E146 - Commerce Today - Your Next Customer is an AI: Preparing for the Agentic Commerce Wave
===

Joshua: [00:00:00] Welcome to Commerce Today. I am Joshua Warren, and this is episode 146. Your next customer is an AI preparing for the ag agentic commerce wave. So I want you to picture a shopper a year or two from now. She doesn't open a browser, she doesn't scroll on Instagram. She says possibly out loud, possibly typing on her phone.

I need hiking boots, waterproof sustainably made. Under $200 delivered by Friday order the best option for me. Her AI does all the rest. It checks product specs, it scans reviews. It confirms inventory. It calculates delivery promises. It places the order. She never sees your website. She never sees your ad.

She never sees your social media posts. Her agent was the shopper. That's agen commerce and it feels a little buzzwordy right now, especially with certain e-commerce platforms out there rebranding themselves around it and everything. But it's honestly not a buzzword. It's not a 10 year maybe moonshot.

It's actually the logical next step from where we already are. Voice assistance [00:01:00] for some have gotten people comfortable asking, find this for me. Recommendation engines taught us to trust machines to narrow down our choices. And just one more step forward and we're buying stuff through agents that act on our behalf.

So here's the really hard question, and this actually goes back to one of the first Commerce Today episodes where we talked about how to make your products visible to ai, but how is your brand gonna show up in that agentic commerce world? And we're not about chasing the latest AI trick. I'm sure you've all seen the ads like I have of all these companies out there already promising that they can make sure your product shows up.

First on chat, GPT. Just like all the people that like to promise, we'll make sure your product shows up first on Google or all the way back through all the different platforms and marketplaces over the years. So we're not talking about chasing those tricks, we're talking about setting a foundation. This can be kind of the boring foundational parts of e-commerce that actually end up make or break your brand whenever [00:02:00] machines become the buyers.

So in my book, the eCommerce Growth Playbook, I talk about three base layers, having clean data realtime system. Integration and earn trust. Those aren't nice to haves in this ag agentic commerce world. They're the price of admission when an AI agent becomes our customers. So I'm gonna break this down into three pillars.

You must get right structured data, realtime availability, and price, and digital reputation and trust. I'll give you some concrete examples of readiness scorecard, and if you reach out to me on LinkedIn, again, my name is Joshua Warren. You'll see a Creatuity gold background behind my headshot. When you search for Joshua Warren.

Reach out to me mentioned that this is episode 146 and you would like the 90 day plan for genic commerce, and I will send that straight to you. So let's dive into it. Very, very first pillar we're gonna talk about is that structured data is the new SEO. Humans really like pictures and clever copy agents, like fa agents, like [00:03:00] facts agents don't browse.

They parse, they ingest product truth through feeds and through APIs. If your data is messy, if it's missing, if it only lives in a paragraph on your PDP, then a lot of agents can't see you. You're absolutely invisible. I'm sure you've tried this, but if you haven't, go to chat. GPT, even a, a free account or no account, just using it.

Not logged in and try first asking it for recommendations of brands that sell your product online. And don't give it your name, don't give it your specific product. But if you sell hiking boots, say, Hey, show me the top 10 places to buy hiking boots online. See if you're listed. Then in a new session, I ask it, Hey, what do you think about and insert your brand name or insert your products and see if it knows anything about them.

See if what it knows is accurate. You may be surprised, especially if you haven't invested in structured data. And structured data means that every important attribute lives in a discreet field with a clear definition. So the material isn't just [00:04:00] premium eco leather that just happens to be mentioned in the description randomly.

It's a field that actually says material equals quote leather origin equals quote Italy. Tanning process equals vegetable certification equals leather working group gold. Waterproof isn't just like a vibe, it's actually a rating. It's IIPX seven. Weight isn't lightweight. It's 1.9 pounds per pair in size nine.

Care isn't quote. Wipe with a damp cloth, it becomes care method equals wipe colon damp cloth machine. Washable equals false. You get the idea. This is usually done through two different things. One structured markup within the htmo on your website. So your e-commerce platform has to support it or your theme has to have been modified, but also the product data itself.

And this is where a real solid PIM product information manager. Earns its keep a PIM gives you a single source of [00:05:00] truth. It enforces attribute completeness and it makes it so easy to say, okay, we actually want to get not just the product description, we now need to list each individual attribute in the HTML code in the specific way.

And a PIM will make that so much easier. It lets you publish that consistent machine readable data to every endpoint, your storefront, your marketplaces, your syndication partners, and yes, the future agent networks that are covered. So a quick test you can run this week. Take your top 50 SKUs for each list.

The top 10 attributes A smart agent would need to answer a buyer's query for hiking boots. That can be waterproof rating height. Outsole component weight, midsole material drop, sustainability cert side range. You get it? Can you export those as clean fields today or are they just buried in the copy?

If it's in the copy, you're at risk. And I will say agents are getting better and better at understanding copy, but they are always going to default to [00:06:00] recommending products that they are most confident about, and these structured data fields give them more confidence. Kind of weird to think about AI confidence, but that is the world we live in.

Another test synonyms and taxonomy agents will resolve intent across different words. It's super interesting if you dive into vector stores and vector databases if you're geeky developer like me more than we have time to get into, but basically. It's not even just simple synonyms agents when they're thinking through vectors, they actually can make connections like dog and cat are related, but different dog and puppy are more similar and kind of y.

And horse is also an animal is also within that family, but they don't just see it as is also an animal. There's exact precise numbers. For the distance between the concept of a dog, the concept of a puppy, the [00:07:00] concept of a cat, the concept of a horse, probably not the best examples unless you're selling pet products, but that applies to literally every word in our English language and in every language because these LLMs often speak all the languages now, which is a little crazy as well.

So basically you no longer have to say. You know, shoe boot, like all the different things about a hiking boot, they will actually understand the concept that a hiking boot is a shoe. So if the user says, I'm looking for hiking boots, they're gonna realize this user is looking for a shoe, specifically a hiking boot.

Now, that relates as well to other things that users might be searching for or the ways that they might specify. What they're most interested in, what's most important to them in this product? So if you're looking for something that is high quality workmanship, that could be phrased a million different ways.

But if you can get that concept into your product attributes, into your product taxonomy the agent will [00:08:00] understand if you call it a certain certification from an industry leader that says that this is high quality workmanship. And the user says, I want a high quality product. The agents can actually connect those two things.

So it's important to have all of that data, those synonyms, that taxonomy all set up so that the agent has as much as possible about your product at its disposable 'cause it's gonna make connections that you would never make, you would never think of, and that your customers would never think of. So a few common failure patterns I see with all of this.

If your attributes are free text and there's no validation so I see this a lot before companies deploy a PIM system. So you might have blue blue all over, case blue, all uppercase, navy blue, midnight and especially if it's midnight, it doesn't say midnight blue. Now all of a sudden it doesn't know if that's the same color or even if that is a color.

So make sure that your attributes, you have some sort of validation rules. And ideally, they're not just free [00:09:00] text where whoever is entering that particular product picks what the color is called. Also, variance without clear linking fields, the machines can't always tell which sizes exist or which colorways might share apparent product unless you set that up well in your product data be careful if you have units all over the place.

So if you use inches in one place, centimeters in another, or maybe you use pounds in one place, ounces in another normalize, especially within product families. Okay, media without structure. So videos and PDFs can be great for humans, but if you just throw four random videos onto the product page that contain everything that you actually need to know before you buy the product, that's not as helpful as actually having URLs, dimensions, alt text, and licensing data within Fields on the page.

Make sure you have compliance information, hazmat flags, battery types, country of origin agents need them to avoid bad recommendations. So five quick wins you could have [00:10:00] in the first 30 days. Define a minimal viable attribute set for your top categories of 10 to 20 fields each that you feel the agents will need.

Make them required in your pim. Normalize your units. Pick one standard per attribute and convert everything. Add unique IDs everywhere. Use G 10 A UPC, and EAN and MPN, at least an internal SKU number. And then link those between parent and child products. Make sure that your key claims are also represented in attribute fields, so if your products waterproof, add a field for the waterproof rating.

Add a field for recycled content percent, if that is, if you're making sustainability claims, add a field for warranty length for care methods and make sure you publish a product JSON feed and keep it in sync with your pim. If you don't have a PIM yet you can start with a clean version controlled export from your e-commerce system.

But I highly recommend in this agentic world looking at a pim. [00:11:00] Now, remember, pretty doesn't matter to an agent, at least not yet. Precision does information, does confidence in being able to reach the user's criteria, what their recommendation does. So the better your data, the more often you're gonna make that short list of the top recommended products.

Now second pillar real time availability is a must. Agents won't tolerate maybe if they're not sure if a product is in stock and can be received in time for when the user has requested it, they're not gonna buy it if an agent is committing to buying on behalf of a person. Especially, you know, and there's a difference here between the little tests you might run in the chat GBT or I've been doing a lot with a perplexity comment browser and you can tell it, find me a product that does this, this, and this.

But then when you actually ask it to buy you a product, suddenly it has to have much more certainty in the information. And so they wanna know, is the item in stock? What price is it at? Can it ship today? What is the delivery promise to this specific zip [00:12:00] code? If your systems can't answer those questions in milliseconds, an agent is gonna skip you and buy from somewhere else that can answer that.

So real time isn't fast, it's real time. It means live. It means no 15 minute delays, no hour delays, no nightly batches. It means an inventory service. And it's reserving stock that's preventing overselling and it's releasing holds. When orders expire, it means a pricing engine that returns the right price for that exact customer, including promos, loyalty tiers, or contract terms.

This sounds like a lot, but if you break it down, it's very doable. So you start with an availability. API start with a simple endpoint that says availability of this skew to this postal code and this quantity be able to return. How many are available, or at least yes or no. Is it available, is it not available?

I recommend agents really like timestamps and that can really help them as they are processing this information. So [00:13:00] also include timestamps. Obviously use this, give as much information via endpoints like this as you're comfortable with. This could become something that a competitor tries to exploit or tries to use for competitive intelligence.

So. Put some protections around this. We want the agents fetching this data. We don't necessarily want competitors scraping this data. Also look at a pricing, API or once you're done with all of this, just a product, API that has all of this data. But again, keep it simple. You can just say, give me the price for this customer group and this sku and provide that price back to the agent, a delivery promise, API.

So if you can say. Basically an agent, if the person said, find me the best deal that I can receive by Friday, they're gonna look at what the total landed cost is for things that will arrive by Friday. So you need an end point that will let the agent see just how fast can this product arrive. I mentioned guardrails, have rate limits, perhaps [00:14:00] some authentication.

This is an area that is emerging very quickly. I would not be surprised if chat GPT and others have a standard, you know, Google produced that sitemap standard and the Google product feed standards that everyone then standardized around many, many years ago. I think we're gonna see something like that come out.

Probably a couple of emerging standards that will compete but look to those to see how can you organize this data, expose it to the agents without exposing it to your competitors. Now, if you're using a, a composable setup, this will be pretty straightforward. Also, if you have a modern order management system or a platform that has an inventory microservice, this should be super simple.

If you're on a monolith. You might need to do a little bit more development work here. So definitely start looking into that. Now also next up, pillar three, your digital reputation and how your digital reputation is your currency. So when all the attributes tie, when the delivery dates tie, let's say you have the [00:15:00] same product at the same price you can deliver to the same person.

How is the agent gonna pick And trust signals are gonna break that tie. Reviews, ratings, defect rates, return rates, warranty claims, customer effort scores, support responses. Agents are gonna weigh all of that. And the interesting thing here is they're not just gonna look at your website. I find it super interesting and sometimes a little scary just how often chat GPT will find a random comment on Reddit about a brand and then suddenly use that.

That will. Heavily weight chat, GPTs feelings about that brand and what it says about that brand. So your digital reputation matters more now than ever before. And it's not something that you can say, you know, oh, we have a good brand. No, you have to earn a good brand one order at a time, and there has to be publicly verifiable proof that it's a good brand.

That's how agents look at things. They're a little less binary. One zero. Than computer programs. They're a little more [00:16:00] human-like in kind of feeling for things. But still, at the end of the day, they are a computer and they're either gonna trust you or not. And it's gonna be based on is there data online they can verify behind that trust?

So start with your reviews. The volume of reviews, the recency of the reviews, but also the specificity of reviews. Before, if you had thousands of reviews, the average customer wasn't gonna read every review before they ordered from you. The average LLM will, and they can even summarize those reviews and find things that really stand out and share that with the users.

So a thousand reviews from last year aren't worth as much as a hundred from last month. That mentioned the exact attributes the buyer asked about. Also, photos and video are gonna increase competence. The agents aren't really parsing videos much yet, but they are absolutely parsing the photos that they see in reviews.

Verified purchase badges can help. Responding to negative reviews helps even more, especially when the fix shows up in the next run of the product. Shows up in future [00:17:00] reviews also syndication matters, and being able to find those reviews on other sites. And also being able to find reviews on marketplaces about your product.

And then. Customer service is also a big part of this. If you can either publish your first response time and your resolution time, or if you have reviews that mention that, agents will notice that. So consider all of these things around the customer experience and your digital reputation. Also, think about the agent experience.

We talk about user experience. Now we're also gonna be looking at. No user experience, the experience that agents have when there's no actual user browsing the site. And this is super interesting and this is changing fast because the recommendation even a few months ago might have been around structured data and APIs, like I mentioned, and that is very important.

But if you haven't used some of the newer tools from chat GBT where it runs in agent mode and it's actually [00:18:00] using its own computer. Go try that. Either try it yourself or find a YouTube video of somebody trying it because it's a little crazy chat. GPT has a web browser chat. GPT browses the site, the user can watch.

In most cases they don't because it's actually pretty slow at using the web browser. At least now I'm sure it'll get faster. So you're having to think about how do we provide all this data in a structured way for agents that are gonna access it that way, but also how do we provide a user experience that is easy for.

An AI using a web browser to access our site. And this is where the technical elements of the site can be so important and technical elements of the design, because I have seen where chat GPT misses it will try to click a button and it's actually off by a few pixels. It's it's interesting, a lot of the accessibility stuff actually applies to chat GPT, which is so strange 'cause that's not what it was designed for or was anyone's mind when the accessibility standards were written.

But chat [00:19:00] GPT will actually have a hard time if the elements are too close. If it requires too precise of a click it's agent mode is gonna have a challenge navigating your site. So really the user experience is. Changing constantly and is evolving constantly of how we need to design sites so that these agents can access them and use them.

And again, I mentioned earlier, experimenting with ChatGPT highly recommend you do some usability tests using ChatGPT's agent mode. Ask it to go to your website, ask it to take a few key actions, see can it add products to cart, can it find product information? What you find out, what it returns might surprise you.

So a few quick things to look at on the agent readiness scorecard. Again, I can actually send you a PDF of this so you don't have to make frantic notes. But there are 10 items on that scorecard attribute, completeness and you score yourself zero to three. I'm not gonna read you every single option, but it's basically about, are the attributes just [00:20:00] embedded in the copy or do you have full validation, full governance.

And a JSON snippet exposed that has all the attributes your ID, hygiene, do you expose a UPC? Is it consistent? Your variant models, how are your variants set up? Are they set up in a way that agents are gonna understand them? Inventory availability, freshness. Is it updated nightly or is it actually updated in real time Pricing, accuracy, delivery promises.

Your reviews, your returns data, your policy transparency. And what's the agent developer experience like around your brand and those scores, those areas, evaluating those 10, scoring yourself on those can really help you decide what do we work on next? What do we start building now so that we're ready for this new world?

Tic commerce. Also have a 90 day plan. I'm not gonna read all of that to y'all. I can send you the PDF, just reach out to me on LinkedIn, send me a dm, say, Hey, just listen to or just watched episode 1 [00:21:00] 46 Commerce today. And I would love the 90 day plan or the agent readiness scorecard from that episode.

So that's it. I hope you found this helpful. I am super interested. I may even do a live stream in the future where we do a little bit of agentic shopping using. Chat GPTs agent mode because this is the brave new world of genic commerce and I am super excited to see how it evolves. If you have any thoughts, any questions, definitely feel free to DM me on LinkedIn or comment on this video, and I'll see you next time.

Thanks. Okay.