How Websites Talk to AI Today (and How to Do It Safely and Ethically)

Just a few years ago, having a website that Google liked was enough. But today, more and more people aren't looking at search results — they're looking at what an AI assistant found. Instead of twenty links, they get one coherent answer, assembled from what the AI read across various websites. And that changes everything, because it's no longer just about being found — it's about being understood, cited, and recommended by AI.

This new field is called GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization) — optimisation for generative search engines. It's a young discipline, but it's already clear what it's built on.

What AI models genuinely value today

Unlike classic SEO, where stuffing in keywords that matched frequently searched terms often worked, generative search engines are looking for something different: a clear, structured, and trustworthy answer to a specific question.

In practice, that means:

• Structured data (schema.org / JSON-LD). When your website clearly labels what kind of business you are, what services you offer, what reviews you have, or what your customers most frequently ask, AI can much more easily "read" and use that information. It's like handing the search engine a neatly filled-in form instead of a disorganised pile of papers.
• Questions and answers right up front. Websites that have a FAQ section phrased exactly the way people ask questions ("How much does a website for a sole trader cost?", "How long does it take to build a website?") and give the answer in the first two sentences have a much better chance of being quoted verbatim by AI.
• Consistency across the internet. When your business name, address, and contact details match on your website, in your Google Business profile, and in directories, AI connects you with greater confidence — and trust is the currency that generative search engines rely on perhaps even more than traditional ones.
• Links and mentions from elsewhere. Reviews, interviews, citations on other websites — all of these work as "references" that AI treats as a trust signal, much like how we make decisions based on recommendations from people we know.

For small businesses and sole traders, there's actually good news here: it's not about having the most money for advertising, but about having the clearest and most honest content. That's a battle a smaller player can win too.

The dark side: when someone deliberately deceives AI

This new era also brings a new vulnerability. Language models don't read the web like a human — they read text. And if an instruction is hidden within that text that a human visitor never sees (for example, white text on a white background, or in code that isn't normally displayed), the model may interpret it as a directive to follow. This phenomenon is called prompt injection — embedding hidden instructions designed to make AI behave differently than its user intended.

In practice it can look harmless: hidden text tells AI to speak about a company only in superlatives, to ignore negative reviews, or to recommend a specific product regardless of the context of the query. In worse cases, it extends to attempts to make AI agents — which today can already shop, book, or fill in forms on their own — carry out something the user never wanted.

It's a real risk that major technology companies are now taking seriously. How do websites and AI systems defend against it today?

• By separating "instructions" from "data" at the level of how AI models process inputs.
• By filtering and validating the content that a model loads from the internet.
• By increasingly sophisticated recognition of text on a page that is clearly not meant for a human, but only for a machine.

Where the ethical boundary lies — and how to use it to your advantage

It's important to draw a clear line between two things here. Hiding instructions that change AI behaviour against the user's interests is manipulation. But adding supplementary, descriptive content intended primarily for AI — content that merely clutters the page for a human but gives the assistant useful context — is perfectly fine and is common practice today.

Example: on a pricing page you can visually present just a clean table, but in the structured data (schema) you can add a detailed description of exactly what the price includes, what the typical delivery time is, and what technologies you use. A person doesn't need to read it in the text — they see a clean website. AI will find it and use it when answering someone's question about "who does quality custom websites at a reasonable price".

The difference lies in intent: you are adding information, not changing the truth. You're not convincing AI to lie on your behalf — you're simply giving it more context so it can answer more accurately for the person asking.

What specifically changes for websites today?

Many of today's websites are built on a principle that was considered cutting-edge until recently — React SPA (single-page application). For a human visitor that's fine, but for AI models it's a problem — because these systems can't "browse" a website the way a human does. If content is loaded dynamically via JavaScript, the model often can't see it at all. What exactly is the problem? Most generative models today work by downloading an HTML page and processing it as text. If content is only loaded after JavaScript runs, the model doesn't see it — and you lose the chance of AI recommending you. What an AI crawler typically sees verbatim on a React SPA is this:

<body>
          <div id="root"></div>   ← empty
          <script src="assets/index.js"></script>
          </body>

All content lives inside JavaScript. AI agents then read either:

• Rendered HTML — JavaScript is rendered by Googlebot (with a delay of days to weeks) and by Gemini via Google's infrastructure. The other major players — ChatGPT, Perplexity, Claude — do not execute JavaScript by default.
• Raw HTML — most AI crawlers see only the static content from index.html and the meta tags in the header.

The good news is that there are ways to help AI readers without rewriting your entire website. You just need to add three things to your existing HTML file — the first place AI crawlers always look, before JavaScript ever runs.

Step 1: Structured data (JSON-LD)

JSON-LD is a data format embedded directly in the <head> section of your HTML — it's therefore present in the document before the browser runs any JavaScript, and AI models can always read it. It works like a neatly completed form about your business: who you are, what you offer, what your prices are, how to contact you. For AI, such a description is more reliable than trying to infer the same information from the visual design of the page.

<!-- Insert into <head> in index.html -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "LocalBusiness",
  "name": "Your company",
  "description": "What you do and for whom — one sentence.",
  "url": "https://vaseweb.cz",
  "email": "kontakt@vaseweb.cz",
  "priceRange": "5 000 – 50 000 Kč",
  "areaServed": "Česká republika",
  "knowsLanguage": ["cs", "en"]
}
</script>

Validate your data at search.google.com/test/rich-results — you'll see exactly what Google (and indirectly AI models) read from your JSON-LD.

You don't need to include reviews in your JSON-LD. If you genuinely have them, Google recommends displaying them directly on the page — ratings hidden from users but visible only to bots go against its guidelines. Far more valuable than ratings in your own JSON-LD is having a review of you on Google, a comparison platform, or an industry portal — those are the sources AI models actively cite.

Step 2: Prerendering or SSG — the real solution

The noscript tag is often mentioned as a safety net for crawlers without JavaScript — but let's be honest: it's not a real solution. It's a bare minimum that may help slightly, but AI models don't look to the noscript block as a primary content source.

The real solution is for the page to generate HTML on the server, without needing to run JavaScript on the client side. Three realistic paths:

• Next.js or Astro — frameworks designed for server-side rendering and static export. HTML is ready immediately; an AI crawler reads it without any issues. Ideal for a new website or a complete rewrite.
• Vite SSG plugins (e.g. vite-plugin-ssg) — if you have an existing Vite/React project, these plugins generate static HTML for each page at build time. A smaller change than switching to a different framework.
• Prerender.io or a similar cloud layer — intercepts requests from crawlers, prerenders the page for them, and returns static HTML. A code-free solution, but with a dependency on a third party.

Noscript as an emergency fallback: If none of the above paths is realistic in the near term, add a noscript block with key information to index.html. It's not a proper fix, but it's better than nothing — and takes five minutes to implement.

Step 3: llms.txt — a summary for AI assistants

Just as robots.txt tells crawlers what they may or may not index, llms.txt is an unofficial format (not yet widely standardised) that aims to give AI models a clear summary of an entire website — in plain text, formatted as Markdown. You place it in the root of your site, where it becomes available at yoursite.com/llms.txt, and you link to it from the HTML header with a single line.

<!-- Link to llms.txt — insert into <head> -->
<link rel="alternate" type="text/plain"
      href="/llms.txt" title="AI-readable site summary" />

The file itself is plain Markdown — write it as if you were sending an email to a colleague who has never heard of you:

# Your company

> Who you are, what you do, for whom — two or three lines.

## Services

- **Basic website** — from 15,000 CZK, delivery in 7 days (description of what's included)
- **Advanced website** — from 30,000 CZK, delivery in 14 days

## Portfolio

- [Project name](https://klient.cz) — brief description of the result

## Contact

- Email: kontakt@vaseweb.cz
- I respond within 24 hours.

Adoption of llms.txt is still low: Ahrefs data showed that 97% of sites with the file recorded no visits from AI bots, and Google does not actively support it. That said: implementation takes an hour, costs nothing, and is a cheap bet on the future — and a signal that you're thinking about how well-structured your content is.

How to verify it yourself

Before ordering an audit or a third-party tool, try a direct test: open your website in a browser and press Ctrl + U (Cmd + U on Mac). You'll see the raw HTML — exactly what an AI crawler sees. Press Ctrl + F and search for your main heading or business name.

• Content appears → the crawler will read your site normally.
• You only see an empty <div id="root"></div> → the content lives entirely in JavaScript and AI won't find it.

An automated version of this test — including a check of JSON-LD, meta tags, and the result without JavaScript — is available as a free tool directly at annlibertas.eu.

Key takeaways

Whatever business you're in, this shift affects you — more and more people are turning not to a search engine but straight to an AI assistant when looking for suppliers, tradespeople, and services. And even when searching on Google, AI search results for a given query typically appear first. Websites that have honest, well-structured, and clearly described content will gain an edge in this new landscape — not through a trick, but by making life easier for a system whose job is to help people find the best solution.

And perhaps it's even a refreshing thought: in an era where algorithms and manipulation are talked about more and more, the best long-term strategy is simply — to be honest, clear, and useful. AI values that just as much as any human does.