payment protocol at all: the agent brings the shopper to the door and the retailer's checkout does what it already does. The same post says Muse uses UCP to talk to Shopify stores and is working on preapproved purchases, so this is where things stand today and not a settled design. It also reports Gemini's shopping carts running on UCP. Where agents do transact, they are doing it on a protocol that already exists.

That leaves the third layer, and here the case for something new is real. A browser session cannot tell a store whether it is talking to a person or to an agent acting for one, or what that agent has been authorized to do. That is why we think an identity and permissions standard such as Poppy, or connectors a retailer sets up directly with an assistant, earns its place. It is a different problem from reading a catalog, and solving it does not require a new way to read a catalog.

The site made the difference, not the protocol

If the route is the agent's choice and the protocols are optional, what does a store control? The quality of what the agent lands on. We had a clean way to measure that, because Muse browses like a person and uses no protocol. Our researchers gave Muse the same shopping queries on both: our agent storefront, and the original website for the same catalog. Each query was its own session, and ten of them are charted below.

2x fasterMuse answered the same shopping queries twice as fast on an agent storefront as on the original websiteAigency agent route study, November 2026

Muse answered the same queries about twice as fast on the agent storefront

  • Agent site
  • Original website
  • All charted queries
    47.5 min
    97.5 min
Total approximate minutes across the charted queries, from sending the prompt to the answer, including queue delay. Each query was a separate session.

Faster on most queries, slower on none

  • Agent site
  • Original website
  • Query 1
    3 min
    3 min
  • Query 2
    4 min
    10 min
  • Query 3
    3.5 min
    7.5 min
  • Query 4
    2 min
    8 min
  • Query 5
    6 min
    21 min
  • Query 6
    4 min
    10 min
  • Query 7
    1 min
    2 min
  • Query 8
    6 min
    12 min
  • Query 9
    12 min
    12 min
  • Query 10
    6 min
    12 min
Approximate minutes per query. Where Muse's time fell in a range, the midpoint is shown.

Muse was faster on the agent storefront in 80% of the queries, level in the rest, and slower in none. The largest gap was a query that took 21 minutes on the original website and 6 on the agent storefront. Query 4 was flagged for an issue on Muse's side; leaving that query out, the saving is 49%. It also opened 16% fewer pages to get there.

Fewer pages, less inference, less load

Speed is the visible part of that result. The less visible part is work. Every page a browsing agent opens is a page its model has to read, and every click is another round of reasoning about what to do next. Muse reached its answers on the agent storefront in about half the time and with 16% fewer pages opened, which means it spent less inference on each answer.

We did not meter tokens, so we are not putting a number on the inference saved. The direction follows from the figures we do have, and it matters on both sides of the visit. For the company running the agent, a store that takes fewer steps to read is cheaper to serve and leaves more room for smaller, faster models. For the store, each agent visit asks its servers for fewer pages, on traffic that is only going to grow.

No protocol was involved in any of that. The agent used a browser both times. The only thing that changed was whether the pages were built to be read by an agent: answers stated plainly, product facts in one place, and nothing that makes a browsing agent wait or hunt. That is what we mean by an agent gateway, and it is the part of this that a store can decide on its own.

What this means for a store

  • You do not pick the route. The agent does, and five agents picked three. A store that is ready on one route only is ready for some of its agent traffic.
  • The page still counts. 71% of sessions read pages, including most of the sessions that had a protocol to lean on. An endpoint on top of a page agents struggle with fixes half the visit.
  • Support what exists before adopting what is new. Browsing, UCP and WebMCP covered every session. None of our sessions would have gone better with a fourth shopping protocol.
  • Watch the identity layer. Knowing which agent is at the door, and on whose behalf, is the part that is not solved, and it is worth following Poppy for that reason.

For how the individual assistants shop, see AI shopping agents, and for the vocabulary used here, the agentic commerce glossary.

Frequently asked questions

Do AI shopping agents use protocols or do they just browse?

Both, depending on the agent. In our test sessions, 62% used the store's UCP endpoint, 19% used WebMCP tools inside a browser, and 19% used no protocol and simply browsed. Each of the five agents stuck to one route in all of its sessions, and most sessions that used a protocol also opened product pages.

Is a new protocol needed for AI agents to shop a store?

Not for finding and comparing products. Every session in our test got through the store by browsing or with a protocol that already exists. The open problem is identity and permissions, meaning how an agent says who it is and what it may do, which is what the Personal Agent Protocol (Poppy) is designed for.

What is the difference between UCP, WebMCP and Poppy?

UCP, the Universal Commerce Protocol from Google and Shopify, lets an agent query a catalog and, where supported, check out. WebMCP exposes tools inside a web page for an agent that is already browsing it. Poppy, the Personal Agent Protocol from Meta and Sierra, covers how an agent identifies itself to a business and what the business permits. The first two are about shopping and the third is about identity.

Does a store have to support every agent protocol?

A store cannot choose which route an agent takes, so being ready on only one route covers only some agents. In our test three agents used UCP, one used WebMCP and one used neither. The common requirement across all five was a store whose pages and data an agent can read quickly.

How much faster is an AI agent on a site built for agents?

In our test, Muse answered the same shopping queries about twice as fast on an agent storefront as on the original website, taking 51% less time, while using a browser and no protocol both times. That is one agent on a storefront Aigency built, so treat it as an indication and not a benchmark.