Every 'GPT' wants to be the only app you will ever need
In September 2025, OpenAI and Stripe launched Instant Checkout in ChatGPT. Any merchant can plug their catalogue and checkout straight into the assistant, and free users were included from day one. Etsy sellers went live first, with over a million Shopify merchants lined up to follow. Walmart signed on in October, and Target announced its own app inside ChatGPT.
The shopper browses, compares and pays inside the conversation. They never touch the merchant's website, and OpenAI charges the merchant a fee on every completed sale. The form keeps shifting (by March 2026, OpenAI was reportedly nudging purchases towards apps that brands build inside ChatGPT), but the direction does not. Buying is moving into the conversation.
This pull towards a single interface is not new, and it is not only Silicon Valley nor did this pull start with AI. For instance, WeChat turned a messaging app into the operating system of daily life in China: payments, bookings, government services, all inside one app.
Closer to home, KBC has been running the same playbook for years. Through KBC Mobile you can buy train and bus tickets, pay for parking and compare energy contracts. For five seasons you could even watch Jupiler Pro League goal highlights in your banking app, through a service called Goal Alert, until the broadcasting rights moved elsewhere in 2025.
Think about that for a second: a bank bought football rights to keep you inside its app. The app opened up to non-customers in 2019, and Kate Coins reward you for using it. Kate, the AI assistant at the centre of it, now handles 70% of customer questions fully autonomously.
Different players, same ambition: become the one place where everything happens, so nobody needs to go anywhere else.
What you lose when the agent wins
For brands, retailers and distributors, the maths is uncomfortable. Every interaction that moves into an agent is an interaction that no longer happens on your owned channels. With it goes the brand connection, the first-party data, and your ability to steer the conversation. You will compete to be mentioned in agent answers, and if you do that well you will be featured now and then. But you decide neither when nor where, and you learn very little about who saw you.
Your customers aren't really yours.
Tony SealeFounder | The Knowledge Graph Guys
Seale said this about handing your customer data to large platforms, and the agent era makes it literal. Push the trend to its end and the brand is reduced to a fulfilment partner behind someone else's interface. That is a precarious place to be, because pure fulfilment is a scale game that players like Amazon already run more efficiently than you do.
Can you simply refuse? Amazon did. In November 2025 it blocked OpenAI's crawlers from reading its product pages, prices and reviews, making the world's largest online retailer largely invisible to ChatGPT's shopping answers. But that stance only works when you have Amazon-scale gravity, where most shoppers start their search with you anyway. In the Benelux, perhaps bol.com could afford it. Almost nobody else can. For the rest of us, opting out of the agent ecosystem means opting out of being found.
Not everyone is equally exposed
The exposure is not evenly spread, and it is worth being honest about where you sit.
Manufacturers hold the strongest position. Someone still has to mill the flour, make the paint, build the furniture. Physical production is one thing the agent cannot absorb, even when it controls discovery.
Brands keep their pull as long as people ask for them by name. An agent can reroute a generic question, but "I want brand X" survives the interface change. The risk is gradual erosion: the fewer direct touchpoints, the weaker that preference gets.
Retailers and distributors whose value is mainly assortment and availability are the most exposed. Those are exactly the jobs an agent does well.
The pattern underneath: the further you sit from making something or knowing something unique, the more replaceable you become.
A better AI model will not save you
A reflex I see in many boardrooms is to look for differentiation in AI itself. Pick the right model, build the smarter assistant. I think that is a dead end. Foundation models are built by a handful of companies with the deepest pockets, and the rest of us rent them. The same intelligence is available to you, your competitor and the agent platforms, as long as your credit card is solvent. There is no edge in something everyone can buy.
Your public content will not differentiate you either, because it gets scraped into the same models, together with everything your customers say about you online.
What does not get scraped is what never gets written down publicly. Your domain knowhow. The intricate understanding of how your business and your customers' problems actually work.
The answer: a map of what only you know
A knowledge graph, stripped of the technical detail, is a map of your world: the things that matter in your domain and how they relate to each other. Not pages, not documents. Meaning.
Take a paint manufacturer. Decades of knowledge about which primer works on which substrate, how humidity affects drying, which colour systems hold up outdoors, which combinations fail and why. None of that lives on a product page. An LLM scraping the public web can describe the products, but it cannot reconstruct the reasoning behind them. That image of your domain has to be built deliberately, with care, by people who know the business. It does not emerge from crawling.
Once that knowledge exists as a graph, it works for you in two ways.
First, it powers experiences nobody can copy. Picture a customer asking which paint to use on a damp, north-facing plaster wall. The generic agent answer lists products it scraped from the public web. Your advisor names the right primer, the drying window between coats and the finish that will hold, and explains why, because the graph knows how moisture, substrate and product relate. Same model underneath, completely different answer. The model provides the language. The graph provides the judgement. A competitor with the same model and without the graph cannot replicate it.
Second, it gives you a control layer. You decide which slice of the graph you publish openly as structured data, so agents cite you accurately and often. You decide which slice you expose through APIs to selected partners or agent platforms, on your terms. And you decide what stays internal, feeding only your own tools. Scraping takes what is on the surface. The graph lets you choose what goes beneath it.
Haven't we heard this before?
If you have been around long enough, this may sound familiar. The semantic web made a similar promise twenty years ago: describe your world in a way machines can understand. For most companies it went nowhere. Two things changed. LLMs can now help build these structures, work that used to take specialist teams years. And it turns out AI needs them: researchers at data.world measured that an LLM answered complex business questions three times more accurately when backed by a knowledge graph, and scored exactly zero on the hardest category of questions without one.
This is not lab theory either. Siemens has run an Industrial Knowledge Graph since 2018, feeding applications from gas turbine maintenance to factory monitoring, and today positions that graph as the foundation that makes AI agents usable across a company. The companies that mapped their domain early were not wasting money on an academic exercise. They were early.
Slow to build, and that is the point
Building this takes time. It needs people who actually know the business, and it does not fall out of a tool you buy. I understand that reads like a drawback. It is the opposite. Anything a competitor can replicate with a bigger budget is not an advantage. The slowness is the proof that it can be defended.
You do not need to map everything at once. Even Siemens advises starting small, with a couple of areas where the value is obvious, and growing from there. In practice that means picking the domain where your knowhow is densest, often the questions your sales engineers and support people answer week after week, and capturing that first.
And a question worth putting to your own team: could your current platform carry a layer like this, or is your knowledge locked inside pages? If you already work headless and API-first, with a well-modelled content platform and structured data, you are closer than you think. I made that case in detail in my earlier piece on content models and ontologies, and everything that has happened since has only strengthened it.
One honest question to close
Here is the uncomfortable closing thought. If you strip away the AI question entirely and ask what unique knowledge your organisation holds, and the honest answer is "not much", then the agent era is not your problem. You were already fighting it out on price and availability, in a market full of lookalikes. AI will simply expose that faster and more publicly than the market did before.
But if that knowledge does exist, and in most companies I meet it does, locked in heads, spreadsheets and legacy systems, then the work is to get it out and give it structure. The agents are coming for the interface either way. What they cannot take is a domain map you built yourself and choose how to share.
Sources
Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol, OpenAI
Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol
OpenAI shifts checkout plans in its agentic commerce strategy, Digital Commerce 360
Amazon quietly blocks more of OpenAI's ChatGPT web crawlers, Digiday
KBC Mobile named best mobile banking app in the world, KBC Newsroom
Tony Seale: The Knowledge Graph Guy, Knowledge Graph Insights podcast
Use Cases of the Industrial Knowledge Graph at Siemens, ISWC 2018
Let's talk about it
I keep coming back to one question with the teams I work with: if an agent asked your systems what your company actually knows, what would the answer be? If you are curious what your domain looks like as a graph, or whether your current content model could carry one, I would genuinely enjoy thinking that through with you. Which knowledge in your organisation would be hardest for a competitor to rebuild?
Book a call with us