Watch: Gemini picks whichever web page it reads first 92% of the time. Only pages rewritten to its measured preferences ever beat that.

Can we make an artificial intelligence write better marketing copy just by giving it a checklist of do's and don'ts? To find out, researchers used a Gemini model to write more than five hundred commercial web pages. Half of the pages were written from a basic brief. The other half used the same brief, but were guided by a list of thirty-one specific text factors known to help or hurt copy. Then, the same model judged the results. At first glance, the guided pages won nearly fifty-eight percent of the matchups. But the test revealed a massive position bias. The model chose whichever page it read first more than ninety-two percent of the time. To find the real winner, they looked at matchups where the guided page was placed in the disadvantaged second slot. In those tests, the guided page won fifteen percent of the time, while the unguided page won zero. When they ran the test again with the model's thinking mode turned on, the heavy position bias dropped significantly. With room to deliberate, the guided pages won over half of the matchups from the second slot. Ultimately, the experiment shows that giving an AI writer a specific list of rules vastly improves its output, but judging those results requires careful attention to the model's own biases.