E-commerce · Case Study

Generating 5,000 SEO Product Descriptions with AI

An e-commerce client had over 5,000 products with inconsistent descriptions that were hurting both SEO and conversion. I architected an AI pipeline using OpenAI to generate SEO-optimised descriptions at scale, with a queue system handling batch processing. The full catalogue was rewritten in 48 hours, organic traffic rose 45%, bounce rate dropped 20%, and roughly 400 hours of manual copywriting were saved.

Results

5,000 in 48 hrs Descriptions generated
+45% Organic traffic
−20% Bounce rate
400+ hrs Manual work saved

The problem

5,000+ products with inconsistent descriptions hurting SEO and conversion.

The approach

I architected an AI pipeline using OpenAI to generate SEO-optimized descriptions at scale, with a queue system for batch processing.

  1. Treated it as a pipeline, not a prompt

    Catalogue-scale generation is a batch data problem. The work went through a queue so it could be retried, rate-limited, and resumed rather than run in one fragile pass.

  2. Wrote to an SEO template

    Each description followed a consistent structure with the product attributes the client actually held, so output was uniform across the whole catalogue.

  3. Kept cost predictable

    Batching and queueing meant token spend was known ahead of the run instead of discovered afterwards.

  4. Measured on traffic, not output

    The goal was organic traffic and bounce rate, so those were the numbers reported — volume of text generated is an input, not a result.

The outcome

5,000 descriptions generated in 48 hours. 45% organic traffic increase. 20% bounce rate drop. 400+ hours of manual work saved.

Stack

Laravel OpenAI API MySQL Queue System

Questions

Is AI-generated product copy safe for SEO?

It is when the output is specific to each product and reviewed, which is why the pipeline writes from real product attributes to a fixed template rather than generating generic prose.

How do you handle 5,000 items without huge cost?

Batch processing through a queue makes token usage predictable and lets the run be rate-limited and resumed, instead of paying for repeated failed passes.

Can this run continuously for new products?

Yes — once the pipeline exists, new or updated products can be pushed through the same queue as they are added.

Building something similar?

If you are shipping AI features or scaling a remote engineering team, I work with founders across the UK, EU, US, and MENA.

Get in touch →