#48 | The Great E-Commerce Split

Imagine your personal AI assistant shopping for you while you sleep. Not just finding things, but evaluating, comparing, negotiating, and making the purchase. Now, imagine the hidden infrastructure required to make that possible.

I was recently at an interesting meetup about AI, with a couple of people from the biggest e-commerce players in Poland. In a dimly lit Warsaw coworking space, three e-commerce veterans leaned forward simultaneously when someone muttered: "We're not building one customer journey anymore. We're building two."

Soon, we will not be designing one e-commerce experience. We will be designing two distinct realities.

One will be for you and me, the physical users who browse a website. The other will be for our agents. Your competitors' agents are being trained right now.

The future of e-commerce is a split reality. There will be the product page we all know, with its polished images, persuasive copy, and human user reviews. This is for human eyes, for emotions, for inspiration. Think of it like a restaurant with both a beautiful menu (for hungry humans) and a detailed nutritional database (for dietitian apps).

And then there will be the hidden, machine-optimized shadow pages.

This second version will be extremely lightweight, almost no visual elements. It will be written in something very much like Markdown, stripped of anything that could dilute an LLM’s understanding. Its purpose is singular: to be crawled and understood by an AI agent tasked with making a purchase. The goal is to inject as much clean, unambiguous context into the LLM as possible, so it can make an intelligent decision. Where humans see lifestyle photography, agents receive:

`{product: electric_shaver, attributes: [wet_dry: true, battery_minutes: 90, warranty_years: 2], warranty_years: 2, price: 99.99, reviews_rating: 4.5, brand: "Philips"}`

The recommenders and listicles will also be different. An agent, tasked by a user to "find the five best shavers for men," will crawl 10 pages, identify popular brands and models, and then access the major e-commerce players to execute the purchase. The time to decision is different for agents; they are not thinking for months. We've seen this movie before - when mobile forced us to create ‘m.domain.com’. This time, it's ‘ai.domain.com’.

This is not science fiction. This is the next battlefield for product discovery.

For years, we have obsessed over Search Engine Optimization (SEO). We tailored our content for Google's crawlers, hoping to please the algorithm and rank higher for human searchers. It was a good job. We got proficient at it.

Now, we must prepare for Agent Engine Optimization (AEO). Brands are already secretly building for this.

The new job will be to understand how these AI agents make decisions and to optimize the entire digital presence of a brand to be understandable for them. There are three critical points to this:

1. Presence: The brand must be presented well across the web, not just on its own site. Agents will roam.

2. Clarity: The information must be structured for machine comprehension, not human persuasion. It must be explicit, not implicit.

3. Accessibility: The data must be available for crawlers. This is a challenge, as some search engines do not want you to go beyond their results page. We may need a different kind of search engine, one built not for people, but for LLMs.

This is an entirely new discipline. A sub-job for marketers and product people to optimize the experience for LLM engines. Early data, still informal, suggests that LLM crawler traffic for some Shopify merchants already exceeds 5% of visits. This is not a future concern; it is a present reality.

The most logical solution will be a new wave of external APIs created solely for LLM integration. Think of connections for n8n, Zapier, or Make.com, allowing agents to plug directly into a service.

And these APIs, they will be different. Imagine really stupid, really simple APIs. They will be highly standardized, without any security by obscurity. They must adhere to the patterns the LLM has already seen in its training set, complete with perfect documentation. They must be predictable.

For some products, like clothing, these agents must also be multi-modal. They will scan pictures, understand them, and make decisions based on a user's pre-compiled profile-heir style, their photos, their preferences. This is a totally different experience, built on simple images that do not dilute the LLM's understanding and clean data feeds.

We are at the level of discussion, who is going to pay for it? There is a proposed model, something like "pay-per-crawl," but nobody knows how it would work.

But enabling this future requires dismantling the very defenses we've spent decades building. Here we find a conflict. Many platforms, like Shopify, have already blocked access for most LLM crawlers. They do not want their data scraped. This creates a strange dynamic. If the official door is closed, the agents will have to come from other places, maybe through some really expensive proxies.

This leads to a philosophical question for product managers: what constitutes the user?

For years, we built systems to fight bots. We would physically alter the UI if we identified a bot - blurring images, changing the location of elements, or blocking them at the network level. These were bad bots, scraping prices and content.

Now, we are talking about opening this very same route for "good bots" - the LLMs - that look technically identical. We have to make this really fluid, really vague distinction between the bots we want and the bots we don't. The job is getting more complicated, just as we get more skilled.

We have some seriously exciting times in front of us. A considerable number of jobs will change. The entire experience has to be re-evaluated. And I am not sure if I want to see the end of it, because the process of getting there is going to be ugly. But it is coming.

So, what should you do tomorrow?

1. Audit your product pages for machine readability. Imagine a blind AI trying to understand your product from scratch. What information is missing or unclear?

2. Prototype a Markdown version of your top SKUs. Just a plain text file with attributes, values, and a clear product description. See how much context you can pack in without any visual fluff.

3. Meet with your dev team about LLM-specific API endpoints. Start the discussion: how can we expose clean, structured data for AI consumption?

The split is coming. The most valuable real estate in your next site redesign won't be above the fold - it'll be in the API response headers.

Are you allocating resources accordingly?

Which side of the page will you be ready for?