From llms.txt to a Complete AI-Ready Architecture

llms.txt was a good start. But it is not enough.

If you work in digital marketing and have been paying attention to the conversations over the past few months, you have probably heard of llms.txt. In short, it is a file you place at the root of your website to tell AI models what your site contains, similar to how robots.txt works for crawlers. We previously wrote about how your content reaches AI answers, and llms.txt follows the same logic: it helps AI understand you.

The problem is that llms.txt works like a table of contents. It tells models what pages you have, but it cannot express relationships between entities, confirm data provenance, or provide semantic context. It is like sending someone the chapter list of a book without a single paragraph from it.

A recent article on Search Engine Journal raises exactly this point and proposes a four-layer framework that goes much further. At difrnt. we analyzed the proposal and started testing it with a few clients. Here is what we found relevant and what actually works in practice.

The four layers of an AI-ready architecture

Think of your online presence not as a website with pages, but as a knowledge base that AI can query. The framework has four components that work together:

1. JSON-LD fact sheets. This is the foundation layer. Every important page on your site should have structured data in JSON-LD format that clearly describes what is on that page: product, service, person, organization. This is not new, but the data shows that pages with structured data are 2.3 times more likely to appear in Google AI Overviews. 2.3 times. That is not a marginal percentage, it is a difference that matters in real business terms.

2. Entity relationship mapping. This is where it gets interesting. AI does not understand your site as a collection of independent pages. It builds a knowledge graph. If you tell it that Dan Toma is CEO at difrnt., that difrnt. is a digital marketing agency, that the agency works with ecommerce and SaaS clients, AI can answer contextual questions about your brand. Without these explicit relationships, AI guesses. And it guesses poorly.

3. Content APIs. Instead of letting AI scrape your HTML, you offer it a structured endpoint where it can pull clean information. It is the difference between reading a scanned PDF and accessing a well-organized spreadsheet. The Model Context Protocol (MCP), created by Anthropic and already adopted by OpenAI, Google DeepMind, and the Linux Foundation, has reached 97 million monthly SDK downloads. That tells you this direction is serious, not experimental.

4. Provenance metadata. Who wrote the content? When? Based on what sources? AI is starting to weigh not just the content itself, but the credibility of the source. Timestamps, authorship, original citations build a trust profile that models use when deciding who to cite.

Why this matters now, not in two years

Researchers at Princeton published a study on GEO (Generative Engine Optimization) showing that structural signals can increase AI visibility by up to 40%. We are not talking about marketing promises here, but data from an academic study.

At difrnt. we see this confirmed in practice. Clients for whom we have implemented comprehensive JSON-LD and clear semantic structure appear significantly more often in AI-generated answers than competitors who have done nothing. It is not magic, it is logic: AI prefers data it can process cleanly, with verifiable sources and explicit relationships.

There is also a historical precedent worth mentioning. In 2012, Schema.org launched as a structured data standard. Companies that adopted it early benefited from rich snippets, knowledge panels, and a visually superior Google presence for an entire decade. The same kind of window is opening now for AI visibility. Those who move first will have an advantage that is hard to recover from.

The concept of Verified Source Packs

One of the most interesting ideas in this new architecture is the concept of Verified Source Packs. In short, brands that structure their data so well that they become preferred information sources for AI models.

Think about this from the perspective of a Romanian brand. If you are an electronics retailer with impeccable Product schema, up-to-date pricing, structured reviews, and clear relationships between products and categories, AI will prefer you as an information source over a competitor that has the same products but presents them in unstructured HTML.

This is not just about technical SEO. It is about brand optimization in the AI era, where data infrastructure becomes a direct competitive advantage. We saw this with a financial services client: after implementing Organization schema with explicit relationships between services, team members, and case studies, ChatGPT started mentioning their brand in answers where previously only larger competitors appeared.

What you can do right now, starting this week

The full four-layer framework is a medium-term project. But there is a minimum viable version you can implement quickly:

• Audit your existing JSON-LD. Run Google Rich Results Test on your top 20 pages by traffic. Check whether you have Product, Organization, FAQPage, and Article schema where appropriate. If you have nothing, start with Organization on the homepage and Article on the blog.
• Create a structured endpoint. It does not need to be complex. A JSON feed with your main products or services, including descriptions, pricing, categories, and relationships between them. This is the first step toward an API-first architecture that AI can consume.
• Add timestamps and authors. Every content page should clearly display the publication date, the last update date, and the author. This seems trivial, but it is exactly the kind of provenance metadata that AI uses to evaluate credibility.

At difrnt. we did exactly this with our own sites and client sites. We did not wait to have a complete solution across all four layers. We started with what we could do in two weeks and iterated from there.

MCP and the future of AI agents

Model Context Protocol deserves a separate discussion because its impact will be massive. MCP is an open standard that allows AI agents to connect to external data sources in a structured and secure way. The fact that Anthropic created it and OpenAI, Google DeepMind, and the Linux Foundation adopted it tells you it will become an industry standard, not just an experiment.

For brands, this means that in the near future, AI agents will no longer scrape websites. They will query APIs and structured endpoints directly. If your brand does not have this kind of infrastructure, it simply will not be visible to AI agents. We explored this direction in more detail in our article about SEO in the era of AI agents.

We are not saying you need to build an MCP-compliant API this week. But you need to start thinking in terms of structured data, accessible endpoints, and explicit relationships between your business entities. That is the direction, and it is closer than you think.

It is not about technology. It is about visibility.

Some parts of this article may have sounded technical. JSON-LD, MCP, content APIs, all of these seem like developer problems. But the real stakes are business stakes: if your brand is not structured for AI, you are losing visibility in a channel that is growing exponentially. And unlike traditional SEO, where you can catch up relatively quickly, in AI visibility the first-mover advantage is considerable.

Take 30 minutes this week and run Rich Results Test on your site. Check what AI sees when it analyzes your pages. If the answer is "almost nothing structured," you have a clear starting point. And if you want to discuss what a complete AI visibility strategy looks like for your business, you know where to find us.