Ask ChatGPT

Open a new ChatGPT chat, send one question, and wait for generation to finish (up to 3 min), then return what the model answered and what it read to get there: answers, the web-search queries it issued, and every source behind the reply. Answers come back as a list, not one string. ChatGPT sometimes serves a paired-response variant that answers the same question twice and asks which reply you prefer; that returns two entries and neither is more official than the other. Read the length before reading answers[0]. Text is as rendered, so inline widget labels appear as visible text and mathematical notation linearizes lossily. Sources are split by how the answer used them: the pool a web search returned, whether or not the answer drew on it, versus the sources the answer actually surfaced on screen. That split is the point — it separates a page that was never retrieved from one that was read and passed over. A page that was retrieved and then cited appears in both groups, which is signal rather than duplication. Treat one run as a single draw, not a measurement: the model may search differently, or not search at all, on the next run of the same question. Repeat a question before concluding anything from it. Works logged out, where anonymous rate limits fail fast with a descriptive error and conversationId comes back null.