Marco Kohnsgrowth14 min read
Generative AI for Growth Hacking: What Our 2023 Interviews Got Right, and What They Missed
I co-wrote one of the first peer-reviewed studies on LLMs in startup growth. Three years on: what held, what broke, and what the paper could not see coming.
In February and March 2023, in the first weeks after ChatGPT appeared, my co-authors and I interviewed founders, growth leads and sales directors at 20 companies, most of them in the US, Germany and Portugal, and asked one question: how are you using generative AI to grow? The answers became a peer-reviewed study in the Journal of Business Research, published in March 2025 with Arash Rezazadeh, Prof. René Bohnsack, Nuno António and Paulo Rita. It is open access, so you can check everything below against the original.
Generative AI for growth hacking means using large language models to run more growth experiments, faster, with the same small team: drafting the copy, the code, the outreach and the analysis that a startup would otherwise buy with headcount. That definition is the paper's and it has held. Most of the specific predictions have not, and the interesting part is which ones.
This is the author's own audit, three and a half years after the fieldwork. Where the interviewees were right, I say so. Where we missed something that now looks obvious, I say that too, because a study of a technology in its first quarter is a photograph, and photographs age.
What did the study find?
Five things, and I will quote the paper rather than my memory of it. We interviewed 20 people in senior roles at companies from pre-seed to Series E, coded the transcripts with the Gioia method, and arrived at: the drivers and barriers of using generative AI for growth; the top use cases for three growth strategies (product-led, sales-led and operational-efficiency-driven); two frameworks, the AI Wheel and the AI Capabilities Framework; three families of risk (organisational, competitive, societal); and a shortlist of LLM use cases mapped to the growth challenges startups named.
The drivers were resource constraints, agility, resilience, and deep customer understanding in a specific niche. The barriers were data privacy, over-reliance on model providers such as OpenAI and Hugging Face, and competitive pressure to be first with an AI-based offering.
The research question, verbatim: "How can startups use generative AI in their growth hacking strategies for effective competition and development?"
| Category | |
|---|---|
| Small, under 100 staff | 12 |
| Medium, 100 to 500 | 4 |
| Large, over 500 | 4 |
| Germany | 6 |
| United States | 5 |
| Portugal | 4 |
| Other (IL, DK, FR, CH, IN) | 5 |
Two readings of the same 20 companies. Most were small, and the sample leaned European, which shaped what we heard about regulation and data. Source: Rezazadeh, Kohns, Bohnsack, António and Rita, Journal of Business Research 192 (2025), Table 1, n = 20 interviews
The sample matters for everything that follows. Twelve of the 20 companies had fewer than 100 employees. Six sat in Germany, five in the US, four in Portugal. The interviews ran 23 to 34 minutes each, face to face and in English, and we stopped at theoretical saturation, which is the qualitative researcher's way of saying the answers started repeating.
Where did the study come from?
From an API key and a lot of blank looks. I met GPT-3 at Techstars Berlin in 2022, when it was a text-completion endpoint with no chat window and no consumer product around it. It was a niche, nerdy topic, and when I brought it up with marketers the reaction was close to "what is this, AI?" I decided to write my master's thesis at NOVA IMS on what it could do for growth and customer acquisition, the problem I had watched every early-stage company struggle with since my first agency at 19.
Then ChatGPT shipped, and the topic went from niche to unavoidable inside a single quarter. The fieldwork happened in that quarter. Some interviewees were already talking about GPT-4 alongside GPT-3, which tells you how fast the ground moved under the study while we were standing on it.
I had started running university workshops on generative AI in 2022 on the back of the thesis, and those grew into 10+ executive-education seminars at CATÓLICA-LISBON and online courses that have reached 1,500+ students. Prof. Bohnsack, who supervised the work, remains a mentor and a collaborator. The paper is the formal version of a conversation we have been having for four years.
Which use cases did the interviewees name in 2023?
Nine, grouped by growth strategy, and every one of them was about producing something faster. For product-led growth: a coding assistant (GitHub Copilot was the example), code verification (OpenAI Codex), and non-technical text such as user stories, acceptance criteria and product documentation. For sales-led growth: promotional content with SEO elements (AdCreative.ai), repurposing one piece across channels (Repurpose.io), and personalised outreach and support chatbots trained on the company's own data. For operational efficiency: AI-assisted ideation for market entry, market research for geographic expansion, and event management assistants.
| Growth strategy | Top use cases the 20 interviewees named, 2023 | Status in 2026, from my own operation |
|---|---|---|
| Product-led | Coding assistant, code verification, generated docs and user stories | Superseded. The assistant became an agent that builds and deploys the whole product |
| Sales-led | Promotional content, channel repurposing, personalised outreach and chatbots | Held, then commoditised. Everyone has it, so it stopped being an edge |
| Operational efficiency | Ideation, market research, event assistants | Held. The least glamorous row aged best |
Two quotes from the paper deserve reprinting because they were more right than the frameworks around them. One founder said it was "now possible to produce an MVP within a week using generative AI". Another, running content for a sales-led company, said: "5% of our articles bring 95% of the traffic, so the most important work is actually automating the process of finding that specific piece of content and making it very good." Read that second one again. In 2023 it was an observation about editorial focus. In 2026 it is the entire strategy, for reasons the next section covers.
What did the study get right?
Three things, and the barriers list is the one I am proudest of.
Over-reliance on model providers. One interviewee told us: "having a product that is dependent on someone else's language model is problematic." In 2023 that read as caution. Since then every operator I know has lived through a deprecation notice, a pricing change or a behaviour shift in a model they built on, and the companies that treated the provider as a swappable dependency rather than a foundation are the ones that slept. The paper filed this under barriers. It should have been the headline.
Human judgement as the scarce input. "LLMs only boost productivity if the user can ask the right questions," said one interviewee, and the AI Capabilities Framework we built around that has three steps: build knowledge of what the tools do, refine it by combining tools, then develop what we called reflective knowledge, the ability to judge an output for accuracy, copyright and hallucination before it goes anywhere. Every content operation I run today is that third step turned into a script. More on that below.
Agents. The paper says autonomous agents "are expected to become more mature and integrated into business operations", and points at AutoGPT and prompt chains as the early form. We were right about the direction and wrong about the timescale by a factor I would rather not compute. The interviewees were planning to integrate these things "within the next five to ten years". It took about two.
What did the study miss?
The cost of production going to zero, and what that does to the value of production.
Every use case in Table 2 of the paper is a way of making more: more code, more articles, more emails, more ideas. The implicit assumption, mine included, was that output was the constraint and generative AI relaxed it. That was true for about eighteen months. Then everyone relaxed it at once.
Stanford's 2025 AI Index puts organisational AI use at 78% in 2024, up from 55% the year before, and reports that the inference cost for a system performing at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024. When a capability is that cheap and that widespread, it stops being an advantage and becomes a floor. The startup that could generate 100 personalised emails from one or two demos, which one of our interviewees described as a breakthrough, now competes with every other startup doing the same thing to the same inbox.
280x
fall in inference cost for GPT-3.5-level performance, November 2022 to October 2024
Stanford HAI, 2025 AI Index Report
Google noticed before most operators did. On 5 March 2024 it introduced a scaled content abuse policy covering pages "generated for the primary purpose of manipulating Search rankings and not helping users", and stated that it applies "no matter whether content is produced through automation, human efforts, or some combination". The old policy asked how content was made. The new one asks why, and how much. Google expected the changes to cut low-quality, unoriginal content in results by 40% and reported 45% once the rollout completed on 19 April 2024. I have written separately about what a manual action under these policies looks like from the inside, because I am working through one for a client, and the ordering of the recovery is the part people get wrong.
So the paper's most cited use case, content at scale for organic acquisition, became a policy violation if done the way the 2023 tooling invited. The interviewee who said the job was finding the 5% and making it very good was describing the only version of the strategy that survived.
How has practice changed in my own operation?
The unit of work moved from the paragraph to the product, and the human moved from writer to gatekeeper.
In 2023 the product-led use case was an assistant suggesting the next line of code. In 2026 I build and ship whole sites with an agentic coding tool: the blog you are reading, the quiz funnel at MyPassion.AI, the ERP selection tool at ERP Pilot. The assistant did not get better at completing lines. It became something that reads a repository, plans a change, runs the build, checks the result and opens the commit. The founder who predicted an MVP in a week undershot.
Content changed in the opposite direction. Producing an article is trivial, so almost all of the effort now sits in what happens before and after the draft. Before: does anyone search for this, can we say something no competitor can, does an existing page already own the term. After: a mechanical gate. This site's own gate script rejects a post that contains an em dash, any of a list of phrases that read as machine-written, a missing or duplicated conversion block, or fewer than two data-bearing visuals. It exists because a read-through misses things a regular expression does not, and because the 2023 interviewee who warned that "if you rely too much on AI without enough human input, the content you produce will no longer be interesting" was describing a failure mode I have since watched happen at scale.
Here is what that looks like in numbers I can read from my own systems today. The MyPassion.AI blog holds 84 published posts, 17 of them published in September 2026 alone, each one through a 367-line gate before it can be committed. Google Search Console for mypassion.ai shows 12,772 clicks from 1,022,526 impressions in the 90 days to 27 September 2026, at an average position of 7.2. Those figures are the site's own, read this morning, and I share them for one reason: the pipeline that produced them is AI-assisted end to end and every post still goes through a human who can say no. That is the reflective-knowledge step from the paper, made operational.
| Category | Percent |
|---|---|
| Coding tasks, Claude.ai | 36% |
| Coding tasks, API | 44% |
| Automation-dominant, Claude.ai | 50% |
| Automation-dominant, API | 77% |
Businesses delegate rather than co-write. Three quarters of API use is automation-dominant, against about half of consumer chat. Source: Anthropic Economic Index, September 2025 report, August 2025 samples of 1M API transcripts and 1M Claude.ai conversations; coding shares rounded from the report's 'little less than half' and '8 points higher'
The pattern in that chart is the pattern in my own work. The Anthropic Economic Index for September 2025 found that a little under half of API traffic is computer and mathematical work, about 8 points above the consumer product, and that 77% of business uses follow automation patterns against about 50% on Claude.ai. Businesses hand the model a task and check the result. The 2023 interviewees imagined a co-pilot. What arrived is closer to a contractor with a supervisor.
Does the AI Wheel still turn?
Partly. The framework mapped generative AI onto activities by funding stage: product content and go-to-market knowledge for product-led growth, automated customer-facing work and personalisation for sales-led growth, productivity and lighter CRM workloads for operational efficiency. The mapping is still a reasonable way to audit where a team uses AI.
What it lacks is a ring for distribution. The wheel assumes that making the thing is the hard part and that the market will find it. In 2026 the constraint is the opposite: attention, trust and the willingness of a search engine or a feed to show your page over ten near-identical ones. If I redrew it, the outer ring would be the channels the company can credibly win, and the AI use cases would sit inside it, chosen by whether they help win that channel rather than by what they can produce. I have set out that argument in full in whether SEO is still worth a solo founder's time, using the search data from four of my own sites.
The AI Capabilities Framework aged better than the Wheel, precisely because it was about people rather than tools. Build, refine, reflect: the third step is now the whole job.
Which risks from the paper have materialised?
The competitive one, faster than the organisational one, and the societal one in a form we did not describe.
We wrote about competitive pressure forcing companies to ship AI features before they understood them. That happened, and the counter-move we observed in a few companies, prioritising explainability over first-mover advantage, turned out to be the mature position. On the organisational side, the data-privacy barrier held up and remains the reason serious companies keep customer data out of consumer tools. One interviewee refused to share a code base or customer data with "any of these open-source tools and language models that sit somewhere in the cloud", and enterprise procurement in 2026 still runs on that instinct.
The societal risk we described was job displacement and skills obsolescence, with a memorable quote from an employee: "this is not going to take my job. I am much better than your tools." What we did not describe is the flooding of every shared channel with plausible text, which is a societal cost borne by readers rather than workers. One interviewee did see it: "this could lead to a significant increase in spam emails that look very human." Two years later Google rewrote its spam policies around exactly that. The employee upskilling argument I made elsewhere rests on the same observation: the person who can judge an output is worth more than the tool that produced it.
What would I ask the same 20 people today?
Not what they use the model for. What they refuse to let it do.
In 2023 the interesting variable was adoption. Everyone was finding uses, and cataloguing them was worth a paper. In 2026 adoption is a given and the interesting variable is restraint: which decisions a company keeps human, which outputs it gates, which channels it declines to flood because it wants to still be readable there in three years. That is a harder study to run, because the answers are less flattering and the interviewees know it.
I would also ask a question the paper did not: what is your dependency plan for the model you built on. The interviewee who called provider dependence "problematic" in early 2023 was describing the one risk that has hit every operator I know since, and the study filed it as a barrier to adoption rather than a design constraint for the product. If I could change one sentence in the paper, it would be that one.
The paper is open access, the full citation sits on my publications page, and the seminars that grew out of it are what I bring to a room when someone books me to speak on AI and growth. If you read it, read Table 3 first. The growth challenges in the left column, over-reliance on paid channels, unclear ideal customer profile, product differentiation, churn, fundraising, are the same ones founders had before the models arrived, and they are the same ones they have now. The tools changed. The problems did not.
FAQ
Who wrote "Generative AI for growth hacking" and where was it published? Arash Rezazadeh, Marco Kohns, René Bohnsack, Nuno António and Paulo Rita. Journal of Business Research, volume 192, article 115320, published online 21 March 2025 as part of the special issue "Application of Machine Learning and AI in Marketing". It is open access under a CC BY licence.
When were the interviews conducted? February and March 2023, at 20 companies in the US, Germany, Portugal and five other countries, with founders, CMOs, VPs of growth and sales directors.
What is the AI Capabilities Framework? A three-step model from the paper for how a startup builds competence with generative AI: build knowledge of the tools, refine it by combining tools, then develop reflective knowledge, the ability to judge outputs for accuracy, copyright and hallucination before using them.
Did the study predict agentic AI? It predicted the direction: autonomous agents integrated into business operations, with AutoGPT and prompt chains as the early form. It expected that to take five to ten years. It took roughly two.