Skip to content

Everyone builtthe checkout.Nobody built the shelf.

A pair of balance scales

AI agents send stores the best-converting traffic in retail. Then they quote a price that’s wrong, a delivery date they invented, and a cake that sold out at noon. Caravanserai is the one integration that lets a store tell every agent what is actually true — and stand behind it.

Two things are true at once, and both are in the public record.

+54%

Agent-referred shoppers convert better than every other channel a retailer has. This is the best traffic in retail right now.

Adobe — source

Adobe Analytics
May 2026

See the source
<1%

And there are almost none of them. That is Etsy's agentic traffic as a share of all traffic, said out loud on an earnings call.

Etsy — source

Etsy Q2 2026 earnings call
12 Aug 2026

See the source
30

Merchants who ever went live on ChatGPT Instant Checkout, against the million promised. OpenAI retired it in March.

Forrester — source

Forrester
Feb 2026

See the source
33%

What in-chat checkout converted at, against a plain link to walmart.com. Walmart measured it, said so on the record, and pulled out.

Walmart — source

Daniel Danker, EVP Walmart, at Morgan Stanley TMT
4 Mar 2026

See the source
0%

Independent shops publishing a delivery promise a machine can act on. Not few — none, in 946 shops we could read. 48.5% say they deliver, in words on the page, which an agent cannot book against.

Caravanserai — source

Caravanserai, The Agent Readability Index v1
1 Sep 2026

See the source

Agent-referred shoppers are the highest-intent, best-converting, highest basket-size traffic a store gets, and there are almost none of them. The gap between those two facts is not consumer appetite and it is not payment infrastructure. It is that stores cannot answer.

Marks identify the source of each figure. Attribution, not endorsement. Every figure is listed with the document behind it in the receipts, and the fifth is our own fieldwork.

Interoperability

We compete with none of these.We speak all of them.

Payment was never the broken part, and every one of these already works. A store integrates once, with us, and answers every agent that arrives — whichever rail it came in on.

Rails an agent can arrive on

Stripe · OpenAIStripe · OpenAI

ACP

Agentic Commerce Protocol

Open spec, Sept 2025

Google

AP2

Agent Payments Protocol

Donated to the FIDO Alliance

Shopify · Google

UCP

Universal Commerce Protocol

Open standard, Jan 2026

Anthropic

MCP

Model Context Protocol

Storefront servers on every Shopify store

Visa

TAP

Trusted Agent Protocol

10+ partners, Oct 2025

Mastercard

Agent Pay

Agentic Tokens

Agent Pay for Machines, June 2026

Coinbase · CloudflareCoinbase · Cloudflare

x402

HTTP 402 settlement

x402 Foundation, Linux Foundation, Apr 2026

Tills a store already runs

Square

Square

Toast

Toast

Clover

Clover

Lightspeed

Lightspeed

WooCommerce

WooCommerce

One integration, every agent — not one integration per AI. That is the whole reason a store with eleven staff can be read by software it has never heard of.

Marks identify the publisher of each protocol and the systems a store already runs. Attribution, not endorsement.

“Crawling and scraping is inadequate to get the full breadth of product data that you need to do a good job of commerce.”

Emily Pfeiffer, Principal Analyst, Forrester

In September 2025 OpenAI and Stripe shipped a protocol and a checkout, and promised a million merchants. By February thirty had gone live. In March it was retired. Every named reason it failed was a merchant-side data problem, and not one of them was payment.

A sieve
  1. 01Stock status went stale.
  2. 02Delivery dates were outdated or invented.
  3. 03Shipping costs were wrong, or missing entirely.
  4. 04No multi-item cart.
  5. 05No sales tax collected or remitted.
  6. 06Merchants would not hand over loyalty, payment data, or the customer after the sale.

Read that list again. It is not a list of things an AI cannot do. It is a list of things a store never told anyone. So it is a product specification.

Two hands passing a blood orange across a crate of fruit

A store is a promise about what is behind the counter. Nothing about that changed when the customer became software.

Photograph · Ron Lach

Illustrative transcript

Today

The agent guesses, then gives up.

I need a birthday cake for Saturday, near me, under $60.
AgentSweet Ash Bakery on Fulton has a 8" vanilla layer cake for $48. Want me to order it?
Yes.
Agent · dead endI can't complete that here — you'll need to call them. Their site doesn't show availability.

This is the ordinary outcome. The store is real, the demand is real, and nothing about the store is legible to the thing being asked.

With Caravanserai

The store answers, and means it.

I need a birthday cake for Saturday, near me, under $60.
AgentSweet Ash Bakery on Fulton: 8" vanilla layer, $48 plus $6.50 delivery. They have 3 Saturday slots left and can deliver to your address between 10am and noon. Held for 15 minutes.
Take it.
AgentOrdered. Sweet Ash confirmed. Saturday, 10am–noon, $54.50. Cancel free until Thursday 6pm.

Same bakery, same question. The difference is a signed quote the store agreed to honour before the agent ever repeated it.

One integration. Every agent. A promise the store agreed to keep.

  1. 01

    Connect the till

    One integration with the system the store already runs — Square, Toast, Clover, Lightspeed, WooCommerce. No replatforming, no new hardware, and your checkout stays yours.

  2. 02

    Publish what's true

    We read live price, stock, fees and fulfilment windows and publish them in the formats agents already speak. One integration, every agent — not one integration per AI.

  3. 03

    Sign the promise

    An agent asks; the store answers with a signed, time-boxed quote. This price, this many left, delivered here by then. An agent can repeat it to a person without lying.

  4. 04

    Stand behind it

    The store honours the quote or it doesn't, and we are the only party who sees the promise and the outcome. Price, stock and cancellation read back from the till; delivery is attested by the buyer or the agent, and counted apart. That record becomes the reason agents send the next customer here instead of somewhere else.

A quote, not a transaction.

Signed, addressed, and good for the next fifteen minutes. An agent can repeat every line of this to a person without inventing anything.

Signing…Expires in --:--
  minting and signing a quote…

Signed in your browser and verified in your browser against the key below. Change one cent and it breaks. That is the difference between a price an agent scraped and a price a store will honour.

Signer key
ed25519:b7b2…0189
Curve
Ed25519 over the canonical body (sorted keys, no whitespace)

Reference implementation. Real signature, illustrative merchant — Sweet Ash Bakery is not a customer. Signed with a disposable demo key.

The window is open because the first attempt just failed, and it starts closing on the fifteenth of September.

15 Sep 2026

Scraping becomes permissioned

Cloudflare has told AI companies to split their crawlers into search, training and agent. From this date it blocks the training and agent ones by default on ad-bearing pages, for free customers and every new site. It is not a total shutdown — a bakery's website carries no ads. But the machinery for deciding which bots get in is now built, on, and defaulted to no. Cloudflare customers blocked 416 billion AI bot requests in five months.

Mar 2026

Shopify closed its own door

Agentic Storefronts switched on by default for eligible US merchants, syndicating catalogs into ChatGPT, Google AI Mode and Copilot — with checkout staying on the merchant's own site. Free. Shopify solved this for Shopify.

1 Jul 2026

Square did it for its restaurants

A ChatGPT app and a Claude plugin. US food-and-beverage sellers with an active Square Online Ordering profile were opted in automatically — orders route into Square Online Ordering, the POS and the kitchen display, paid by Order by Cash App, no commission. It is a catalog inside two assistants: no addressed commitment, no record of what was honoured, nothing for a seller off Square. Two platforms have now sorted this market by till rather than by merchant.

Mar 2026

The first attempt failed in public

Instant Checkout retired. Atlas retired four months later. The field is clear in a way it was not six months ago and will not be in twelve.

Now

Both numbers are true at once

Agent traffic is the best-converting channel in retail and under one percent of volume. That is only briefly true. It is the whole window.

A cardboard parcel left beside a white door

The whole thing is only worth building if the box turns up on the day it was promised.

Photograph · Tima Miroshnichenko

We are picking the first stores now.

A small number of merchants whose shelves change by the hour, and the people building the agents that have to read them. We do the integration ourselves, against the till you already run.

I run a store.

Tell us what you sell and what you run the till on. You do not write code, you do not replatform, and you do not change checkout — we connect to the system you already have.

One email when there is something real. Nothing else.