A curve-fit EA is one whose parameters were adjusted until its backtest looked good, capturing the accidents of the period along with any pattern in it. It is the most common kind of EA on sale, and three things give it away before any money is spent.
The first is the gap between the backtest and the live record. A fitted strategy produces a beautiful backtest and a live record that starts below it and stays below it, because the edge in the backtest was never in the market. A real strategy runs modestly below its backtest, by roughly the size of the costs the backtest could not model, and stays there. If there is no live record, this tell is unavailable, and that is itself the answer.
The second is the number of inputs. A strategy with a genuine edge can usually be described in a sentence and needs a handful of settings. A strategy with forty inputs, several of them at oddly precise values, has been tuned, and each input is a dial that was turned until the past looked right. Ask how the values were chosen, and what happens to the results when each is moved slightly. A real edge is not fragile to small changes; a fitted one collapses.
The third is the response to a losing month. A vendor who issues new "optimised settings" after a drawdown is fitting again, this time to the last few months, and the new settings will describe the recent past as well as the old ones described the older past. A real strategy has losing months and its settings do not change.
None of these is proof. Together they are enough. A strategy that shows all three is a description of a past that has already ended.