Models

Within a row the prompt is the same, so the figures compare directly: the bar is websites per 1,000 words, on a scale shared by every answer to that prompt; the thin line is the middle answer from all models on that prompt. Across rows, prompts differ (from under 1 to 17 websites per 1,000 words), which is why the column heads compare each answer with the other models on its own prompt. Click a model’s name to show its answers on the map.

What this counts, and what it doesn’t. A website is counted once per answer, however many of its pages are linked. The count says nothing about whether a website is reliable or how the answer uses it.

The shaded band is “about the same” (within 10%). avg is the middle answer from the other models on the same prompt; −50% and +100% are equally far from it. Rows with one answer are lighter. K3 is not shown: none of its answers had captured source links.

Both scales run from one fifth to five times the others’ middle answer; an answer beyond that sits on the edge as a dashed circle. Choose a reasoning level to show only those answers. One answer says little, so those maps are lighter.

4 comparisons so far. Reasoning levels are each app’s own labels and are not comparable between vendors. More runs at every level on the same prompts are planned.