Ask ChatGPT for the best AI automation agency in Serbia for a small business, and it puts Veles Digital first. Ask Claude the same question, and it cites our site as a source. SEMrush scores the site at 99% AI Search Health and 96% Site Health. None of that happened by accident, and none of it is the same work as ranking on Google.
This is what we changed on veles.digital to become the answer, the work usually called AI search optimization or answer engine optimization, and which parts of it you can copy for your own business.
Answering an AI is not the same as ranking on Google
A search engine hands you ten blue links and lets you pick. An AI engine reads the pages itself, decides which ones it trusts, writes a single answer, and names a few sources. You are no longer competing for a position on a results page. You are competing to be the sentence the model quotes.
That changes what you optimize for. Under the hood, ChatGPT, Perplexity, and Google's AI Overviews work by retrieval: they pull a set of candidate pages, rank them for relevance and trustworthiness, and lift passages from the best ones into the answer. Three questions decide whether you make the cut. Can the engine read your full page without running your JavaScript? Can it tell what you are and what you do without guessing? Can it find a clean, direct answer it can quote without rewriting? Get those three right and you become quotable. Miss any one and you are invisible, no matter how good the business is.
What we actually changed on the site
Five things, in rough order of how much they mattered.
We render the whole page to HTML before anyone asks for it. Most agency sites are React apps that build themselves in the browser. An AI crawler that does not run JavaScript, and many do not, sees an empty shell. Every page on veles.digital is prerendered to static HTML at build time, so the crawler gets the complete page, structured data and all, on the first request. You can test your own site in ten seconds: open it with JavaScript disabled. If the page is blank, that is exactly what half the AI crawlers see.
We describe the company to machines, not just to people. The site carries a schema.org graph that spells out, in a format built for this, who we are: an organization that is also a local business in Belgrade, the two founders as real people linked back to the company, a service catalogue, a list of answered questions. A model does not have to infer any of it. It is stated. Naming the founders as people matters more than it sounds, because engines weight content from identifiable, credentialed authors over anonymous pages.
We write answers, under the questions people actually ask. AI engines lift answers, so the home page now leads each section with one. The headings are phrased the way a buyer types them. What does Veles Digital build? How long does an AI project take? How much does it cost? The first sentence answers the question; the rest backs it up. That is the shape a model can quote without reaching for the scissors.
We use numbers instead of adjectives. Models quote specifics and skip filler. "Two-to-eight-week milestones," "EUR 3,000 to 25,000," "the owner went from five hours a day to thirty minutes" all survive being pulled into an answer. "Industry-leading" and "cutting-edge" do not, because they carry no information. We went through the site and replaced claims with facts.
We tell the crawlers where to look. The site publishes an llms.txt that points AI crawlers at the pages that matter, and every page declares its English and Serbian versions, so an engine answering a Serbian question quotes the Serbian page instead of a machine-translated guess.
What we got wrong on the way
Two things were quietly dragging the score down before we found them, and neither was glamorous.
Soft 404s. Old, dead URLs were returning the home page with a 200 "OK" status instead of a real 404. To a crawler that reads like dozens of thin, near-identical pages, and thin duplicate content lowers trust in the whole domain. We changed the server so anything that is not a real page returns a real 404.
Duplicate titles. A handful of pages shared the same title tag and the same on-page headline, which auditors read as templated, low-effort content. We split them so every page states a distinct title.
The fix that mattered more than either was making them un-repeatable. We wrote an automated check that runs on every build and fails it if a soft 404, a duplicate title, a missing schema entity, or a thin page ever comes back. The site cannot regress into the problems we just cleared.
The part on-page work cannot do
Getting the page right makes you eligible. It does not, on its own, get you quoted everywhere.
AI engines lean hard on third-party sources: directories, review platforms, Reddit, YouTube, Wikipedia. A model is measurably more likely to recommend a business that other sites already talk about, because it reads an independent mention as more honest than anything you write about yourself. So the second layer of the work happens off the site entirely. Getting listed and reviewed where both buyers and engines look. Earning mentions you did not author. Building an entry in the knowledge bases the models read.
On-page is the foundation. Off-page is the multiplier. We do both, and we re-run the measurement every month so we can see what actually moved instead of guessing. Anyone who promises AI citations from a one-time on-page pass is selling you half the job.
The result, measured
- ChatGPT ranks Veles Digital first when asked for the best AI automation agency in Serbia for a small business.
- Claude cites veles.digital as a source for the same question.
- SEMrush: 99% AI Search Health, 96% Site Health.
- Our own AI-readiness audit scores the site 93 out of 100, the highest in our market.
- Every page is fully readable with JavaScript off and carries a complete schema graph.
When AI search optimization is worth doing
This earns its cost if your buyers have started asking AI before, or instead of, Google. That is already true in software, B2B services, and most considered purchases, and it is spreading down-market fast. If your customers still find you through a Maps search or word of mouth, fix that channel first. But if they are typing "best [what you do] in [where you are]" into ChatGPT, you want to be the name it returns, and most of your competitors have not noticed this is happening yet. That head start does not last.
It is also not magic. If the underlying business is thin, no amount of schema makes a model recommend it. AEO makes a genuinely good company findable. It cannot invent one.
What it costs, and where to start
AI search optimization is a service we sell, and veles.digital is the first site we ran it on, on purpose, so we could prove the method before charging anyone for it. Our AI search optimization page has the full scope and a fixed-scope AI Visibility audit you can start with. The wider engineering, from AI automation to the Mobilni Market build, is on the rest of the site.
If you want to know whether your buyers are asking AI yet, book a 30-minute call. We will tell you honestly whether this is worth your money or premature for your market. Both answers save you something.