[thoughtful] AI Change Desk | Episode 17: Merchant Control Check

If you run ecommerce, or growth, or digital product, or honestly just anything adjacent to online buying right now, there is a new category of sentence you are going to hear a lot this year.

It sounds something like this:
Good news, we showed up in the AI answer.

And, look, that might be good news.

It might also be the start of a much more annoying conversation.

Because the question is not just whether you showed up.
The question is what exactly happened after that.

Who supplied the product data?
Who controlled the comparison?
Who owned the checkout?
Who got the attribution?
Who is on the hook when the product title is wrong, the discount does not apply, the loyalty benefit disappears, or the customer support team suddenly gets a ticket that began in a surface nobody has really been monitoring?

That, I think, is the real story now.

Back in Episode 14, we talked about AI shopping as a discovery surface.
The point then was simple.
Do not confuse visibility with conversion.
Do not confuse the model mentioned us with this is working.

This week I want to tighten that question.
Because now we have a clearer picture of what these systems are becoming.
OpenAI is leaning into product discovery in ChatGPT while pulling checkout control back toward merchants.
Shopify is turning AI shopping into something merchants can actually administer from a real platform, with real settings and real attribution.
And Google is pushing shopping deeper into personalization and protocol infrastructure at the same time.

So this is not really an AI shopping is here episode.
We did that already.
This is a merchant control episode.

And yes, I know, merchant control sounds like the kind of phrase created by a very sincere committee in a conference room with bad coffee and excellent lanyards.
I know.
Unfortunately, it is also the real job now.

__INTRO_MUSIC__

Welcome back to AI Change Desk.
I'm Michael.

If you have been with me over the last few episodes, there is a continuity line here that I want to make explicit.
Episode 14 asked whether AI shopping visibility meant anything.
Episode 15 asked who owns the artifact after the chat.
Episode 16 zoomed out and asked what happens when AI starts looking like infrastructure, security, and national capacity instead of just product rollout.

This week brings all of that back down to the customer edge.
Once a journey starts in an AI surface, who actually owns the sale, the attribution, and the policy that follows?

That is the question.
Not the screenshot.
Not the demo.
Not the keynote sentence.
The question.

Let me say this as clearly as I can.
The new thing is not that people can discover products in AI interfaces.
That part is already obvious.
The new thing is that the architecture underneath discovery is getting more defined.

And when architecture gets more defined, operating questions stop being hypothetical.
They stop being sometime-later problems.
They become this-quarter problems.

Here is the simple version.

OpenAI is saying ChatGPT can be a place where people explore, compare, and figure out what to buy.
Shopify is saying fine, if that is true, merchants need a manageable way to participate, track, and keep their own checkout and customer relationship intact.
Google is saying this is going to be more personalized, more connected, and more protocol-driven than traditional shopping search.

That means the operator problem is getting sharper.
Not bigger in a vague sense.
Sharper.

It is no longer enough to ask, are we visible?
Now you have to ask:
Are our products represented accurately?
Are our policies carried over cleanly?
Can we reproduce the result in QA?
Can we isolate the traffic and measure it honestly?
Do we know whether this channel is helping us, or just flattering us?

And, look, I know this is the stretch in the episode where I start sounding like the least fun person in the meeting.
I get that.
But this is exactly how expensive confusion usually enters the building.
It does not arrive wearing a cape.
It arrives wearing a dashboard.

Before we get into the company moves, let me translate three phrases because these terms can get abstract fast.

First, merchant-controlled checkout.
That just means the customer may start in an AI surface, but the actual purchase still happens in the merchant's own checkout environment.
That matters because checkout is where a lot of real business rules live.
Taxes.
Payments.
Shipping logic.
Promo codes.
Returns language.
Fraud handling.
Loyalty.
Support expectations.
All of that.

Second, personalization.
That means the answer is not purely universal.
It can change depending on what the system knows about the user, what they asked, what they connected, what preferences the system inferred, and what context was available.

Third, protocol.
That is just the agreed way systems exchange information.
If AI shopping becomes more protocol-driven, that means more of the commerce workflow depends on structured data, compatible standards, and clean handoffs between platforms.

Which, to be fair, sounds very neat until one of the handoffs quietly drops something important.
Then it stops sounding neat and starts sounding like a support ticket.

So let us start with OpenAI.

On March twenty-fourth, twenty twenty-six, OpenAI published a product post called Powering Product Discovery in ChatGPT.
The framing matters.
It does not just say, look, richer shopping exists.
It says ChatGPT is becoming a place where people start shopping in order to explore, compare, and figure out what to buy.

That is a subtle but very important distinction.
This is not just a transactional story.
This is a discovery and decision-support story.

OpenAI says the new experience supports more visual browsing, comparison, and up-to-date information in one place.
The company also says it is expanding the Agentic Commerce Protocol to support product discovery, with more complete and relevant information coming directly into ChatGPT.

That alone would be interesting.
But the more important part, in my view, is the merchant side.

OpenAI says it wants to offer merchants options for how they convert consumers.
And crucially, it says the initial version of Instant Checkout did not offer the level of flexibility it wanted to provide.
So OpenAI is allowing merchants to use their own checkout experiences while it focuses efforts on product discovery.

That is a real design signal.
That is the platform basically saying the cool demo is not enough.
The actual commerce system has too many moving parts.
If merchants are going to take this seriously, they need more control at the moment of conversion.

Which is, honestly, a fairly adult conclusion.
And I mean that as a compliment.

Because this is what happens when novelty meets reality.
Reality says, okay, wonderful, the chatbot found the lamp.
Now tell me who owns shipping, who honors the coupon, how the tax is calculated, what happens if the item is out of stock, whether the customer is logged in, whether their rewards apply, and who gets yelled at when the answer implied something that turns out to be wrong.

That is the actual work.

The OpenAI help article adds another layer that matters for operators.
It says product results are selected independently by ChatGPT, are not ads, and are not influenced by OpenAI partnerships.
It also says the system uses the query and context, including things like memory or custom instructions, when deciding what to show.

Now pause there for a second.

That means two things are true at once.

One, OpenAI is drawing a line between ads and product results.
Two, the result set is still model-mediated and context-sensitive.

Which means if you are a merchant or operator, you should not confuse not ads with fully deterministic or easy to audit.
Those are different things.

The same help article says the model can simplify titles and descriptions to make results easier to read, and that some labels like budget-friendly or most popular are model-generated, not guarantees.
That is useful product behavior for users.
It is also the kind of thing that should make any serious merchant or merchandising lead sit up a little straighter.

Because now the representation layer matters.
Your product is not just being displayed.
It may be interpreted.
It may be summarized.
It may be grouped.
It may be compared.
And it may be contextualized in a way that is helpful for the user and mildly terrifying for anyone who has spent years trying to make product language, claims, and disclosures consistent.

And yes, I realize that is not the most romantic way to talk about shopping innovation.
But innovation without control is just a more stylish route to confusion.

There is another detail in the OpenAI post that is worth noticing.
OpenAI says retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, and Wayfair have integrated for discovery.
It also says Shopify Catalog is already integrated into ChatGPT and that no extra work is required from individual Shopify merchants for product data integration.

That is a real scaling move.
It suggests the ecosystem is trying to reduce the amount of custom integration work required for discovery.
But reducing integration work at the top of the funnel does not remove downstream complexity.
It just moves the pressure.

So the OpenAI story, operationally, is this.
Discovery is getting richer.
Product representation is getting more model-mediated.
Checkout is moving back toward merchant control.
And the clean line between product experience and commerce operations is disappearing.

Now let us bring Shopify into it, because this is where the story stops feeling conceptual.

Shopify's March twenty-fourth announcement is blunt.
It says millions of merchants can sell in AI chats, and it frames Agentic Storefronts as a way to manage participation from Shopify Admin.
ChatGPT is the clearest live example in the current packet, while surfaces like Microsoft Copilot, AI Mode in Google Search, and the Gemini app are better understood as earlier-access storefront channels, not something every store should assume is already live.

That is not just a market prediction.
That is platform packaging.

And once a platform starts packaging something, the operator questions get more concrete.

Shopify says hundreds of millions of ChatGPT users can shop from Shopify merchants, that users can complete purchases via an in-app browser, and that merchant customizations carry over: brand experience, pricing logic, payment methods, checkout customizations.
It says orders flow into the admin with ChatGPT referral attribution and that merchants remain the merchant of record, retaining ownership of customer relationships and data.

That is the line I want to sit with.
Merchants remain the merchant of record.

Because that is the whole episode in one sentence.

If the merchant remains the merchant of record, then the merchant also remains the owner of a whole bunch of downstream obligations.
Pricing accuracy.
Product policy accuracy.
Post-purchase support.
Returns.
Refunds.
Fraud handling.
Loyalty.
Customer communication.
And maybe most importantly, explaining what happened when the journey did not begin on the merchant site.

That is why I think Shopify is the clearest sign that this story has moved from interesting trend to actual operating surface.
The help documentation makes that even clearer.

Shopify's help page for the ChatGPT agentic storefront says the storefront acts as a discovery-focused referrer platform.
It says the purchase happens on the online store checkout through a ChatGPT in-app browser or a new tab on web.
It says all the merchant's customizations, branding, selling strategies, and payment methods are supported because the sale happens in the merchant's checkout.

That is hugely important.
Because it means the AI surface is upstream from the conversion, not a total replacement for the merchant environment.
And if that is true, then AI commerce should be treated as a channel that influences and originates journeys, not just a magic storefront floating in space.

The help page also says stores need relevant policies completed: terms of service, privacy policy, return and refund policy.
That might sound small, but it is not.
That is the platform quietly admitting what every operator eventually learns.
Policy hygiene becomes distribution hygiene.

If your policy layer is weak, your channel quality is weak.

And I know, I know, that is deeply unglamorous.
Nobody wants to gather the team and say, great news, the future of AI shopping requires much better return policy hygiene.
That is not exactly cinematic.
But it is true.

Shopify also says there is no built-in checkout setting for ChatGPT in the admin because it is a product discovery channel.
If a merchant wants to completely hide a product from being discoverable by ChatGPT, there is separate guidance for hiding products from discoverability by AI channels.

That matters too.
Because it introduces a new category of merchandising control.
Not just, is this live on the store?
But, is this discoverable in AI channels?

That is a different question.
And in many organizations, I suspect nobody officially owns it yet.

Which is usually how the fun starts.

There is another line in the Shopify release that I think operators should really pay attention to.
The company says products stay synchronized across surfaces, with real-time inventory and pricing, and that brands are syndicated and shoppable while still owning the purchase journey through their online store.

Again, that is the promise.
And it is a reasonable one.
But it also creates a testable operating burden.

If you are going to believe that promise, you should probably audit it.
You should probably check whether your top products really are synchronized the way you think they are.
You should probably verify that pricing, inventory, policy language, and image quality survive the trip.

Because if there is one thing ecommerce is exceptionally talented at, it is discovering that real-time means something slightly different in five different systems.

Now let us bring Google into the picture.

Google's contribution to this story is a little different.
It is not just saying here is a sales channel.
It is pushing two ideas at the same time.
Shopping answers can become more personalized, and commerce interoperability can become more protocol-driven.

Those are related, but not identical.

First, personalization.

In its March seventeenth post on Personal Intelligence, Google says the feature is expanding in the United States across AI Mode in Search, the Gemini app, and Gemini in Chrome for personal Google accounts.
Google also says this connected experience does not apply to Workspace business, enterprise, or education users.
It describes shopping recommendations tailored to a user's recent purchases, preferred brands, and style.
It says users can choose whether to connect apps like Gmail and Google Photos, and can turn those connections on or off.

Operationally, that means shopping answers can become more context-rich.
Maybe better.
Maybe more useful.
Maybe more likely to feel magically relevant.

It also means the result becomes harder to reproduce.

That is the key operator point.

If one user gets a shopping answer shaped by recent purchases, brand preferences, and connected account context, and another user gets something different, then your old QA habits start to break.
You can no longer assume that one screenshot tells you the universal truth.
You can no longer assume that a single test query fully represents how the system behaves.

And if you run growth, merchandising, customer experience, or product operations, that should make you just a little uncomfortable.
In a healthy way.

Because personalization changes what review means.
It changes what ranking issue means.
It changes what this product is showing up means.

You now have to ask:
For whom?
Under what context?
With what data connected?
In what geography?
In which product surface?

And this is the part where I become, once again, the least fun person in the room and ask whether anyone can reproduce the result twice.

Second, protocol.

Google's March nineteenth update on the Universal Commerce Protocol is, in some ways, even more important for operators.
The company says UCP now includes a cart capability so agents can save or add multiple items from a single store, a catalog capability so agents can retrieve select real-time product details like variants, inventory, and pricing, and identity linking so shoppers can receive benefits like loyalty pricing or free shipping when logged in on integrated platforms.

That is not a cosmetic update.
That is the commerce stack getting more structured.

It means the future version of AI shopping is not just the model says some shopping things.
It is there is an underlying system for how agents talk to merchant data, carts, identity, and benefits.

Which sounds very clean.
And maybe it will be, eventually.
But in the near term, it means more rails.
More capabilities.
More conditional behavior.
More dependencies.
More reasons for one customer journey to look perfectly normal and another one to be weird in a way that nobody catches until after it hits support.

Google also says adopters can customize which capabilities to support.
That is great for flexibility.
It is also how fragmentation happens.

Because once platforms and merchants support different subsets of capabilities, you no longer have one neat commerce surface.
You have a matrix.
You have different levels of support for cart, catalog, identity, loyalty, and maybe eventually post-purchase experiences.

So the Google story is not just shopping gets more advanced.
It is shopping gets more variable.
And variable systems require stronger operations.

Now let us talk about what I think is the hidden trap here.
Measurement theater.

This was already a problem in Episode 14.
Now it is a bigger one.

Because the easier these surfaces become to screenshot, the easier it is for teams to tell themselves a very flattering story.
Look, we showed up.
Look, the answer mentioned us.
Look, we are present in the future.

And, um, that may all be true.
It may also mean almost nothing.

Because the real questions are harder.

Did the customer click through?
Did the journey reach merchant checkout or stay in some controlled platform surface?
Did pricing match?
Did availability match?
Did the coupon work?
Did loyalty carry over?
Did the referral get attributed cleanly?
Did the product return rate change?
Did the support burden increase?
Did customer trust actually improve, or did we just create a more sophisticated way to mis-measure intent?

This is why I think finance, ecommerce, growth, and customer experience need to stop treating AI shopping as someone else's experimental toy.
It is a cross-functional channel question now.

And maybe that is the real joke here.
The future arrives, and instead of replacing org charts, it somehow creates even more reasons for them.

But seriously, this matters because the accounting logic and the customer logic can diverge really fast.

Here is one example.
Suppose your product shows up beautifully in an AI answer.
The customer clicks through.
They land in merchant-controlled checkout.
The sale happens.
Wonderful.

Now tell me:
Does analytics treat that as referral traffic, AI-originated traffic, or something else?
Does the attribution logic put it in the same bucket as search?
Does paid media start claiming credit because of later touchpoints?
Does the customer support team know that the journey began in ChatGPT or AI Mode?
Does merchandising know which products are consistently being surfaced and which are not?

If the answer is no, then you do not have an AI shopping strategy.
You have an AI shopping anecdote.

And anecdotes are lovely.
I have many.
They should not run your operating model.

There is a second measurement trap here too.
Personalization makes visibility less universal.

If Google or other AI surfaces shape recommendations based on user context, then we rank well becomes a much fuzzier statement.
Maybe you rank well for certain users.
Maybe not for others.
Maybe only when certain signals are connected.
Maybe only when the customer expresses a specific kind of intent.

So the analytics layer has to mature.
Not theoretically.
Practically.

You need an AI-originated traffic view.
You need to separate it from ordinary referral or search buckets.
You need to know whether the journey was discovery-only, assistive, or conversion-capable.
And you need a weekly review that includes somebody from outside growth, because growth teams alone are structurally tempted to fall in love with top-of-funnel novelty.

I say that with affection.
And experience.
And maybe mild emotional scarring.

There is one more thing here that I think is under-discussed.
Post-purchase ownership.

A lot of the excitement in AI shopping naturally centers on discovery and comparison.
That makes sense.
Those are the visible parts.
The interesting parts.
The parts that demo well.

But the discipline of commerce has always been that the expensive truth often appears after the click.

Returns.
Refunds.
Support.
Delivery issues.
Policy misunderstandings.
Product mismatch.
Expectation mismatch.
Loyalty frustration.

If a customer begins in an AI surface and the surface simplifies, summarizes, or rearranges product information on the way to purchase, who owns the interpretation gap?

That is not a rhetorical question.
That is a policy question.
A legal question.
A customer-experience question.
A trust question.

The Shopify help page's emphasis on store policies is not random.
It reflects the fact that these channels are not purely merchandising channels.
They are policy-sensitive channels.

And I think a lot of teams are still treating AI shopping like a prettier search layer.
It is not just that.
It is a distributed control layer.

That means the sales path, the policy path, and the support path can now start in one place and finish in another.
And any time that happens, ownership has to be named very explicitly.

Otherwise you get the classic enterprise move, which is everybody nodding thoughtfully while secretly assuming another team has this covered.

That move is undefeated.
It is also very expensive.

So if I were running this as an operator, not as a commentator, what would I do by Friday?

First, I would publish a one-page AI commerce surface map.
Nothing fancy.
Just one page.
For each relevant surface, I would name:
The discovery source.
The product feed owner.
The checkout owner.
The analytics owner.
The merchant-of-record assumption.
The returns and support owner.

Because one of the fastest ways to make a problem expensive is to leave ownership verbal.
Write it down.

Second, I would audit twenty-five priority SKUs.
Not your whole catalog.
That is how good intentions become next quarter's problem.
Pick twenty-five that matter.
Then check title consistency, image consistency, price consistency, availability consistency, shipping and policy language consistency, and whether legal disclosures show up clearly enough in the product description.

If you cannot trust those twenty-five, you definitely cannot trust the bigger story you are telling yourself.

Third, I would create an AI-originated traffic view in analytics.
Separate.
Visible.
Weekly.
Do not bury it in generic referral traffic.
Do not let it hide inside search.
Do not let three teams define it three different ways.

Fourth, I would make a category decision on merchant-controlled checkout.
For some categories, maybe merchant-controlled checkout should be a hard requirement.
For others, maybe the business can tolerate more experimentation.
But make that a named decision, not an accidental outcome.

Fifth, I would assign one owner for post-purchase policy across AI-originated journeys.
Not forever.
Not for every policy topic on earth.
Just for this surface.
Someone needs to own the question: if the customer came through an AI answer, what promises, disclosures, loyalty mechanics, and support expectations are we willing to stand behind?

And sixth, I would run the worksheet.
Yes, of course I made a worksheet.
I know.
I hear myself.
I would also like to be the kind of person who does not say run the worksheet with a straight face.
And yet ... here we are.

But genuinely, a worksheet helps because this is one of those topics where teams think they agree until they start writing down who owns what.
Then suddenly the air gets very quiet.
Which is useful information.

So let me bring this together.

Episode 14 asked a discovery question.
Episode 15 asked an ownership question.
Episode 16 zoomed out to a capacity question.

This week, Episode 17 is asking a control question.

If AI shopping starts the journey, who controls what happens next?

OpenAI is leaning toward a model where discovery happens in ChatGPT and merchants regain more checkout control.
Shopify is making AI shopping feel like an actual administrable channel, with referral attribution and merchant-of-record logic intact.
Google is making shopping more personalized and more protocol-driven, which could make the experience more useful for users and more complicated for operators.

And the common thread is this.
These are no longer isolated product updates.
They are pieces of a new commerce operating layer.

The teams that do well here will not just be the ones who show up in the answer.
They will be the ones who can explain, with a straight face and real evidence, how the product got there, how the customer moved forward, how the measurement works, what policies apply, and who owns the messy parts after the click.

That is not glamorous.
It is not especially tweetable.
And it is definitely not the kind of thing that gets turned into a triumphant keynote line without a lot of editing.

But it is how this actually becomes durable.

So by Friday, I would not ask, did we show up?
I would ask, can we control the journey well enough to trust what this channel is doing?

If the answer is yes, great.
You have the beginning of a strategy.

If the answer is no, that is also useful.
That means you are still in screenshot season.
And screenshot season, while fun, is not the same thing as operations.