Opinion
AI has turned marketing into a moving target
AI has compressed the lifespan of marketing tactics, forcing teams to replace fixed annual plans with a continuous cycle of testing, measuring and retiring what no longer works.
The first banner advertisement appeared in 1994 and returned a 44 percent clickthrough rate. Two decades later, banners were wallpaper. Every marketing tactic decays once enough people copy it. What has changed is the speed: tactics now spread and expire quickly enough to make an annual marketing plan outdated long before the year ends.
Growth used to reward whoever found the opening first, and the reward lasted. Hotmail added a signup line to every outgoing message and rode it for years. Dropbox paid users in storage for referrals. Tactics like these bought enough runway for growth to become a profession built around funnels, attribution models and quarterly plans. All of it rested on a quiet assumption: a tactic keeps working long enough to plan around.
Copying a tactic used to take real work. You had to notice it, understand why it worked and hire someone capable of running it. Now it can begin with a screenshot pasted into a chatbot, and the interval between a tactic being effective and being everywhere has collapsed. In our own tracking of 412 tactics since 2023, a cold email opener lasts roughly seven months before inboxes adapt and spam filters finish the job. Other teams will measure this differently depending on what they run and where. I have not met anyone who still measures it in years.
That seven-month lifespan breaks the planning cycle. A company writing an annual marketing plan is committing twelve months of budget and headcount to components with a shelf life of months. The plan is not wrong when it is written. It is wrong by the second quarter, and most organisations have no mechanism for noticing, because their reporting is built to show whether the plan was executed rather than whether it still works.
The closest working model I have found for this environment comes from security research. A security team assumes that every opening is temporary: once it becomes widely known, systems adapt and the window closes. Marketing teams now face the same pattern as channels saturate, feeds are reweighted, inboxes learn and communities change their rules. The relevant lesson is to organise around decay: expect tactics to stop working and build a process capable of detecting that moment quickly, without treating audiences as systems to exploit.
So the asset has moved from the playbook a team found to the rate at which it can produce the next one and retire the last. One name for this emerging model is Agentic Growth Hacking: using AI to observe, test and adapt across channels continuously rather than simply producing more marketing content.
The more consequential role for AI agents is not better copy but broader, continuous presence. Attention markets now resolve inside windows too short for human working hours. LinkedIn settles much of its distribution decision about a post within the first eighteen minutes, and a reaction in minute five carries several times the weight of one in minute forty. When we ran an agent across eleven communities inside that window, the difference in reach was not marginal. The mechanism was unremarkable: the agent was simply present in eleven places at once, which no person can be.
Financial markets went through a version of this thirty years ago. When decision windows shrank below human reaction time, machines took over execution and people moved to designing the strategies. Marketing is beginning the same reordering: less manual operation of each channel, more experiment design, boundary setting and decisions about what agents should do.
A second shift may prove even more important. A growing share of buying research now ends inside an AI-generated answer rather than a page of links, which means part of the audience being written for is a model summarising on a person's behalf. After six weeks of filling documented citation gaps on Wikipedia, we saw answer engines quote those citations back to buyers who never visited a search page or website. The vocabulary for this channel is still unsettled, as is the question of what legitimate writing for machine readers should look like.
This also raises a question the industry has not answered. Agents operating across attention windows at machine speed can describe both a research programme and a spam operation. The tests I would apply are whether the recipient asked for the contact, and whether the community would endorse the behaviour if it saw the whole picture. But individual restraint is not enough: if enough companies point agents at the same communities, those communities fill with noise and close. Platforms will adapt, and they should. Discovering that something works does not remove the responsibility to ask what happens when everyone uses it.
Most of this also fails. The experiments that produce a number worth writing about are a small fraction of those that get run, and reporting only the survivors misrepresents the work. Honest measurement makes the picture harsher because the useful checkpoint is rarely launch day: a Reddit comment that is live on Tuesday can be gone by Sunday, and much of what a team produces never survives long enough to count.
None of that changes the direction. Every era of marketing has invented its own jobs, and nobody planned to become a search specialist until a search engine made it a career. The roles forming now are recognisable in outline: someone who designs small tests and reads decay curves, closer to a quantitative analyst than a campaign manager; someone who supervises agents, approves what needs a human and decides when an experiment is shut down; and the still-unnamed work of writing for machine readers. None of these appeared in a marketing curriculum five years ago.
For a company deciding what to do, the useful move is smaller than a strategy. Pick one mechanism in one channel, state what you expect to happen, choose when you will measure it and agree in advance what result would cause you to stop. Then observe what changes, form the next hypothesis and test again at a scale small enough to limit the downside. In an Agentic Growth Hacking lab, agents expand the number and speed of those tests, while people retain responsibility for judgment, brand boundaries and the decision to stop. A team organised this way spends less effort defending an annual plan and more effort learning before the market moves again.
Finding a tactic tells you what worked last quarter. The more durable advantage is the capacity to keep finding, testing and retiring tactics before the market moves again.
Mickey Haslavsky is Founder and CEO of enso.
