← Blog

Fake review detection on a budget

Spotting fake reviews doesn't require a big-tech budget. For most SMEs, a modest, well-configured setup is enough.

By René Theuerkauf

Online reviews steer what people buy. That makes them a target: fake reviews — written to praise or to sabotage — distort the picture, and generative AI now produces them faster and cheaper than ever. Large platforms throw serious resources at the problem. But what about the small and medium-sized businesses that live and die by their ratings and don't have a machine-learning team?

The question

My research with Prof. Ralf Peters asked a practical version of the problem: how good can fake review detection be on a limited budget, and how should you configure it? Rather than chasing state-of-the-art accuracy at any cost, the goal was feasibility — a setup an SME could actually run.

What the results suggest

• You don't need the biggest model. Detection quality that is good enough for practice is reachable with modest, well-chosen configurations — the last few points of accuracy cost far more than they return.
• Domain matters. Models tuned to the kind of product and language you actually deal with outperform generic ones. Specificity beats size.
• Data quality carries the result. Clean, representative training data does more for detection than model complexity does.
• It's an arms race. As AI writes more convincing fakes, detection has to keep adapting — a setup is never "done".

What it means in practice

If you run an online shop or a review-driven service, the takeaway is encouraging: fake review detection is not reserved for platforms with vast budgets. A focused, domain-specific setup — with attention to data quality and honest evaluation — gets you most of the way. The engineering effort is real but bounded.

Wondering whether this applies to your reviews? Get in touch — happy to give an honest read on feasibility.