The marketing loop: research, make, measure, learn

It is natural to judge an AI-native firm by the AI it uses. Which model writes the copy, which one makes the images. It is the wrong place to look, for two reasons that are both visible in the public data.

The first is price. Stanford's AI Index found that the cost of running a model at the level of GPT-3.5 fell more than 280-fold.¹ The investment firm Andreessen Horowitz, tracking the cheapest model that reaches a fixed quality bar, put the decline at about ten times a year, or a thousand times in three years.² The research group Epoch AI measured the same trend task by task and found declines of between 9 and 900 times a year, depending on the task.³

The second is sameness. The same Stanford report found that the gap between the best model and the tenth best had halved in a year, and that the top two were separated by 0.7%.¹ A year on, the 2026 edition described the performance gap between American and Chinese models as having effectively closed.⁴

Put the two together. Every firm can rent models of nearly the same quality, at a price that keeps falling. Whatever a firm does that others cannot copy, it is not that.

The best evidence on how often ideas work comes from software companies, which test changes on real customers at very large scale. At Microsoft, only about one third of the ideas tested improved the metric they were designed to improve.⁵ Bing, running more than 250 experiments on a typical day, reported the same: fewer than a third moved their metric.⁶ These were ideas from experienced teams who expected every one of them to work.

The same research shows why the testing is worth it. One Bing test, a small change to how ad headlines were displayed, raised revenue by 12%, worth more than $100 million a year in the United States alone. It had sat in a backlog for months because nobody thought it important.⁷

Nobody can pick the winners in advance. The firms that win are the ones that find out fastest.

These are product experiments, not ad campaigns, but the lesson carries across. If most ideas fail and the rare one pays for all the rest, then the number of good ideas you can test in a month matters more than how confident you are in any one of them.

Advertising adds a second pressure: even a winning ad wears out. A Meta analysis of purchase campaigns found that after four exposures to the same creative, a person's likelihood to purchase had dropped to roughly 60%.⁸ Meta has an interest in advertisers making more ads, so read the figure with that in mind. The direction is not in doubt: what works this month will work less well next month.

A marketing loop is the same four steps, repeated, with each turn starting from what the last one found.

1. 01Research. What your customers say in reviews, what competitors are running, what is selling, what the season is doing. This decides which ideas are worth making.
2. 02Make. Ads and emails built from what you really sell, your approved claims and your visual style. Each piece is made to test a stated idea, so a result means something.
3. 03Measure. Results read back as they land and joined to the piece that produced them: this hook, this photograph, this offer, this audience.
4. 04Learn. What worked and what did not, written down as a record for your brand, and used to plan the next round.

None of the steps is new. Agencies have always researched, produced and reported. What AI changes is the cost and speed of the first three, which makes it possible to go round the loop in days where it used to take a quarter.

Most marketing set-ups have all four steps and still do not have a loop, because the steps are not connected. The breaks are at the joins.

• —Research that nobody uses. A competitor review delivered as a slide deck at the start of the engagement and never opened again.
• —Work with no idea behind it. Ten ads that differ in ways nobody chose on purpose. When one wins, nobody can say why, so nothing is learned.
• —Results that stop at the report. A monthly dashboard the client reads, while the people making next month's work never see it.
• —Learning that lives in a person. The account manager knows what works for your brand. Then they leave.

Measurement is its own weak point. A study of 663 advertising experiments on Facebook compared the true effect of each campaign, measured by a randomised test, with what standard non-experimental methods estimated. For upper-funnel outcomes the estimates were off by a median of 115 percentage points, when the true lift was 28%.⁹ A loop that learns from wrong numbers learns the wrong things, so how a firm measures matters as much as how fast it makes.

If models are cheap and alike, and production is fast for everyone, the lasting difference between firms is what each turn of the loop leaves behind.

That is a record specific to your brand: which claims your customers respond to, which photographs sell, which offers pay back and which only move sales forward, how quickly your ads wear out, what your customers ask before they buy. No model comes with this, and a competitor cannot rent it. It only comes from running the loop on your brand, and it gets more useful every month.

It also means the record should be yours to see. A firm that keeps what it learns about your brand in a form you can read is building something for you. One that keeps it in its staff's heads is building it for itself.

This is how Lightvector is built. Every brand has a hub that holds its facts, proof and style; every piece of work is made from it and tied to the idea it tests; and every result is written back as a learning for the next round.