Deep Thought Said 42. Your AI Deep Research Might Too.

Seven and a half million years.

That's how long Deep Thought, the most powerful computer ever built, spent working on the Ultimate Question of Life, the Universe, and Everything. (If you've read The Hitchhiker's Guide to the Galaxy, you know where this is going.)

On the big day, the crowds gathered. Deep Thought delivered its answer “with infinite majesty and calm.”

Forty-two.

Uproar. Seven and a half million years for a two-digit number? Deep Thought didn't flinch. The answer was correct, it said. The trouble was that nobody had ever worked out what the question was.

They knew the subject: Life, the Universe, and Everything. They'd just never pinned down what they were actually asking.

I've been thinking about Deep Thought since digging deep (did I just say that?) with AI deep research.

Deep research tools are remarkable. Give ChatGPT, Gemini, Claude, Copilot, or Perplexity a research question, and within the hour you'll have a report built from dozens (often hundreds) of sources that looks like it came from a research firm.

But they share one habit with Deep Thought. They will always give you an answer.

Ask a vague question like “How big is the quality control market?” and you'll get a large, impressive number that tells you nothing about who will buy from you.

Ask a sharper one, like “How many mid-sized Ontario manufacturers still run quality control on spreadsheets?”, and you hit a different wall. Nobody has published that figure. So the tool builds one from whatever it can find: a Statistics Canada count here, a vendor survey there, a reasonable-sounding assumption to glue them together. Each piece might be useful. The total is a guess, delivered with the same confidence as everything else in the report.

Forty-two was precise, too.

Why it's so tempting

If you run a B2B company with no senior marketer on the team, you know the moment. The board meets in two weeks and someone will ask how big the opportunity is. Or a competitor you'd never heard of turned up in a deal last month.

There's nobody down the hall to ask. In my experience, a commissioned market study starts around US$10,000, often runs to US$50,000, and can go much, much higher. It also takes weeks. Off-the-shelf reports are cheaper (Frost & Sullivan's list at roughly US$2,500 to US$7,000), but they answer the question the publisher chose to ask, which may not be yours.

So you open a deep research AI tool, type the question, make a cuppa, and come back to a 30-page report with footnotes.

That's a sensible instinct. I use these tools all the time too, though I spend far longer on the question than on the cuppa. Used with care (and I can't emphasize that enough), these tools are one of the most useful things AI can do for a company your size. Used carelessly, they hand you 42.

Three ways it hands you 42

1. It answers the question you typed, assumptions included.

Say you give it the full picture: who you are, what you sell, who you sell to, and where. Then ask for your top competitors. The tool will find companies that describe themselves the way you do online, and it'll make a tidy job of it. What it won't tell you is that your buyer's real alternative might be a spreadsheet that someone built years ago, a co-op student, manual processes, or doing nothing for another year. Those rarely show up in a competitor list, because the question assumed a competitor is another vendor. And in plenty of B2B deals, doing nothing is the toughest competitor in the room.

2. It won't look where you didn't point it.

Deep research is thorough inside the boundaries you set. Ask about manufacturers and distributors, and you'll get a solid picture of manufacturers and distributors. It might name a neighbouring segment in passing, buried somewhere in the middle of the report. It won't stop to investigate it, and if the segment with the most pain is one you've never considered, there's a good chance it never comes up. Leave the boundaries loose and you get the opposite problem: pages of tangents that have nothing to do with your decision. Either way, where it looks is down to you.

3. Your bias comes along for the ride.

Try this one: “What frustrates finance leaders most about their forecasting tools?” Reasonable question. It also assumes finance leaders use forecasting tools, and that they're frustrated with them. The research will find sources that fit, because there are always sources that fit (and now and then, it invents them). Then it hands your assumption back to you with citations attached, and it looks a lot like a finding.

I learned this one firsthand. I ran the same research brief through five different tools, and five polished, research-grade reports came back. All five agreed with me. That should have been the warning sign. Buried in my prompt was an assumption I hadn't framed neutrally, and every tool built its research around it.

Garbage in, garbage out. It just comes out beautifully formatted.

That's the one that worries me most. A hunch that you know is a hunch keeps you careful. A hunch with twenty footnotes feels like fact.

And when you act on it, 42 gets expensive. You tell the board which competitor to worry about, then spend two quarters responding to a company your buyers never seriously considered. Or you price against the vendors on the competitor list while deals keep slipping to “we'll stick with the spreadsheet.”

What it's good for (and it is good)

None of this means you should avoid it. It does a few things very well:

• A fast first map of a market: who's in it, how they describe themselves, and what's changed in the last year or two.

• Data points that test your assumptions. If you believe a segment is growing and three credible sources say it's shrinking, you want to know before the board meeting.

• Things you didn't know to ask, if you ask it directly what you might be missing.

Look at what's on that list. Inputs. Raw material for a sharper hypothesis. The report won't label them that way. It presents everything with the confidence of a conclusion.

How to ask Deep Thought a better question

Most of the work happens before you hit send.

Start with the decision you're making. “Research the quality control software market” gives the tool a topic and nowhere to aim. Try this instead: “We're deciding whether to focus our sales team on Ontario manufacturers with 50–200 employees next year. What would have to be true for that to be the right call, and what evidence is there either way?” Now it knows what the answer is for, and it has a reason to look on both sides.

Write down what you believe going in. Who you think you're for, what they struggle with, and who you think you're up against. Then ask the research to test that hypothesis, including a section that argues against it. Just keep that hypothesis out of the question itself. Ask “What's the evidence on X?” instead of “Why does X happen?” If it can't find anything that disagrees with you, be suspicious.

Brief it like the researcher it is. Trust Insights teaches a framework called CASINO: Context, Audience, Scope, Intent, Narrator, and Outcome. Tell it what you're working on, who'll read the report, the boundaries (timeframe, geography, sources), why you need it, what role it should play, and what the finished report must include. Gathering the context can take an hour or more, depending on the question (time well spent: the more it knows, the better the report).

Make it show its working. Ask for a section on where the evidence is thin, where sources disagree, and which numbers are estimates built from other estimates. Ask for a citation next to every claim. Then click through on the ones you're leaning on and check that the source exists and says what the report claims it says. More often than you’d think, it doesn’t. The link usually works. It’s the report’s account of the source that goes wrong.

Bring your own judgment. For a big decision, run the same brief through multiple tools, because they'll probably use different sources. They'll disagree in places. Merging the reports and deciding what to trust takes time and experience, and that part's on you.

Turn the findings into interview questions. Don't skip this one. Every surprising finding, every number you're relying on, and every assumption the report confirmed becomes a question you can ask a customer.

If that sounds like a lot to add to a week that's already full, it is. It's also work I take on for clients: I sit down with you to gather the context, write the brief, run it through several tools, merge and fact-check the reports, and come back with a sharper hypothesis and the questions to test it with customers. What you get is a much better starting point for your customer conversations, which is where the answer lives.

The bigger computer

So what did Deep Thought recommend?

A bigger computer. One powerful enough to work out the question. It would be so vast that living creatures would form part of its operating system. They called it Earth.

The good news is, your bigger computer is much smaller, and you already have access to it: your customers.

Use the research to build a sharper hypothesis about who your best-fit customers are, what they're struggling with, and what makes you different. Turn that hypothesis into questions. Then ask those questions, out loud, to ten people who could buy from you. They'll tell you what no report can: what finally pushed them to look for help, whether there's money to fix it this year, and the words they use to describe the problem.

The research gets you to a better question. Your customers give you the answer.

So before you run your next deep research report, write down the decision it's meant to inform. Then book ten customer conversations before you act on it.

Forty-two is a perfectly good answer. Just make sure you know the question.