Going Stellar

Home/Work

Selected work

What we have
actually done.

All of it is work on our own group of e-commerce brands — the honest framing, and also the useful one.

Client names, budgets and margins stay private, so the numbers here are the ones that describe the work itself. Our own brands are not clients, and paid social on the group is largely run by an outside agency — where that matters, it is said in the case.

Selected work

Six pieces of work, with the numbers behind them

Measurement

The CPA spike that never happened

A brand’s cost per acquisition appeared to jump and the budget was cut in response. The forensic pass showed reported conversions collapsing from 158 to 44 a day while real orders held steady at 100–110. The performance was fine; the measurement was broken. The budget went back.

158 → 44reported conversions/day
~105real orders/day, unchanged
Paid search

Proving brand-ad cannibalisation

An account suspension gave us a natural experiment. With brand ads running, organic click-through fell from 57.5% to 22.0%. When the account was suspended the ads stopped — just ten paid clicks that week — and organic recovered to 74.6%. We had been paying for clicks we already owned.

57.5% → 22.0%organic CTR with brand ads on
74.6%after they stopped
Machine learning

Predicting which parcels come back

In markets where most orders are cash on delivery, refusals are the biggest silent cost. We trained a model on the real order history with an honest temporal split: the riskiest fifth of orders contains 39% of all refusals, which is what a confirmation queue needs to be worth running.

0.686AUC, temporal split
1.9×lift on the top 20% risk
Attribution

Teaching the ad platforms about cash on delivery

Cash-on-delivery stores using third-party checkouts were sending no purchase event at all — Meta showed the pixel as inactive and nobody had noticed. We built server-side feeds into both Meta’s Conversions API and Google’s Data Manager API, keyed on the real order, idempotent, and reconciled on a schedule.

Profitability

Moving scale decisions off the dashboard

We built one profitability engine — revenue minus goods, shipping and marketing, ex-VAT, delivered orders only — down to the individual SKU, and let it set the breakeven for every store. Campaigns that looked expensive against a target CPA turned out to be the ones worth scaling; others that looked fine were losing money on every order.

SEO

Executed, not audited

Three stores in one category: alt text written for 1,200 images that had none, unique descriptions with metadata and titles for 467 products, 31 collections reworked and 24 new ones created. On another store, 55 SEO titles that carried registered trademarks were rewritten to remove a real risk of feed suspension.

1,200images given alt text
467products rewritten

Contact

Tell us what is not working

Write in English or Romanian. A person reads it — usually the same person who would do the work — and answers within one business day.