The promise was simple: produce personalised marketing content at scale without growing headcount. Two years after LLMs became mainstream in marketing teams, what is actually happening?
What has changed in practice
The most advanced teams are no longer using LLMs to produce content from scratch. They use them to scale a reference piece into multiple variants adapted to specific segments.
A marketing director at a European retailer describes the process: his team writes a reference email in two hours. The model then produces 12 variants adapted to as many customer segments, adjusting tone, key arguments and calls to action, in 20 minutes.
Measured results
On email campaigns, teams that have adopted this approach report open rate increases of 8 to 15% compared to non-personalised campaigns. Conversion rates improve more modestly, between 3 and 7%.
On e-commerce product pages, the impact is stronger. Several retailers have documented conversion increases of 15 to 25% on pages rewritten with contextual personalisation (based on entry channel, customer profile, season).
What does not work
Autonomous long-form content generation remains disappointing. Blog articles, white papers or high-editorial-value content produced entirely by LLM without substantial human editing consistently underperform content written by humans.
Another documented limitation: excessive personalisation. Several brands found that very granular personalisation (more than 50 variants of a single email) produces side effects. Brand voice consistency erodes, and the most exposed customers begin to perceive the automated nature of the content.
Key takeaway
LLMs are effective scaling tools, not original creation tools. Their ROI is real and measurable on short, transactional content. They require human oversight on anything that touches brand identity.