A supplement store Google could barely read, and AI had never heard of.

The starting point

What was actually wrong with the old site?

Almost everything a search engine reads, and nothing a customer would notice. There was no Google Analytics and no Search Console, so nobody could answer basic questions like which pages people landed on or what they searched to get there. Google Merchant Center did not exist, which meant no Shopping listings at all. Thirteen pages carried two H1 headings each because the theme printed a hidden duplicate. One FAQ schema block was broken by a single stray quotation mark, so Google was reading half of it and discarding the rest. Meta titles were missing or still in English on a Bulgarian store.

The content was thinner than the catalogue deserved. Four products, and not one page explaining what berberine is, what ashwagandha is, or why the dose on the label matters. A brand selling supplements made in Bulgaria had no page saying who made them.

The lead

Why did we build this for AI answers and not just Google rankings?

Because the shape of search changed and most sites are still built for the old one. AI Overviews now appear on roughly half of all queries, and AI Mode answers without showing ten blue links at all. When Google or ChatGPT builds an answer, it does not read your page. It retrieves a passage.

That single fact rewrites how you write. A retrieval system embeds the chunk, not the page, so a sentence that says "it reduced their symptoms by nearly half" names nothing and matches nothing. Every sentence has to stand alone. Name the substance, the number, the body, the date, inside the sentence.

So every article on Biovaya now opens on a definition, carries question-shaped H2 headings with the answer in the first forty words, and cites its sources by name: the European Medicines Agency, EFSA, the French agency ANSES, the Court of Justice of the European Union, the World Health Organization, the New England Journal of Medicine. Twenty-one URLs now carry FAQ schema.

Here is the part most supplement brands miss. AI assistants cite what they can check. A page claiming "clinically proven" cites nothing, so nothing cites it back.

The unfair advantage

What does Bulgarian search have that English search doesn't?

Room. Search a health question in English and you compete with WebMD, Healthline and twenty years of optimised content. Search the same question in Bulgarian and Google frequently returns no People Also Ask box at all, because there is not enough structured Bulgarian content for it to build one.

That absence is the opportunity. Question-shaped headings and FAQ schema, written properly in Bulgarian, are close to free upside in a market where almost nobody does it. The same is true of the AI answer engines: they are answering Bulgarian health questions today from whatever Bulgarian-language source they can find, and the supply is thin.

Biovaya's competitors are optimising for the words they think customers use. We wrote for the words Bulgarians actually type.

The build

What did we actually build in ten days?

Twenty-four pages and eighteen articles. An encyclopedia of twelve ingredient pages, each one covering a single substance: what it is, where it comes from, the dose, who should avoid it. Four solution pages and a hub, built around the problems people search for rather than the products we sell. A new homepage, an about page carrying the founder's real story, and a where-to-buy page. All four product pages rebuilt.

Underneath that: Google Analytics 4 and Search Console connected and verified, Merchant Center set up and the feed submitted, an XML sitemap filed with indexing requested, and robots.txt opened to the AI crawlers that matter, GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot. Blocking those is a common accident and it makes a brand invisible to exactly the systems now answering questions on its behalf.

We also automated the review flow through Klaviyo, two emails per order, and ran it back across the existing customer base. Reviews are not decoration. Genuine, dated, first-party reviews are one of the few signals that separates a real brand from a drop-shipper, and both Google and the answer engines can tell.

The tooling

How does an agency that doesn't build on WordPress rebuild a WordPress store?

Honestly, WordPress is not our stack. We build on modern frameworks and we would not have chosen WooCommerce, Woodmart and twenty-five plugins for this brand. But Biovaya was already there, with a live catalogue and an order history, and a platform migration would have cost the client months it did not have. So we kept the platform and changed the way we work on it.

Every change ran through Claude, Anthropic's AI, driving the site's own REST API directly rather than a person clicking through the admin panel. That sounds like a technical footnote. It is the reason twenty-four pages and eighteen articles moved in ten days without a single broken URL.

The method is what matters, not the tool. Each change is specified precisely, applied programmatically, then read back from the live site and compared against what was sent. After we rescaled the logo, we checked every one of 234 images and 72 internal links on the affected pages before calling it finished. When we replaced the product cards inside fifteen articles, we confirmed all fifteen matched byte for byte after saving. A person doing that by hand takes a week and starts missing things by hour three.

The judgement stays with us. Which claims are lawful, which pages deserve to rank, what a Bulgarian buyer actually searches for at eleven at night: none of that is automatable, and none of it should be. What is automatable is the part where mistakes happen.

The uncomfortable part

Why did we delete most of the selling copy?

Because it was illegal, and because deleting it made the site more visible, not less.

Food supplements in the EU may only carry health claims from an authorised list. Biovaya's pages were full of the kind of sentences every supplement brand writes and almost none may. Google Shopping had already noticed and rejected most of the catalogue.

So we went to the source, pulled the authorised claims from the Bulgarian text of Regulation 432/2012 word for word, and rebuilt every product statement on top of them. It turns out the law lets you say the commercially valuable thing, as long as you credit the right ingredient. Zinc contributes to the maintenance of normal testosterone levels in the blood. That sentence is authorised, checkable and it belongs to the mineral, not the plant.

Then the useful part. A page that says "clinically proven" and cites nothing is invisible to an answer engine. A page that names a regulation, quotes a 2024 committee decision and links the primary source is exactly what a citation looks like. Compliance and AI visibility turned out to be the same job.

The measurement

What happens now, and what will the data change?

This is why the analytics went in first. From here the site stops being rebuilt on judgement and starts being tuned on evidence.

Search Console tells us which Bulgarian queries the new pages are appearing for and where they sit, so we refresh the pages that are close rather than writing new ones nobody asked for. The Generative AI report shows impressions inside AI Overviews and AI Mode, which are separate targets.

Analytics 4 needed one custom job. Its built-in AI channel misses referrer-less traffic, so we group ChatGPT, Perplexity, Claude, Gemini and Copilot into their own channel and report conversions rather than sessions, because AI referrals arrive further down the decision than search traffic does.

On top of that, a monthly run of tracked prompts in both Bulgarian and English, asking the assistants the questions a customer would, and logging whether Biovaya appears and which page gets cited.

Then we act on it. That is the whole point of the ten days.

The honest bit

What nearly caught us out?

We got a piece of the regulatory reasoning wrong and had to redo five pages six days after building them.

Roughly 1,500 health claims for botanical ingredients sit unassessed in EU limbo, neither approved nor rejected. We had treated that gap as room to speak carefully, which is what most of the industry does. Then we read the Court of Justice ruling in case C-386/23 properly. It closed the gap in April 2025: while the assessment is unfinished, botanical claims touching psychological or behavioural function are not permitted at all.

So we rewrote the four solution pages and the hub. Nobody made us. No client complained, no regulator wrote, and the pages had already been signed off. We found it in our own work and fixed it inside a week.

We are telling you this because no agency admits anything, which is precisely why you should ask every agency you speak to what they got wrong on their last project. If the answer is nothing, they either are not looking or are not telling.