BEAM

Seedlight BEAM: one place to run your whole eCommerce, with AI agents that know your business →

Journal

Notes from the build

How we plan, build and grow eCommerce platforms. Written from real projects, not theory.

Latest

AI Engineering01

Claude can drive a computer and a browser: what shipped, and what it means for a store

Computer use left beta on 19 August 2026, and a separate browser tool launched the same day. Two different things share that name, though, and only one of them touches your machine. The specification, the costs, and the risk the documentation names outright.Szymon Żynda10 min readRead
AI Engineering02

Claude Fable 5.1 and Mythos 5.1: specs, benchmarks, and what changes for us

Anthropic released two models on 1 September 2026. Full specifications, every published benchmark, a comparison against the rest of the Claude family, the new pricing, and an honest answer on what it changes in how Seedlight works.Szymon Żynda
AI Visibility03

42 percent of our AI crawl was not AI. How to measure yours

For a month we believed AI models had fetched our content 6,600 times. After filtering, 3,934 remained. Three measurement mistakes that inflate AI traffic figures almost everywhere, and how to check your own logs.Szymon Żynda
AI Visibility04

We checked 24 Polish stores: can an AI agent buy from them?

We crawled product pages across twenty-four Polish stores and checked nine fields a shopping agent needs to complete a purchase. Three of them appeared nowhere, and seventeen stores could not be read at all. Method, limitations and the raw numbers.Szymon Żynda
eCommerce05

AI product recommendations: how much data you need before they earn anything

Recommendations learn from interactions per item, not from total orders. A calculation you can run on your own numbers, four approaches with the data threshold for each, and why measuring without a control group always reports success.Szymon Żynda

Collections

Pick a thread and follow it

AI EngineeringHow platforms actually get built with AI: agents, discipline and the craft behind AI-native engineering.Browse articles AI VisibilityBrowse articles eCommerceBuilding stores that sell: platforms, migrations, checkout, product data and everything in between.Browse articles AutomationsAI workflows that take hours off teams: descriptions, translations, feeds and support.Browse articles B2BWholesale online done right: ERP integrations, contract pricing, ordering portals and self-service.Browse articles Global ExpansionSelling across borders: new markets, languages, logistics and channel strategies in Europe.Browse articles MarketplaceMulti-vendor platforms and marketplace channels: liquidity, settlement, feeds and operations.Browse articles ScalingWhat breaks as stores grow, and how to build so it does not: performance, teams, processes.Browse articles

All articles

AI Engineering01

Claude can drive a computer and a browser: what shipped, and what it means for a store

Computer use left beta on 19 August 2026, and a separate browser tool launched the same day. Two different things share that name, though, and only one of them touches your machine. The specification, the costs, and the risk the documentation names outright.Szymon Żynda10 min readRead
AI Engineering02

Claude Fable 5.1 and Mythos 5.1: specs, benchmarks, and what changes for us

Anthropic released two models on 1 September 2026. Full specifications, every published benchmark, a comparison against the rest of the Claude family, the new pricing, and an honest answer on what it changes in how Seedlight works.Szymon Żynda11 min readRead
AI Visibility03

42 percent of our AI crawl was not AI. How to measure yours

For a month we believed AI models had fetched our content 6,600 times. After filtering, 3,934 remained. Three measurement mistakes that inflate AI traffic figures almost everywhere, and how to check your own logs.Szymon Żynda8 min readRead
AI Visibility04

We checked 24 Polish stores: can an AI agent buy from them?

We crawled product pages across twenty-four Polish stores and checked nine fields a shopping agent needs to complete a purchase. Three of them appeared nowhere, and seventeen stores could not be read at all. Method, limitations and the raw numbers.Szymon Żynda9 min readRead
eCommerce05

AI product recommendations: how much data you need before they earn anything

Recommendations learn from interactions per item, not from total orders. A calculation you can run on your own numbers, four approaches with the data threshold for each, and why measuring without a control group always reports success.Szymon Żynda9 min readRead
Automations06

AI assistants for online stores: what works today and what is still too early

One label covers two different products: the advisor your customer talks to, and the assistant working in your back office. Buyers ask about the first, the second pays back sooner. Why that is, how to test your catalogue in an hour, and the three cases where an advisor earns from day one.Szymon Żynda9 min readRead
AI Visibility07

What an AI agent sees when it comes to buy in your store

An agent never looks at your page, it receives a record. Field by field: what has to be in that record before a purchase can close, three silent failures worse than a missing field, and the standards decision (ACP, UCP, the Shopify agent flow) that nobody will make for you.Szymon Żynda11 min readRead
B2B08

B2B ordering portal: what it must do before your reps stop retyping emails

A portal only empties the sales inbox once it closes a whole order: this buyer’s price, a reorder from history, a pasted SKU list, the credit limit, real availability dates, approval on the buyer side, a PO reference and an honest status when the ERP goes quiet. Each requirement with the condition it fails without, plus an acceptance test to run on a demo.Szymon Żynda12 min readRead
eCommerce09

Who should build your eCommerce platform: a freelancer, an agency, a software house or an in-house team

Four delivery models held to the same standard: when each one is right, how it bills, what is left when it goes away and where it usually breaks. With a decision table, the questions to ask before signing, and the things builders do not publish about themselves. Checked 17 August 2026.Szymon Żynda12 min readRead
AI Engineering10

How an agency works with Claude Code on a client’s store

An AI agent writing code on your platform is a trust question, not a productivity question. What agents are never allowed to touch, how a change reaches production, how you approve it without reading pull requests, and what stays yours when the contract ends.Szymon Żynda10 min readRead
Global Expansion11

Medusa multi-language: how translations work and what breaks at scale

Since version 2.12.3 Medusa stores translations natively, which settles where multilingual content lives and none of what actually breaks in bulk translation: markup, brand names, units and field limits. Here is what the Translation Module covers as of 17 August 2026, the six failures that repeat, and the checks that catch them before publication.Szymon Żynda9 min readRead
B2B12

Framework agreements in German B2B: what a Rahmenvertrag settles

German wholesale buyers start with an agreement, not an order. What a Rahmenvertrag usually settles, which billing models follow from it, where German law fixes payment periods and late payment interest, and what your platform has to carry so the terms survive contact with real orders.Szymon Żynda11 min readRead
eCommerce13

What an eCommerce build costs on every platform, and who publishes the number

Public implementation ranges for eight eCommerce platforms, and next to every figure the column nobody else prints: who published it. Including a section on the numbers you should not trust. Checked 13 August 2026.Szymon Żynda12 min readRead
Automations14

AI agents running your store: what they do on Monday morning

An AI agent in a store is really two things: a workflow triggered by an event, and an agent answering questions about live data. Concretely: what happens when 200 products land, when a feed rejects 30 items, and when a customer asks where the parcel is. Plus the permission model we insist on before any AI touches an order panel.Szymon Żynda10 min readRead
eCommerce15

BEAM + Medusa vs Shopify Plus and an agency: two routes, not two products

A comparison of offers rather than engines: what you get on day one, who runs it, whose platform it is and what you keep if you part ways. With a three-year model built from substitutable assumptions, and our published price ranges.Szymon Żynda10 min readRead
eCommerce16

Shopify alternatives for growing brands: seven routes and one test

Seven realistic directions for a brand that has outgrown Shopify, six of which we do not sell. Each one with two numbers, platform cost and build cost, because every option on this list has to be implemented by someone. With an honest section on when to stay.Szymon Żynda13 min readRead
eCommerce17

Medusa vs WooCommerce: free to start, paid to maintain

Both licences cost nothing. The difference sits in the build, the running cost and the price of the first change after launch, so we split the bill into three explicit layers. Twelve dimensions side by side, plus an honest section on when WooCommerce is the cheaper answer. Vendor data checked 11 August, build ranges 12 August 2026.Szymon Żynda14 min readRead
eCommerce18

Medusa vs Shopify: which one fits, and when switching is a mistake

Medusa and Shopify compared across eleven dimensions: licence, cost to implement, cost to change after launch, checkout, API limits and lock-in. Figures checked at source, current as of 12 August 2026.Szymon Żynda10 min readRead
eCommerce19

Black Friday: an engineering checklist for your platform

Black Friday 2026 falls on 27 November. An engineering schedule: what to measure in September, what to rebuild in October, what to freeze in November.Szymon Żynda12 min readRead
Marketplace20

Marketplace Take Rate: How Much Commission to Charge and How to Defend It

What a marketplace take rate is made of, what real platforms charge, and how to work out your own ceiling from one seller order.Daniel Pawłowski13 min readRead
Scaling21

Your eCommerce team after launch: who you actually need

After launch you need four functions covered, not four new hires. Which roles scale with order volume, and which are better bought than employed.Szymon Żynda11 min readRead
AI Engineering22

The agent did exactly what we wrote: why AI automations fail quietly

AI automation rarely fails loudly. It follows the instruction literally rather than the intent, so the same defect lands in every single output. Three measured cases from our own work, applied to descriptions, attributes and feeds.Szymon Żynda9 min readRead
Automations23

AI catalog translation without an agency: where translation ends and localization begins

Bulk catalog translation is one of the best first AI projects in a store: it runs on data you already own, and you catch every mistake before it goes live. On one condition, that you know where translation ends and localization begins, because that is where most failed rollouts break.Szymon Żynda8 min readRead
Global Expansion24

Entering the German market: an eCommerce checklist

Translating your store is the smallest part of selling in Germany. The rest is packaging registration, VAT and OSS, the Impressum, product data in the format German channels expect, payment methods you may not offer yet, and returns run as a process. Here is the checklist in the order that works.Szymon Żynda9 min readRead
AI Engineering25

eCommerce skills for AI agents: why we open-sourced ours

We have published a set of AI agent skills specialised in eCommerce: catalog and feeds, structured data and AI visibility, 2026 compliance, migration and a pre-deploy gate. What is inside, how to install it, and why the safety rules in a skill matter more than the list of steps.Szymon Żynda8 min readRead
Automations26

AI for order support: automating „where is my order" (WISMO) questions

„Where is my order" (WISMO) is the most common and most repeatable category of contacts in eCommerce. How eCommerce customer service automation takes it over in a draft-first model: AI writes a draft with order context and status, and a human approves or escalates.Szymon Żynda9 min readRead
eCommerce27

POAS, not ROAS: measure profit from ads, not revenue

ROAS shows the revenue an ad touched, not the result. Two campaigns with identical ROAS can leave one in profit and the other in loss, because margin differs. POAS (Profit on Ad Spend) measures what actually stays: definition, formula, a worked example and the scaling trap.Szymon Żynda9 min readRead
AI Visibility28

How AI describes your brand: Brand Visibility as a new KPI

How ChatGPT, Perplexity and Gemini describe and recommend your brand is becoming a metric of its own, because customers increasingly ask an AI assistant instead of Google. How to check for yourself what a model says about your brand, what shapes that picture, and what you can actually fix, without promising the AI will start recommending you.Szymon Żynda8 min readRead
AI Engineering29

Marketplace feeds with Claude Code: from export to validation

Same catalog, but every channel wants its feed its own way: different field names, different formats, its own taxonomy and hard rejection rules. Here is how Claude Code maps catalog columns onto a channel's fields, fills gaps, validates types and values, and flags offers that will be rejected before you upload. On a copy of the data and with a diff review.Szymon Żynda9 min readRead
AI Visibility30

Zero-click commerce: the end of the click and what it means for your store

More and more often the customer gets an answer to a buying question right away, in the results or from an AI assistant, and never lands on your site. What that means for your store, why it is not the end of the world, and how to become the source of the answer the model shows the buyer instead of losing traffic.Szymon Żynda9 min readRead
AI Visibility31

Agentic commerce: how to prepare your store so AI recommends it

Agentic commerce means AI agents shop on the buyer behalf: they search, compare and increasingly close the purchase. You prepare your store for this not with a campaign, but by making it readable to a machine. A concrete checklist, plus how to check whether AI already recommends you.Szymon Żynda9 min readRead
AI Visibility32

GEO for eCommerce: how to land in ChatGPT and Perplexity answers

GEO is not a separate kind of magic detached from SEO. It is the same good SEO, extended to visibility in generative answers. What to actually do so your brand and products show up in ChatGPT and Perplexity, and how to measure that presence, without promising miracles.Szymon Żynda9 min readRead
AI Visibility33

Structured data for AI: the Product, Offer and FAQ agents look for

For search engines and AI agents to understand your store unambiguously, you need a handful of schema.org types in JSON-LD: Product, Offer, AggregateRating and Review on real data, BreadcrumbList, FAQPage and Organization. Which ones matter, how to implement them correctly, and the mistakes that draw a penalty instead of visibility.Szymon Żynda9 min readRead
AI Engineering34

Claude Code for eCommerce: cleaning a catalog CSV step by step

A messy catalog CSV export can be cleaned in a few passes: split variants, validate EAN/GTIN codes, normalize units and colors, repair encoding, generate slugs, flag duplicates and gaps. Here is the full process and how to brief it, on a copy of the data and with a diff review before you save.Szymon Żynda9 min readRead
AI Engineering35

Agentic engineering: how we build ecommerce platforms with AI agents

An AI agent writes code fast, but the bottleneck is verification, not writing. We show the process where gates, an adversarial review and a human on irreversible actions deliver speed and safety at once.Szymon Żynda8 min readRead
Automations36

eCommerce automation: where to start and how many hours it really saves

Where do you start automating an online store, and how much time does it genuinely take off your team? Instead of one magic number: a qualification rule, seven repeatable tasks with hours before and after, and the work that always stays with a person.Szymon Żynda10 min readRead
B2B37

B2B contract pricing: the layers of price and which rule wins

B2B pricing is not one discount but several models running at once: price levels, customer groups, individual prices, volume tiers and framework agreements that fix terms before the first order is placed. Here is the order in which the rules resolve, what a framework agreement has to settle, and what your platform and ERP need to carry.Szymon Żynda12 min readRead
Scaling38

What Breaks at 10x Traffic: The Engineering of Scaling an eCommerce Store

When traffic suddenly spikes, a store rarely fails for a single reason. The database gives out first, then the missing cache, and finally the synchronous integrations. Here is what breaks as traffic grows and how to prepare your platform for the peak before it arrives.Szymon Żynda11 min readRead
eCommerce39

Which ecommerce platform? Criteria, not rankings

There is no best ecommerce platform, only one that fits your business. Instead of a ranking: nine criteria that actually decide your migration target, plus an honest when SaaS is enough.Szymon Żynda9 min readRead
eCommerce40

The checkout that stops losing carts: a 10-point audit

Over 70% of carts end up abandoned, but the recoverable ones have repeatable, concrete causes. A 10-point checkout audit: what to check, the typical mistake and how to fix it.Szymon Żynda10 min readRead
eCommerce41

12 signs your store has outgrown its SaaS platform

Twelve signs a SaaS platform is limiting your store, what each one really costs, and a simple framework for when to migrate and when to stay.Szymon Żynda12 min readRead
B2B42

Wholesale digitisation in 90 days: a real plan, not a mega-project

Digitising wholesale does not have to be a year-long project with an uncertain end. Here is a 90-day plan in three phases where each ends with something that works, not a slide deck.Szymon Żynda6 min readRead
Marketplace43

The marketplace cold start problem: supply or demand first?

A marketplace has a chicken-and-egg problem: sellers will not come without buyers, buyers without listings. Here are proven ways to solve the cold start, with the honest condition under which none of them work.Daniel Pawłowski5 min readRead
eCommerce44

The true cost of owning your eCommerce platform: a 3-year TCO

SaaS looks cheap in month one and expensive in year three. We break down the total cost of ownership: subscriptions, fees and apps versus building a platform you own.Szymon Żynda11 min readRead
B2B45

B2B store and ERP integration: an architecture that does not hurt

The most expensive word in wholesale is retyping. Here is the store-to-ERP architecture, sync rhythms and error handling that make orders flow on their own.Szymon Żynda10 min readRead
Automations46

AI product descriptions done right (without hurting SEO)

A description generator is not a chat window. Here is the workflow that holds brand voice across thousands of SKU, passes human review and meets Google guidelines.Szymon Żynda9 min readRead
AI Engineering47

AI-native is not vibe coding

We build every project with Claude Code. But AI without engineering discipline is chaos in production. Here is how we combine both.Szymon Żynda5 min readRead
eCommerce48

You are not buying hours: why fixed scope wins

Hourly billing shifts all the risk to the client. In BEAM every stage has a fixed price and a defined outcome. Here is why we price this way.Szymon Żynda4 min readRead
eCommerce49

Migrating a store without stopping sales

The biggest fear of replatforming: "what about sales during the move?". Our migration playbook answers with specifics, not reassurances.Szymon Żynda6 min readRead
More to exploreExample buildsBEAM platformFAQRSS

Newsletter

The Journal, straight to your inbox

New articles and lessons from real builds, every now and then. No spam, unsubscribe with one click.

By subscribing you agree to receive the newsletter. Details: privacy policy.

YouTube

Daniel & Simon | Commerce & AI

The founders talk commerce and AI: what works in real stores, what breaks and what is coming next.

Watch the channel ↗

Prefer specifics about your platform?

A 30-minute call with the team that ships these builds weekly. We come prepared with an initial read of your situation.

Book a Blueprint →