One of my most common use cases for AI is to research topics before I have to talk about them in public. Here are tactics I use to get results that go beyond the first page of a Google search.
Don't let it answer from memory
An AI will tell you something that stopped being true a year ago. Its training stops at a cutoff date, and it often can't tell whether what it knows is still true, so it will confidently give you an outdated answer. That's why I often remind it to look up today's date and tell it not to trust its memory until it has web search results to back it up.
I also remind the AI to read the actual page before it states anything, which helps catch stale stats, misattributed quotes, or numbers that aren’t in the source they’re credited to.
Say what you're after, and set a floor
I give the AI the subject I’m interested in, and a few areas I already care about, with the instruction to treat those areas as a floor, not a ceiling, and look beyond them. Without that line, it’s satisfied once it has answered what I asked, and stops digging.
I also leave out my own answer. If I tell it what I expect to find, it leans toward finding it, so I keep my read out and decide afterward, based on the results of the research.
Tell it what counts as a find
If I just ask for "anything interesting," what comes back is rarely enough, because the AI decides for itself what phrases like “interesting” or “surprising” mean. So it helps when you define what you are looking for. I often want findings that are new, that overturn a belief, or that I can act on. Background is fine to include, but it can't be all that comes back. And the best findings tell me what I'd do differently knowing them.
Approach it from different angles
Instead of asking the AI to research the topic all at once, I split it into a handful of narrower questions, each chasing a different angle: what's changed in the last year, how the people who do this for a living do it, where the experts disagree, the real numbers, who makes the money. I've included more examples in the prompt below. They're suggestions, not a fixed list, so use the ones that fit.
If your plan can run AI agents, you can do all of this in parallel. The agents run every question at the same time and pass the busywork (opening pages, copying quotes, checking links) to cheaper, faster models, so you're not paying top rates for it.
If you're in a regular chat window instead, the by-hand version works just as well. Open a new chat for each question and go through them one by one, using a fresh chat each time so each answer doesn't build on the last. It's slower, but you land in the same place.
Make it grade and cite everything
Every source gets a grade: A for primary (a filing, official data, the source itself), B for established press, C for a blog or an estimate with no method shown. That way I can tell which facts are solid enough to say and which I should double-check.
C isn't banned, it's labeled. A weak source can still be a useful lead, so I keep it. The label just means I won’t let a claim rest on it without tracing it back to something firmer first.
Every fact is accompanied by a link, a date, and the grade. Estimates get flagged as estimates, and when two solid sources disagree it gives me both instead of picking. Some numbers, chased back far enough, turn out to rest on nothing anyone measured. That's how they get caught, before I repeat them.
Ask what it couldn't find
At the end, I ask it for a list of what it looked for and couldn't verify. An AI that says "no reliable source exists for this number" just saved me from a number I would have repeated.
Once everything is back, I go through it and decide what to use. This final judgment call stays with me.
The full prompt
All of these tactics as one prompt. Fill in the two brackets, and if your plan can’t run AI agents, run one angle at a time.
First, check today's date. Don't rely on your training data for anything recent. Search the web and read each source before you cite it.
I'm researching [SUBJECT]. These areas matter most, but don't stop at them: [YOUR AREAS].
Background information is fine to include, but you must go beyond that. I want findings that are new, that overturn a common belief, or that I can act on. For each one, note what it would change about how I'd approach the subject.
Work one angle at a time, whichever apply:
- What has changed in the past year, and where the debate stands now
- How experienced practitioners handle it, step by step
- Where credible experts disagree, and the strongest case on each side
- How things got this way, including the approaches that failed
- How the same underlying problem is solved in another field
- The most-cited figures, traced back to their original source
- First-hand accounts, in people's own words
- Widely held assumptions that don't hold up
- Where the money comes from and where it goes
Hand the mechanical work (fetching pages, pulling quotes, checking links) to a cheaper, faster model, and keep the strong model for the reading and the judgment, so the run doesn't cost more than it needs to.
Grade each source: A for primary (filings, official data, the source itself), B for established press, C for blogs or unsourced estimates.
Give every fact an inline link, a date, and a grade. Label estimates as estimates. Where two solid sources conflict, present both.
Close with a list of anything you could not verify.The 30-second version
If you give your AI a generic research request, you get generic and shallow results.
Don't let AI answer from memory. Make it check the date and read the actual pages it cites.
Say what you're after, give it a few areas as a floor (but not a ceiling), and keep your own answer to yourself until the end.
Run several narrower angles instead of one big prompt, and use the ones that fit.
Make it grade every source, put a link and date on every fact, and list what it couldn't verify.
None of this works because the AI is smart. It works because each tactic assumes the AI is about to be confidently wrong, and catches it before I do. Getting a pile of research has stopped being the hard part. What’s important now is being able to stand behind it when someone asks where it came from.
me+machine