Backend work
Back to gallery

AI-assisted workflow for marketplace cards

This case is not about “auto-generating text”; it is about a controlled content workflow: brief, product facts, forbidden claims, editorial accountability, publication, and search-result measurement.

Abstract 3D illustration of an AI assistant, a text document, and editorial tools.
Role
Experiment setup, process design, and measurement
Scale
72 marketplace product cards
Method
AI draft + human editorial review
Result
55% less time, 32.5% lower budget; positions tracked for 16 weeks
Across 72 cards, the team built a production flow around “AI draft + editor”: preparation time fell by 55%, budget fell by 32.5%, and search positions improved while editorial and legal control stayed in place.

Search-position trend for control keywords

A control group of 7 keywords. The lower the number, the higher the card appears in marketplace search. The chart does not frame AI as stronger than copywriting; it shows the applicability boundary: the copywriter track reached a stronger final position, while the AI-draft-and-editor track became a faster and cheaper production option.

Update start AI draft + editor Copywriter from scratch
Keywords
7
AI + editor
209 → 63
Copywriter
276 → 32
04

Business task

The team needed to update descriptions faster, but the category did not allow loose product claims. Each text had to combine search terms, product facts, marketplace restrictions, and readable language.

05

Why not full automation

The main risk was not generation speed but publication quality: forbidden wording, unsupported product properties, repetitions, and the wrong tone could create legal and catalog-quality issues. A human editor therefore remained responsible for the final text.

06

Experiment design

The experiment covered 72 cards and two production tracks: copywriter from scratch and AI draft followed by editorial review. The comparison was based not on subjective text quality, but on search position, production time, and cost.

07

Brief and guardrails

Before drafting, older descriptions were reviewed, search queries were collected, product facts were separated, and forbidden phrases were listed. Each brief included tone, target length, keywords, and restrictions so the model would not drift into generic claims.

08

Editorial QA

The first drafts exposed common failures: repetition, missed stop words, and overconfident product details. The checklist was tightened before scaling, and prompt rules were adjusted after the first publications.

09

Result measurement

Search positions were tracked weekly. In the 7-keyword control group, the copywriter track reached a stronger final position, while the AI track provided a workable trade-off between speed, cost, and quality control.

10

Commercial takeaway

The case shows process economics, not editor replacement: AI drafts are useful when a team needs to process a long tail of cards faster while keeping humans accountable for facts, category restrictions, and publication quality.

Working through something similar?

Discuss a similar task