Sellphy.

DEMO sells clothes, so the variants are sizes and colors and the button says add to request. Alcona sells fence panels, so the variant is a length dropdown and the button says add to calculator. Gymleco sells gym equipment on a dark theme with feature cards and a specs table.

I wrote the page once. Each shop dresses it from the dashboard, so a seller changes how the page sells without touching code, and a fix I make lands in every shop the same day.

The products screen is a table: thumbnail, status, price, discount, variant count, sizes and categories, with filters across the top. Ticking rows opens bulk edit, so repricing forty shirts before a sale is one form, not forty.

Page Preview takes any URL from the shop and shows what Google, Facebook and X will display for it, with character counts on the title and description. A title under 30 characters gets a warning. The audit table underneath lists every field it checked. Sellers fix the snippet before a page goes out instead of finding out weeks later in Search Console.

The catalog basis is what keeps the page honest. Anchored to eight selected products or a category, the AI writes about things the shop sells, at the prices in the database. The mood presets, editorial, typographic, narrative or catalog, pick the layout family, so the result lands inside the shop's design system rather than beside it.

The blog assistant works the same way. The seller gives a topic and a notes-and-facts field for what the AI must not invent, and it proposes a plan, drafts the post and runs an editorial check. The seller decides what stays. Slug, indexing and the featured image sit in the publishing panel on the right, and Page Preview checks the result before it goes out. Custom pages built this way rank about 75% higher on Google than the pages sellers were putting up before, and a week of content fits in an evening.

The field for what the AI must not invent is the design decision I argued for hardest. A shop owner knows the return window and the fabric weight; the model does not, and it will fill the gap with something plausible. Making the seller write the facts down first is a small chore that removes confidently wrong copy before it is written.

The analytics tour is ten steps. Step four, in the capture, explains the visits chart: the dark line is total page views, the light one is unique visitors per day. Because there are no cookies, unique visitors reset at midnight UTC, and someone who comes back every day for a week counts seven times.

That is the price of counting without cookies, and the tour says so in plain words where the seller will read it. A seller who has never run a shop gets through the first week without a support ticket.

Bulgarian sellers often have the keyboard in Latin mode, so the search reads both alphabets. In the capture the seller typed porachki and the results are for Поръчки, orders, with the matched letters marked. Each hit lists its sub-sections, so the seller lands on the part of the page they wanted rather than the top of it.

I made the same trade everywhere in the dashboard: fewer clicks and less to remember. A tour explains a screen once, and after that search means nobody has to keep a map of the menu in their head. The site editor lets you type into the page instead of finding the field for it. The products table lets you tick rows and edit them together. A seller with a shop to run notices those things before they notice a feature list.

The home, shop, category, about, services and legal pages work this way. Cart, checkout and the dashboard stay dynamic. When a seller saves a change, only the affected paths are rebuilt, so an edit shows up in seconds without touching the rest of the shop. Category filters would normally defeat the cache, so a proxy folds the filter into an internal path while the public URL stays readable.

Once the storefronts were files I no longer needed Vercel. The whole platform runs in Docker on a Hetzner VPS behind Caddy, and the hosting bill is 83% lower than before. Moving all of my own shops onto it saves me 25 or more hours of maintenance a month, which is time that now goes into the product.