Why pay for this when ChatGPT is right there?

This is the objection we hear most, so it gets the first guide and a fair fight. On 9 July 2026 I asked one question twice: once to ChatGPT, in a logged-out chat with nothing configured, and once to the Nolemy pipeline. Same question, same day. Both answers are below, and both are reproduced exactly.

The question was “Does Uber Eats refund missing items?”, a claim thousands of people live out every week. ChatGPT got it exactly as typed: one message on chatgpt.com, default model, no account. The pipeline got the same claim and did what it always does: searched nine community platforms, killed the marketing, kept 56 admissible quotes, and compressed them into a signed record.

Read the two columns again, slowly. The ChatGPT answer is competent. It is also unfalsifiable: zero dates, zero sources, zero named platforms, and a confidence expressed as maybes (“often”, “may”, “can depend”). You can act on it. You can never check it. The record is the opposite kind of object. Here are three of its 56 quotes, exactly as people typed them:

Those three quotes span June 2026 back to 2021. Same mechanism every time: the refund system works until you have used it, then an account flag turns every future claim into not eligible, evidence be damned. That pattern, repeated across 56 reports on nine platforms, is what the record leads with. It is the single most useful fact about this claim, and the chat window knew it too. Which brings us to the interesting part.

To be fair to ChatGPT, the flag mechanism is in its answer. It sits mid-list, as one factor among several: “the order details, the evidence available, and your account history.” Four words, no explanation, equal billing with paperwork. The 56 people in the record describe “your account history” as the whole game: the wall that ends refunds for anyone who has used the system more than a few times.

That is the fracture this whole guide exists to show. A chatbot weighs text by how often the internet repeated it, and the internet repeats the policy far more often than the outcome. A record weighs text by testimony density: 56 people, sorted with the pain pattern first, because the pain pattern is what you were actually asking about. Same fact, opposite placement: one buried it under a reassurance, the other led with it, because 56 people put it first.

My experiment is one claim. The published tests point the same direction. In April 2026, WIRED asked ChatGPT to cite WIRED’s own product picks in three categories: TVs, headphones, laptops. All three answers were wrong, including headphones WIRED had never tested. OpenAI’s own announcement of its shopping research feature, November 2025, scored the purpose-built tool at 52% product accuracy on multi-constraint queries, against 37% for ordinary ChatGPT search. Those are the vendor’s numbers, about the vendor’s flagship.

The academic result is harsher. ByteDance’s ShoppingComp benchmark (November 2025, 558 scenarios, 35 domain experts) put GPT-5.2’s product-retrieval F1 at 17.76%. Human experts scored 30.02%. One scope note before you quote any of this: these numbers measure shopping tasks, and the Columbia Tow Center’s March 2025 finding that AI search tools botched more than 60% of citation queries is about news sourcing, a different task. Each number counts only for what it tested.

Now the limits, because this entry owes you its own. For settled knowledge, a chat window is the better tool: syntax, recipes, the boiling point of things. Use it there; we do. The instrument changes the moment a question has a counterparty, someone who profits from your belief in a particular answer. Refund windows. Cancellation flows. “Rated Excellent.” On those questions the average of the internet is the marketing budget talking, and you need the fracture instead: the measured gap between the promise and the 56 people who lived it.

One of those 56 might be wrong. All of them together, dated and linked and agreeing on the mechanism, are a fact about the world, and no model produces that fact from memory. Ours is rented, same as everyone’s. The 56 receipts are what you pay for.