How to integrate with Instinct
There are no developer docs for this. Instinct, the AI shopping assistant, launched to invites in August 2026, raised a billion dollars at a ten billion valuation a month later, and it works by opening your website and using it like a person would. Straight up to the Pay button, then a human confirms.
Nothing to sign up for, then. Integrating with Instinct means two pieces of work instead, and agentic commerce only becomes a payments question in the second one. (If you're on Shopify and also wondering about Meta's assistant, how to integrate with Muse has a different answer: you're probably already in it and didn't notice.)
One: make the site usable by something that can't improvise
Higher bar than "the website works". Lower bar than building an API.
Every product needs its own address, showing its facts immediately. Not after a click, not on hover, not in a modal. Anything that only appears when a human interacts with it may as well not be there.
Facts as text rather than pictures: price, stock, options, shipping, delivery time, returns window, sizes, materials. Nearly half the shops we scan are missing machine-readable product data, and half hide shipping cost until checkout. Hiding it doesn't protect your margin, it just takes you out of the comparison.
Then the checkout, which needs to work without guessing. Ambiguous address fields, errors that don't say what's wrong, a required box that only appears once you've filled a different one, a bot challenge on add-to-cart. Each one is mild friction for a person and a dead stop for a robot.
And policies as findable text. Returns, shipping, warranty, how to reach you. 36% of shops we scan have a returns policy that's incomplete or hard to find.
Then go and buy from your own shop while refusing to be clever. Fill only what's required, never guess at a label, never re-read an error you already passed. Where you stop is where the robot stops.
On Shopify, the readability half of that list is what our free app checks and fixes. The checkout test you should still do yourself, because it's the part that surprises people.
Agentic commerce, when the machine is the customer
Here's the part that actually needs MPP, and the part most articles blur.
MPP is the Machine Payments Protocol, released by Stripe and Tempo on 18 March 2026. It lets a robot pay for something without a human entering a card. It works on an allowance: the robot gets a spending limit up front, then draws against it in small amounts, settling in normal currency or stablecoins.
It was designed for what robots buy on their own account. An API call. An hour of compute. Access to a dataset. A small recurring charge.
One question decides whether any of this is yours:
Is the robot buying for a person, or for itself?
Buying for a person, someone wanted a coat and sent their assistant, and your job is readable data plus a checkout that completes. A human approves the payment. MPP doesn't come into it and adding it wins you nothing.
Buying for itself, the robot needs something from you to finish its work. Your signup form assumes a human with an inbox and a credit card, and that assumption is now a wall between you and a customer who can't climb it.
The second case is worth sitting with longer than most merchants do. The test isn't "am I an API company". It's "could a machine be my customer". A data provider, a research service, a monitoring tool, an image library, a verification service: each has a robot-shaped buyer who can't currently buy, and a pricing page written for a human who'll never read it.
What building for machine customers means in practice
Publish what you sell and what it costs in a form a machine can parse, because it can't work out your pricing from marketing copy.
Sell an allowance rather than a cart. Limit up front, spend as you go, settle after.
Price in small units: per call, per row, per second. A 49-a-month plan is a human invention, and a robot weighing one job against it will just leave.
Let it prove who it is without a browser. Keys or signed requests. No email confirmation loop, no CAPTCHA.
Handle retries safely. Robots retry, and a double charge a person would spot and dispute can sit unnoticed for months, which is worse than the charge.
Publish your limits: how fast, how much, and why you'd refuse, in a form something can read and respect.
MPP is one of several here. ACP does AI checkout, Google's AP2 does permissions, x402 does payments over the web. Pick by the problem you've got, not by what's loudest this week.
Where these numbers come from
Every percentage on this page comes from our own scanner, run against the last 220 ecommerce sites we audited before 2026-09-28. It's an automated pass over live pages: what loads, what's in the markup, what a crawler can and can't reach. Not a survey, not an estimate, and not a sample we picked to make a point. Where a check didn't apply to a site, that site isn't counted in its denominator, which is why the totals differ per row.
We'll re-run it as these assistants change. If a number here is stale by the time you read it, tell us and we'll fix it.
Which problem do you actually have
Nearly everyone reading this has the first one. It's also worth more money.
- Can a robot read your products? 48% of shops we scan: not properly
- Can it tell who you are? 48%: no company tag
- Can it find shipping and returns? 50% and 36%: no
- Can it finish checkout unaided?
- Would you notice if one visited? 46%: broken analytics
- Could a machine be your customer, and can it pay you?
Six only matters once one to five are true. A robot that couldn't read your product page never reaches your payment system.
Not sure if you've got problem one or problem six? Twenty minutes will tell you, and guessing could cost you a quarter. Book a call and we'll work out where your shop actually sits and whether machine customers are real for your business or a distraction. On Shopify, install the free app first and bring the results.