What AI Changed in Growth Marketing — and What It Didn't

Long enough to have an answer

I first wrote about using a language model in growth work in spring 2023, six weeks into experimenting. I wrote a follow-up a year later mapping where it had earned a permanent place.

Three years is long enough for the answer to settle. The genuine changes are narrower than the 2023 discourse promised and deeper than the sceptics allowed — and, more usefully, they're now specific enough to state plainly.

---

What genuinely changed

1. Reading at scale became free.

This remains the largest real shift and it's still the least discussed, because it doesn't produce a demo.

Every company sits on text nobody has read: years of sales call transcripts, support tickets, closed-lost reasons, review sites, community threads. Previously this was permanently inaccessible — not expensive, impossible at the budget of a Series A company.

Now it's an afternoon. In practice this has changed how I do customer research more than anything else in a decade. I no longer build positioning from six interviews and hope they're representative; I build it from every recorded customer conversation the company has, and then use interviews to test what I found.

2. The cost of producing competent output went to zero.

True, and its consequences were mostly not what people predicted.

Competent content stopped being a differentiator, because everyone has it. Search results filled with adequate, unsourced, indistinguishable articles. The result is that the scarce thing shifted decisively toward what a model cannot produce: original data, named first-hand experience, a specific argument someone is willing to be wrong about in public.

Companies that responded by increasing volume are worse off than they were in 2023. Companies that responded by publishing less and making it more specific are meaningfully better off.

3. Operational overhead compressed.

Reformatting, drafting, restructuring, extracting, querying. Individually trivial and collectively a real change in how much a small team can carry. This is where the productivity gains actually landed — not in strategy, in logistics.

---

What didn't change at all

Deciding what to do. Still the bottleneck, still unassisted in any meaningful way. A model will give you a competent summary of every option and cannot tell you which four of five things to stop doing, because that judgement requires knowing your constraints, your politics, and what you're willing to bet.

Getting an organisation to act. Every failed growth engagement I've seen in three years failed for organisational reasons: no owner, no agreement between sales and marketing, no budget authority, competing priorities. None of that is a text problem.

Knowing whether something worked. Attribution, measurement, and causal inference are exactly as hard as in 2021. Arguably harder, given how much low-quality content now sits between a company and its buyers.

Trust. Buyers still buy from people and companies they believe. The cost of seeming credible collapsed, which — predictably — made actual credibility more valuable, not less.

---

Where the value moved

The clearest pattern across three years: value moved from producing things to knowing which things are worth producing, and from analysing data to having proprietary data to analyse.

For a growth marketer, that means the skills that appreciated are:

• Framing the question well enough that the answer is usable
• Having original evidence nobody else has — your own experiments, your own customer research, your own outcomes
• Judgement about what to stop doing
• Getting other functions to change what they do

And the skills that depreciated are the ones I'd have listed as valuable in 2020: producing volume, formatting, summarising, first-drafting, competent execution of a known pattern.

That's uncomfortable if the second list is what your role is built on. It's also been visible for three years now, which is enough warning to have acted on.

---

The rule that survived

I proposed a heuristic in 2023: it helps when I already know the answer and need it expressed, varied, or stress-tested; it costs me time when I don't know the answer and hope it does.

Three years and several model generations later, I'd change one word. The models are far better, the range of things I'll hand over has widened considerably, and the boundary has moved — but it hasn't disappeared, and it's in the same place.

Judgement still doesn't transfer. Everything downstream of judgement mostly does.