Designing for the AI-Native Web: Schema, llms.txt, and Being Citable
How to make a site citable by AI assistants in 2026: what structured data and llms.txt actually do for the AI-native web, and where the advice goes wrong.
Most "AI SEO" advice in 2026 is selling a file. Add llms.txt, the pitch goes, and the assistants will start quoting you. We build for the AI-native web across finance and property, and we run this site as our own test bed, so let me save you the detour: the file is the least important part. The work that actually gets you cited is older, duller, and already sitting in your codebase.
Here is what we've learned shipping it.
Two audiences now read every page
A page used to have one reader that mattered: a search crawler building an index a human would later query. That reader is still there. But a second one has arrived — an assistant fetching your page mid-conversation, extracting a claim, and either citing you or paraphrasing you into oblivion.
These two readers want different things. The crawler wants a well-structured, well-linked site it can rank. The assistant wants a single page it can lift a true, self-contained sentence from without reading the other forty. Designing for the AI-native web means writing for both at once, and understanding that they do not use the same signals.
That distinction is where most of the confusion about llms.txt comes from.
llms.txt: useful, but not where you think
llms.txt is a proposed Markdown file at the root of your domain — a curated map of your most important content, written for machines. The theory is tidy: hand the model a clean index and it stops guessing.
The problem is that in 2026 the file lives a split-brain existence, and whether it helps depends entirely on which machine you're trying to reach.
Where it does nothing
Google has been blunt. Its official AI optimisation guidance, updated in May 2026, states that llms.txt is "not needed for AI Overviews, AI Mode, or any other generative AI Search feature." John Mueller compared it to the old keywords meta tag and noted that "no AI services used it and bots did not request the file." Gary Illyes confirmed Google isn't pursuing it.
The reason is structural, not political. Google's AI surfaces draw from the same index as classic search. If you're already indexed, the file adds nothing; if you're not, the file won't save you. For the single largest source of AI answers on the planet, llms.txt is inert.
So if someone tells you a text file will get you into AI Overviews, they're wrong, and the vendor said so themselves.
Where it genuinely helps
The other half of the market went the opposite way. Anthropic explicitly recommends llms.txt in its guidance for agent-facing documentation. OpenAI maintains the file for its own SDKs and protocols. Perplexity has been observed surfacing its contents directly. Chrome's Lighthouse 13.3, shipped in May 2026, added an "agentic browsing" audit that checks whether the file exists at all.
The pattern is clear once you stop looking for a universal answer. llms.txt is an agent and documentation convention, not a search one. If you run a docs site, an API, or anything an autonomous agent needs to navigate under a time budget, it earns its place. If you run a marketing site hoping to appear in a chatbot's answer, it's close to decorative.
Adoption data backs this up. Across top-100 domains in regulated sectors — financial services, healthcare, legal — publication rates still sit under 10 per cent. The measured uplift in citations, where it exists at all, is a "moderately positive correlation" on Anthropic and Perplexity, unquantified and non-causal. That is not nothing. It is also not a strategy.
Structured data is still the real lever
While everyone argues about a Markdown file, the thing that actually makes a page machine-legible has been sitting in the spec for a decade: schema.org structured data, expressed as JSON-LD.
An assistant extracting a claim from your page has to answer three questions fast — what is this, who said it, and when. Prose makes it infer all three. A BlogPosting block hands them over as fields: headline, author, date published, publisher. You've turned a paragraph it has to interpret into a record it can trust.
This is the part of "AI SEO" that is boring, verifiable, and works. Every article on this site ships a JSON-LD block. We validate it in the build — a page doesn't deploy if its structured data doesn't parse. That isn't a growth hack. It's the difference between an assistant attributing a sentence to you and attributing it to no one.
If you do one thing after reading this, make it this: mark up your pages with valid structured data and validate it in CI. It costs a schema block and a test. It outranks every text-file trick in actual effect.
Entity clarity beats clever markup
Here's the lesson that cost us the most, and it isn't in any of the guides.
When we launched this site, the single hardest problem wasn't indexing — every page was indexed within weeks. It was that search and language models alike kept spell-correcting our name. "MarcTech" became "MarTech," the marketing-technology category, because that entity is vastly better established in the training data. No amount of on-page markup fixed it, because the problem wasn't on the page.
A model's sense of who you are is built from everywhere you appear — business registries, professional networks, third-party citations, the consistency of your name across all of them — not from a file you host yourself. You cannot declare your own identity into existence. You corroborate it from the outside.
That's the uncomfortable truth about being citable: the highest-leverage work often happens off your own domain. Structured data makes a page legible. Entity consistency across the open web makes you legible. The second is slower, you control it less, and it matters more.
What we actually do now
The approach we've settled on, in order of effort-to-payoff:
First, valid JSON-LD on every page, validated in the build pipeline so it can't silently rot. Second, self-contained paragraphs — each one should survive being lifted out of the page and still state something true on its own, because that is exactly how an assistant will use it. Third, ruthless entity consistency: the same legal name, the same descriptions, the same links, everywhere the business appears. Fourth, and only if you're shipping docs or an agent surface, an llms.txt file — built for the tools that actually read it, not as a wish cast at Google.
Notice what's missing: tricks. There's no prompt-injection in your alt text, no keyword-stuffed hidden div. Those get you filtered, not cited.
The honest limit
None of this is measurable the way classic SEO is. There's no rank tracker for "times quoted in a conversation you'll never see." The feedback loop is slow and mostly invisible, and anyone selling you a precise AI-citability score is selling you a number they made up.
So we treat it as hygiene, not growth. Valid structured data, clean extractable prose, consistent identity — these are things a serious site should do regardless, and they happen to be exactly what the AI-native web rewards. The upside is real but unquantified. The cost is a schema block and some discipline.
That's a trade worth making. A text file that the biggest search engine on earth has publicly told you it ignores is not.
Context Windows Aren't Memory: Architecting State for Long-Running AI Agents
Why a bigger context window won't fix a long-running AI agent, and how we architect external state for AI systems that run operations for weeks, not minutes.
Small-Model Economics: Why the Frontier Model Isn't Always the Right Call
When a smaller model beats the frontier model in production. Cost, latency and failure-mode trade-offs from applied AI R&D in Australia, finance to voice.
CPS 230 When Your Service Provider Is an AI System
APRA's operational risk standard never mentions artificial intelligence. It doesn't need to. If your AI sits inside a bank's critical operation, the obligations arrive through the contract — and the transition period has already expired.