You have ChatGPT open right now. Here is why that might not be enough.

Should I just use ChatGPT for market research? You probably have it open in another tab right now. You could paste in your product, your price, your market, and last month's traffic, and have a thoughtful answer in about ten seconds, for nothing. So what could a paid tool possibly add.

Nothing, sometimes. For a lot of the thinking a founder needs to do, the raw model is genuinely enough, and if that is what you need then you should not pay us.

We are also not going to argue that the general models are weak. They are extraordinary, and MarkLens is built on one of them. Anthropic's models do the reasoning underneath every read we produce. We are not a competitor to the thing in your other tab. We are a particular way of using it. The model is the kitchen. It is not the meal. If that raises the obvious follow-up, which is what stops anyone else renting the same kitchen, we answered that one separately and just as plainly.

So the difference between asking the model yourself and running a read is not intelligence. The model is not missing any. The difference is what a general model does by default when you ask it whether your idea is good, and what it will not do unless you know how to make it. There are three of those defaults. None of them are flaws.

First, what the raw model is genuinely better at

This part matters more than the contrast that follows it, because a contrast built on a fake concession is worth nothing.

It is far more flexible. It will reason about anything, in any direction, and follow you wherever you go. MarkLens does one job. The model does a thousand, and it does most of them well.

It is free, or close to it, and instant. We are neither. That is a real advantage and not a small one, particularly early, when the question is still forming and you want to ask it forty times.

It talks back. You can push, disagree, reframe, and chase a thread sideways for an hour. A fixed output cannot do that. When what you need is something to think against rather than an answer to act on, the conversation is the whole value, and a verdict would just get in the way.

We mapped this tooling category into five shelves once, by the job each kind of tool does. The general model is arguably an informal sixth: it does a bit of every job, on demand, for free. That is a genuinely strong position, and it is why the objection in the title is a fair one rather than a lazy one.

So if the model is that good, and we run on it, what is actually different.

Default one: it is built to encourage you

Modern models are shaped by human feedback. People rate responses, and the model is trained toward the ones people rate well. Over millions of those judgments, helpful gets shaped into something close to agreeable: supportive, constructive, willing to find the angle where your plan works.

That is not a defect. It is what the training selects for, and most of the time it is exactly what people want. Nobody wants a search engine that opens by telling them their question is bad.

But it means that when you ask a general model whether your idea is good, the default gravity is warmth. It will locate the promising reading. It will lead with what could work. The hard version of the answer is usually available, but you have to go and get it, and you have to already suspect it exists.

You can watch it happen. Hand a general model a product with weak sales and the reply will usually open with what is working, move to what could be adjusted, and place the possibility that the thing simply is not wanted somewhere near the bottom, phrased kindly. Every item on that list may be true. The ordering is still doing work on you, and the ordering came from the training rather than from your numbers.

For the question that actually matters after launch, which is why nobody is buying, that is the wrong default. You do not need a model rooting for you. You need one that will say the thing you have been working around for six weeks. A tool that agrees with you is pleasant and expensive, because the bill arrives later, as spend against a diagnosis that was never true.

MarkLens is tuned against that default on purpose. The verdict is built to withhold the encouraging read, and the confidence attached to it is calibrated to the evidence rather than to your hopes. When the evidence is thin, the read says so and scores lower, which is not a weaker answer. It is a more honest one.

Default two: it reasons from training, not from your market this month

Ask a general model about your category and you usually get an answer assembled from what it already knows. That answer is often good. It is also general, and it is describing the category as it appeared in the material the model learned from, not the field as it stands the week you are asking.

Your competitors changed their pricing in March. Two of them stopped shipping. A marketplace changed how it ranks listings in your category. None of that is in a general answer unless something goes and looks.

MarkLens looks every time. Every read runs live searches against the current field, four on a diagnosis and five on a forecast, and the verdict is built only from what those searches actually returned. Then the confidence is calibrated to that evidence: a handful of weak sources produces a low-confidence read and a capped score, and the read tells you which of the two it is standing on.

The gap between a fluent answer and a grounded one is invisible right up until it costs you, because they read identically. A wrong answer about your market arrives in the same calm, well-organised prose as a right one, with the same air of having considered the matter. Grounding is the only thing that separates them, and grounding is work a general model will not do unless you drive it there yourself.

Default three: it follows your framing

Ask the same question three ways. Ask it hopefully. Ask it worried. Ask it flat. You will tend to get three different answers, because the model is responding to how you asked as much as to what you asked. Lead it, and it follows. That is a feature nearly everywhere else, and it is what makes it such a good thinking partner.

Which means the answer you get from a raw chat is partly a picture of your own framing handed back to you. A founder quietly hoping for a yes will phrase their way toward one without ever deciding to. Not by lying. By choosing which facts to include, which worry to leave out, and which word to use for the thing that is not working.

Try it deliberately some time. In one conversation, ask whether you are right that there is real demand here. In a fresh one, ask whether you are wrong to worry that there is none. Supply identical facts both times. The two answers are rarely identical, and neither of them is dishonest. Each is a reasonable reply to a different question, and you asked a different question twice without deciding to.

MarkLens holds the same evidentiary bar no matter how the question arrives, and returns a verdict you did not get to shape. That is less pleasant, and it is the point. An instrument is only useful because it does not bend to the hand holding it.

Notice what none of those three defaults are. None of them is the model being bad at its job. Each one is the model being very good at a different job from the one you handed it: being useful across everything, rather than being right about the one specific thing you cannot afford to be wrong about.

When you should skip us and use the model

Three cases, and we mean all three.

Your question is not our job. MarkLens answers whether something will sell and why it is not selling. If you want to draft positioning, argue with a strategy, write the launch email, or think out loud for an hour, the model is better at that than we are and it is not close.

You have the discipline to do this yourself. This is the honest one. If you will supply the evidence rather than letting the model infer it, prompt deliberately against your own optimism, force it to search the live field instead of answering from memory, and hold it to a verdict it did not want to give, you can get a real read out of a raw model without us. That is not a trick we know and you do not. Most founders will not do it, and the reason is not capability. It is that running that process honestly on your own product, in the week the numbers are bad, is the hard part. We sell that discipline pre-built. We are not selling access to something you cannot otherwise reach.

The model is free and we are not. If budget is the binding constraint and you have the discipline above, the honest recommendation is the free tool. Spend the money on the fix instead.

The honest test

So here it is, and you can apply it in about a minute. If you will supply the evidence yourself, push the model off its instinct to encourage you, make it check the market as it stands today, and hold your own feet to the fire when the answer is one you did not want, then use the model. Genuinely. It is excellent, it is free, and we are built on it.

If you would rather that discipline sat inside the instrument, so that you cannot quietly steer your way to the answer you were hoping for, that is what MarkLens is. The same reasoning underneath. A different set of defaults on top.