Your Reviews Are Written by LLMs and It's Ruining Local E-Commerce

Building trust in Nigerian e-commerce has always been an uphill battle. We already fight the dreaded "what I ordered vs what I got" on a daily basis. Now, add cheap GPT scripts flooding product pages with synthetic praise, and the whole feedback loop starts feeling completely broken.

Last week, a friend in Akure asked me to recommend a laptop power bank. I hopped onto a popular local marketplace, clicked a top-rated listing, and immediately hit a wall of text that smelled like pure, unadulterated ChatGPT output. Every single review sounded like it was written by an overly eager corporate PR intern who had never actually touched the product.

The Prompt Engineering of Fake Feedback

The issue isn't just that vendors lie—sellers have pushed fake claims since the days of physical markets. The problem is that lying has reached automated scale.

A few years ago, review manipulation was manual work. A vendor would pay a handful of people to manually post copy-pasted comments. Simple string-matching, rate limiting, or IP throttling could catch a chunk of it.

Now? A vendor writes a 15-line Python script, hits an LLM API with a prompt like "Write 30 unique positive reviews for a 20,000mAh power bank, varying sentence length and tone," and pushes them to the platform using headless browser automation.

The result is a wall of vague, polished fluff:

"This item exceeds expectations. Outstanding quality and exceptional utility. Highly recommended for daily use."

It sounds polite, but it gives you zero useful information.

Real feedback from someone sitting in traffic in Owerri or debugging code during a blackout in Jos looks very different. Real human reviews have texture. They say things like: "The fast charging actually works on my Redmi, but the short cable inside the box died after three days so I had to use my own." Real reviews contain specific, unpolished details that cheap automated prompts rarely bother to capture.

Why String Matching Won't Save Us

From a developer standpoint, filtering this garbage out is turning into a messy problem. Old-school spam filters looked for duplicate text blocks or rapid-fire submissions from the same IP address. But generative AI breaks traditional rule-based moderation because every API call yields syntactically unique output.

If you're building a product platform today, basic moderation inputs won't cut it anymore. You have to handle trust at the database and architecture level:

• Hard-link reviews to fulfillment data: If a comment isn't tied to an actual completed order ID with verified delivery status from the logistics partner, it shouldn't touch the product's aggregate rating. Period.
• Vector embeddings over string matching: Instead of checking if text matches word-for-word, pass incoming reviews through an embedding model and measure cosine similarity against recent posts. If fifty reviews cluster in the exact same semantic vector space—saying the exact same three points using slightly different synonyms—you're looking at a script.
• Behavioral telemetry: Real users browse around, abandon carts, click images, and type with natural human pauses. Bot scripts hit endpoints directly or trigger DOM events in mechanical sequences. Tracking mouse movement, keystroke dynamics, and session flow tells you more than the text itself.

The Real Cost of Synthetic Trust

In an environment where Sapa is real and people are carefully counting every Naira, losing money to a fake-review trap stings twice as hard. If buyers realize that 5-star ratings on local platforms are just synthetic noise generated by a merchant running a script from an office in Onitsha, they stop trusting digital marketplaces altogether. They go back to buying through direct WhatsApp calls where they can at least hold a real human accountable.

Regulators in other countries are passing laws against fake AI testimonials, but legal policies don't clean up your database tables. As software builders in Africa's growing digital economy, solving this comes down to what we actually code into our backends.

If we don't build systems that filter out synthetic noise, we're just building shiny catalog apps full of phantom products that nobody trusts.