International Ecommerce Run by AI 2009-2026
Loviux
27 stores across 24 countries run by an AI agent. From January to September 2026, orders grew 7×, and in September ChatGPT was already its top sales channel, with no paid advertising.
- Magento Development
- International Ecommerce
- Applied AI
- Content Automation
- AI Chatbots
- Programmatic SEO
01The challenge
The challenge
Loviux is an intimate wellness ecommerce business that has worked with Kiwop since 2009. It wanted to operate like a major international ecommerce player (over 70,000 products, 24 countries, each with its own language, currency, taxes, and carriers) while staying exactly what it is: a two-person team. The bottleneck wasn't the platform, it was the hands doing the work: writing and maintaining product pages in 18 languages took weeks per batch, stock from multiple suppliers changed daily in different formats, and every customer query had to wait until someone was free to read it.
02The solution
The solution
We solved it in four layers. Platform: Magento 2 with a Hyvä frontend, a custom checkout, Varnish, and zero-downtime deploys with automatic rollback, shipping orders the same day they come in. Internationalization: 27 store views from a single backend, each with its own domain, currency, taxes, and shipping rates, full hreflang, 4,500 programmatic SEO landing pages, and sitemaps covering 265,000 URLs. Operational AI: a content pipeline that writes and maintains product pages in 18 languages (from weeks to a single night per batch), stock synced with suppliers several times a day, tracking and emails in the customer's language, and per-order margin reports built by cross-referencing supplier invoices. Customer-facing AI: a model-agnostic chatbot (Anthropic and OpenAI, interchangeable) that recommends products with linked cards, looks up orders after verifying identity, and replies in 18 languages from a 3 KB widget that doesn't touch Core Web Vitals. And since August 2026, a fifth layer: an AI agent that runs day-to-day operations (it monitors orders, payments, and servers, writes and translates content, handles customer email and shipping issues, and fixes errors) while one person approves money and business decisions. Since the relaunch on January 12, 2026, the store has opened twelve new markets, published comparisons, guides, and "The best…" pages in all its languages, and enriched its product feeds for Google's AI search. The growth in orders coincides with those changes and with ChatGPT emerging as a channel, with no paid advertising; no experiment tells us which of them weighs most.
Tech stack
How it's built
Front end
- Hyvä Fast Magento front end
- Tailwind CSS Styling system
Engine
AI
- Anthropic API Claude models via API
- OpenAI API GPT models via API
Data
- Varnish HTTP cache
- OpenSearch Search engine
03Architecture
How a request flows
-
A new product enters the catalog from the supplier feed and is flagged as pending content
-
That night, the AI pipeline writes the full product page and publishes it in the 18 languages across the stores; what used to take weeks per batch is now done in a single night
-
Stock is synced with supplier feeds several times a day (mixed XML, CSV, and JSON formats, normalized into a single format)
-
Every order triggers tracking and transactional emails in the customer's language, and the package ships the same day (weekday orders placed by 5 p.m. CET)
-
Supplier invoices are cross-referenced with each order to calculate the real margin, order by order
-
If a customer asks, the chatbot replies in 3-8 seconds in their language: it recommends products with cards linked to the catalog, or looks up their order after verifying their identity
-
Above all of this, an AI agent reviews operations every day: if it detects a problem (a customer charged twice, for example), it proposes the refund, which a person approves, and deploys the fix the same day
What is AI and what is deterministic
AI components
Since August 2026, an AI agent has run day-to-day operations (orders, payments, servers, content, customer email, and incidents) and records every decision in a work log. Language models (Anthropic and OpenAI, interchangeable) write and maintain product pages in 18 languages, generate the content for programmatic SEO landing pages, handle customer support conversations, and classify incoming queries. Catalog copy that used to take weeks per batch is now written in a single night, cutting editing time by 90%.
Deterministic components
Everything transactional runs on conventional code: stock synchronization with suppliers, prices, taxes and shipping rates per country, the checkout, identity verification before showing order details, and matching supplier invoices to orders for the margin reports. AI never touches prices or stock: it only reads them.
What the system does not do
- The chatbot never shows order details without first verifying the customer's identity; that check runs on the server, not in the model
- It only recommends real products from the catalog, with cards linked to their page: it cannot invent products, prices, or availability
- Every conversation is logged in a review dashboard with its cost per chat; the team audits the bot's responses
- The AI layer is model-agnostic: Anthropic and OpenAI models can be swapped through configuration, with no dependency on a single provider
- The chatbot widget weighs 3 KB and loads lazily: AI never degrades the store's Core Web Vitals
- Every generated product page goes through automatic validation before publishing: banned phrases, character limits on titles and metas, heading and FAQ structure, and mandatory mention of every product variant
- Regression detectors watch the published catalog (language leaks, for example) and trigger surgical regeneration of only the affected products; the pipeline only touches active products with stock
- No automated process can move money: a technical block prevents it, and a person approves every refund, purchase, or campaign
- The agent records every decision in a work log and, when unsure, asks before acting; the store publishes its figures and its mistakes every month
04Measured impact
Results
- 7× Orders in September vs January 2026
- 44% Of September orders came from ChatGPT, ahead of Google
- 1 Person supervises the AI agent that runs the store
- < 0,01 € Per AI support conversation
Verifiable data
Metrics in detail
- Orders: September vs January 2026
- 7×
- Revenue: September vs January 2026
- 11×
- Orders, January-September 2026 vs 2025 (from a very low 2025 base)
- 7×
- Revenue, January-September 2026 vs 2025 (same low base)
- 7×
- Change in average order value: growth comes from more orders
- -5%
- Stores with sales (January-September, 2025 → 2026)
- 6 → 23
- Countries shipped to (January-September, 2025 → 2026)
- 6 → 23
- Revenue outside the Spanish store (January-September, 2025 → 2026)
- 55% → 87%
- Orders arriving from ChatGPT (Aug 6-Sep 30, 2026)
- 35%
- Orders arriving from ChatGPT in September 2026, the top channel ahead of Google
- 44%
- Orders from paid advertising (Aug 6-Sep 30, 2026)
- 0
- Countries with a live store
- 24
- Languages for catalog and support
- 18
- Store views from a single Magento backend
- 27
- Products in catalog
- 70 000+
- Live URLs
- 1 300 000
- Programmatic SEO landing pages
- 4 500
- Lighthouse desktop / mobile / CLS
- 100 / 91 / 0
- Cost per AI support conversation
- < 0,01 €
- Tokens per chatbot message (input / output)
- ~4 300 / ~350
- Customer response time (previously, hours)
- 3-8 s
- People supervising the AI agent
- 1
- Editorial pages written by the AI in summer 2026 (18 languages)
- 148
- Product Q&As generated by the AI
- 8 000
Methodology: Orders, revenue, stores, and channels were queried on October 1, 2026 in Loviux's production database. An order counts if it was placed by a person (no tests, internal orders, or bots), is paid, and was not cancelled or fully refunded; revenue excludes VAT, includes shipping, and deducts partial refunds, in the euros actually collected (orders in other currencies at the amount the payment gateway settled or at the ECB reference rate for that day). The website was relaunched on January 12, 2026, so January mixes days from the old site and the new one. The multiples against 2025 start from a very low year, which is why the headline figure compares September with January 2026. September is still within the withdrawal period and may drop slightly. The source of every order has been recorded since August 6, 2026 (last click, completed with server logs when the browser hides it): there is no channel data before that, and ChatGPT is the only AI assistant that has brought in orders. None of this proves which change caused the growth, because there is no control group. Catalog, stores, countries, and URLs are counted in the backend and in the sitemaps; performance was measured with Lighthouse on the production templates; cost and tokens per conversation come from the chatbot's review dashboard at API pricing, and the chatbot passed end-to-end tests before launch. The agent's figures come from the monthly report and the work log the store itself publishes (loviux.es/informe-mensual-ia); the agent has run Loviux since August 3, 2026.
05The process
The process
-
Platform
Magento 2 with a Hyvä frontend, a custom checkout, Varnish, and zero-downtime deploys with automatic rollback
-
Internationalization
27 store views from a single backend: domain, currency, taxes, and shipping rates per country, full hreflang, and 4,500 programmatic SEO landing pages
-
Operational AI
A pipeline that writes and maintains the catalog in 18 languages, syncs stock with suppliers, and calculates the real margin on every order
-
Customer-facing AI
A model-agnostic chatbot with recommendations linked to the catalog, order lookups with identity verification, and 18 languages, in a 3 KB widget
-
Ongoing expansion
New countries launch in days, not months (the latest were Hungary and Norway), with a conversation dashboard and cost-per-chat tracking to audit the AI
-
AI operator
Since August 2026 an AI agent runs the store: it monitors orders, payments, and servers, writes and translates content, handles email, and fixes errors; one person approves the money
06Showcase
The platform
“On deadlines, they do not just deliver on time, they regularly deliver early. Things we expected to take weeks have landed in days; the AI chatbot went from idea to live on all our stores in a couple of days. Honestly, sometimes we joke that we do not know what these guys are running behind the scenes to move this fast.”
Case chapters
Let's talk.
Your next project?
Tell us your challenge. We'll respond within 24h.
- NDA available
- Response <24h
- Phased proposal
Your first meeting is with a Solutions Architect, not a salesperson.
Talk to an architect