Playbook, buying

Research a purchase without falling for weak claims

A chat model is a great research assistant and a lazy fact-checker. We make it hunt for reasons not to buy.

The Curio Team · 2 min readWritten with AI from the sources below and checked against them before publishing.

Ask a model which product is best and you get a tidy list that reads like every review roundup ever written. Ask what breaks, and the conversation changes. We research purchases backwards now: disqualifiers first, recommendation last. Deciding with AI works the way deciding always has. The quality of the answer follows the quality of the question.

Reviews need the same skepticism. Fake ones got common enough that the FTC finalized a rule against fake reviews and testimonials in August 2024, with civil penalties, and it explicitly covers AI-generated reviews. Stars are a signal that someone wanted stars. Patterns across specific, dated complaints are the actual data.¹

The model itself will fill gaps with plausible guesses. Anthropic's guidance for reducing hallucinations is blunt about the fix: give the model permission to say it does not know, and make it cite sources for factual claims. Prices and specs shift constantly, which makes purchase research exactly the terrain where a confident wrong answer thrives.²

One more habit that costs nothing: date-stamp the research. Ask when the model's knowledge ends, and put the current month in your prompt. A recommendation built on a two-year-old product lineup misses the revision that fixed the hinge, or the price drop that changes the answer entirely.

Five steps, roughly ten minutes, before any purchase that stings.

1. Ask for disqualifiers, not winners. Try: list the five most common reasons people return or regret this product. The failure modes teach you more than the feature table ever will.
2. Make it argue the other side. Ask for the strongest case against buying at all, then the case for the model one tier cheaper. If the cheap option survives that argument, buy the cheap option.
3. Demand a source for every factual claim, then click two of them yourself. A spec the model cannot source is a rumor with good posture.
4. Read reviews for patterns, not scores. Three specific complaints about the same hinge beat two hundred five-star one-liners, and recent reviews outrank old ones after any product revision.
5. Sleep on anything over 100 dollars. A day of distance is the cheapest fraud protection there is, and a deal that cannot wait a day was never a deal.

A confident answer is not a checked answer.

We also keep the model away from the final click. It gathers, compares, and argues; the checkout button stays human. That boundary sounds fussy until the first time a misremembered spec nearly costs you a restocking fee.

None of this slows you down much, and it compounds. The disqualifier habit alone has killed more bad purchases in our house than any spreadsheet ever did. The model does the gathering. We keep the judgment. That division holds for a 30 dollar gadget and a 3,000 dollar sofa, and it is the entire skill.