Every document QA system must decide whether to feed page images, extracted text, or both, yet this choice is rarely examined in isolation. We fix the prompt, judge, and scoring pipeline while varying four commercial model endpoints, two corpora, and two context regimes (gold evidence pages versus the full document). When the image budget permits, page images consistently yield higher accuracy across all document lengths, but the latency and cost of images grow steadily as documents lengthen, whereas text latency remains roughly flat. Text and images also fail on different questions; exactly one representation is correct on 19%–25% of items, indicating that neither subsumes the other. Exploiting this complementarity, a lightweight TF‑IDF router that reads only the question text gains 2.6 points over an always‑text baseline and cuts median latency by about 30% relative to an always‑vision baseline on a document‑disjoint held‑out split.
Review