# AI GTM Is Just Fast Experimentation

Inside the new GTM playbook where pricing, distribution and even the customer can keep changing.

By Paras Madan | 2026-09-17


A few years ago, GTM was something you could plan around a relatively stable product. You found an ICP, worked on the positioning, picked a few channels, built a sales or product-led motion around them and then spent the next few years making that machine better.

AI has made that much harder. Products are changing every few weeks, new model capabilities can suddenly make old features irrelevant, pricing keeps moving and even customers are still figuring out what they actually want from many AI products.

So I think AI GTM is increasingly becoming much simpler: how quickly can you test an assumption, get a real signal and change what you are doing?

This does not mean companies should stop having a strategy. You still need to know roughly who you are building for, what problem matters and why someone should care, but almost everything after that should probably be treated as something that can change.

## Pricing can change as you learn

Lovable is probably one of the clearest examples. In her first year leading growth there, Elena Verna and the team [changed pricing around 10 times](https://stripe.com/sessions/2026/how-lovable-turned-pricing), testing annual plans, credit rollovers, top-ups and removing per-seat pricing as they learned how people were actually using the product.

That is unusual when you compare it with how SaaS companies traditionally treated pricing. Pricing used to be one of those big decisions you revisited after a year or two, while Lovable was treating it almost like a product feature that could continuously improve.

And that I think is much closer to how AI GTM works today!

## More activity is not more learning

There is an important difference between experimenting fast and simply doing more things. Launching five campaigns every week is not experimentation if you do not know what each one is trying to prove.

A useful GTM experiment starts with an assumption. Maybe developers are not converting because the free plan runs out too quickly, enterprise buyers like the product but do not understand how to deploy it, or users in one country love the product but the pricing does not make sense locally.

You change one thing and watch what happens. If the result is bad, that is still useful because you learned something before spending another six months building around the wrong assumption.

The advantage is not being right every time. It is making being wrong cheap.

This becomes particularly important in AI because the market can move faster than the company's planning cycle. A feature that drove signups six months ago might now be available everywhere, while a use case nobody cared about can suddenly become possible because the models got better.

## Let the signal lead the experiment

Cursor gives a good example of this at a completely different level. Its user base in India tripled in a year to more than 3 million developers, and Cursor also found that India had more power users than any other market, measured by agent requests per developer.

Instead of simply continuing with its existing global plans, Cursor launched Cursor Start, an India-only ₹649 monthly plan with local pricing and UPI payments. [Cursor’s India launch](https://cursor.com/blog/cursor-start-india)

The interesting part is what they said after launching it. They are starting with India, learning from how developers respond and then deciding whether the model should move into other markets.

The signal came first: people in India were already using the product heavily. Then came the hypothesis that more of them might pay if the price and payment method fit the local market, and only after that came the experiment.

This is probably how more AI companies will have to operate. ICP, pricing, packaging, onboarding and even sales motions cannot always be treated as permanent decisions.

## Distribution can come from the team itself

One thing I find interesting about Lovable is that the experimentation does not stop at pricing. Look at how their team uses LinkedIn.

A recent analysis of their employees found around 45 people at Lovable with more than 10,000 followers, with the top 15 adding up to roughly 750,000 followers. The founder posts constantly, their growth leader already had a large audience before joining, and recruiters, community people and engineers all post regularly.

Even joining Lovable becomes content. For several employees, their announcement that they had joined the company became one of their highest-performing posts.

When Lovable has a major announcement, people also do not simply reshare the corporate account. Different employees tell the same event from their own perspective: the founder story, the hiring story, the product story or the personal story.

Now, I don't think the takeaway is that every AI company should tell its employees to become creators. The more interesting point is that Lovable is treating its own team as another distribution surface that can be built and experimented with.

Hiring someone with an audience can affect distribution, an employee talking about how they work can affect hiring, and a founder explaining something publicly can affect product discovery. These things previously sat in different parts of the company, but now they can all become part of the GTM system.

## GTM is part of the product loop

This is where the definition of GTM starts becoming much wider. A pricing change can be a GTM experiment, a free tier can be one, an onboarding change can be one, giving product credits at a hackathon can be one and launching a cheaper plan in India can be one.

Even hiring someone who already understands and speaks to your customer can eventually become part of distribution. This does not mean everything a company does should suddenly be called marketing, but it does mean the things that affect growth are increasingly spread across product, pricing, community, hiring and sales.

Earlier, the sequence was often simple: build the product, hand it to marketing and then distribute it.

AI companies are increasingly running those things at the same time. You build something, put it in front of users, watch what happens and then use that information to decide what gets built next.

GTM becomes part of the product loop itself.

## Constant experimentation can become chaos

There is one obvious problem with all of this. If you tell a team to experiment constantly, it is very easy to end up with chaos.

Pricing changes every month, messaging keeps changing, three channels are launched and abandoned, and after a while nobody remembers why something was tested in the first place. That is not fast GTM.

The experiment still needs a hypothesis, a signal you care about and enough time or data to know whether anything actually changed. Even Lovable's pricing experiments are not simply random changes; the team has talked about running tests with real customers, looking at conversion and usage behaviour and building the infrastructure that lets them change monetisation without rebuilding everything each time.

So the goal is not maximum experimentation. The goal is maximum learning.

## The playbook has to keep changing

AI products are becoming easier to build, which also means individual features can become easier to copy. If two companies can eventually build similar functionality, the advantage may increasingly come from how quickly each one understands what should happen around the product.

Who is actually getting value from this? What should be free? What will people pay for? Which users should talk to sales? Which market behaves differently? What makes someone share the product and what makes them come back?

There probably will not be one permanent answer to most of these questions. And that is why I don't think AI GTM is really about finding one perfect playbook.

The playbook itself has to become something the company can keep changing.

You can start with the wrong pricing, pick the wrong channel or even begin with a slightly wrong customer. The dangerous part is not being wrong; it is taking six months to find out.