What marketplaces taught me about working with AI

Before I worked with language models, I spent years running stores on marketplaces like Trendyol, Hepsiburada and N11, and selling cross-border into other platforms. I didn’t expect that experience to be relevant to AI. It turned out to be some of the best preparation I had.

Marketplaces and AI systems have something important in common: you don’t control them, you can’t see inside them, and they reward people who learn to work with feedback instead of fighting it.

1. You optimise for a system you can’t see

No seller knows exactly how a marketplace ranks products. You form hypotheses, test them, and watch what happens: a better title, clearer images, a price change, faster shipping. Over time you build intuition for what the algorithm responds to.

Working with AI models is the same skill. You don’t know precisely why one prompt works better than another. You test, compare outputs, keep what works and drop what doesn’t. People who expect a manual get frustrated. People used to experimenting feel at home.

2. Inputs decide everything

On a marketplace, the quality of your listing data determines most of your results. Wrong category, missing attributes, vague descriptions: the best product in the world will still sink.

AI is identical. Messy product data, inconsistent naming and half-filled spreadsheets produce messy output. The unglamorous work of cleaning and structuring data is still where most of the value is. It was true for listings and it’s true for prompts, agents and automations.

3. Scale exposes small mistakes

Managing a few products by hand is forgiving. Managing thousands of listings across several platforms is not: a small error in a template is repeated everywhere at once.

The same happens when you automate with AI. A slightly wrong instruction, copied into a workflow that runs hundreds of times a day, becomes a big problem fast. Marketplaces taught me to test on a small batch, check carefully, then scale.

4. The platform changes, so you need a system, not a trick

Every seller has a story about a tactic that worked brilliantly until a policy update killed it overnight. The sellers who last are the ones with solid fundamentals: good products, good data, good service.

AI tools change even faster. A specific trick that works with today’s model may be irrelevant in six months. What lasts is the system around it: clear processes, clean data, a human who reviews the important outputs, and the habit of measuring results.

5. Customers still decide

Algorithms rank you, but customers buy from you. Reviews, returns and repeat orders always came back to whether the product and service were actually good.

That’s my main filter for AI projects too. Not “is this clever?” but “does the customer, or the colleague using it, end up better off?” If the answer is yes, the technology will find its place. If not, no amount of automation will save it.