3 years of Intent: A retrospective and revolution

What started off as a concept in Intent has morphed into a movement. Hence the capital “I”.

Maybe it’s the cocktail party effect at play, but whenever I hear “Intent” my ears prick up.

We’re the authority and have been for the past three years. Whenever the word gets mentioned, regardless of how tenuous the context, I immediately assume whoever is saying it is talking about us. That’s some subjective marketing recognition, subconsciously owning a word, isn’t it?

It’s not just that I see it creep into day to day lexicon when brands talk about user behaviour. Nor that martech vendor homepages have started to use it, specifically over the past two years. Of 25 complementary martech vendors I reviewed, the likes of Bloomreach, Wunderkind, even Salesforce, the word Intent appears on 13 of them. Two years ago it appeared on none of them. (Thank you, Wayback Machine.)

But it’s because reading Play Bigger by Al Ramadan et al. explains how categories get created. Of the four internal and external signals that define whether you’ve created a new category, we’ve satisfied three.

Intent was created in 2024, so not even three years ago, albeit with years of research before it. Six years of building and running User Conversion, the UK’s largest conversion rate optimisation consultancy, was the litmus test of what eCommerce brands really wanted under the hood. Writing The Person in Personalisation, theoretical research involving over 145 personalisation experts, cemented it.

I felt like I knew what brands wanted. A way to measure success that wasn’t as short termist, binary or retrospective as conversion rate. Nor something as long termist, and dare I say intangible, as lifetime value. Something that reflected how users oscillate between buying stages and measured their journey and growth, firmly putting the person back in personalisation.

Despite all that research and experience, nothing could have prepared me for the evolution Made With Intent would go through.

We have recently announced a fundamental moment in that evolution, codenamed “Intent Next”. Our v4.0. So, please indulge me while I share the past three years of learning and the iterations we’ve gone through as a business.

Looking back, there’s a pattern in all of it that I couldn’t see while it was happening. Every version of this product did the same thing.

v1.0. Intent identification

You’ll notice we don’t use the “p” word much. Part one of making personalisation achievable was moving the metric away from the things that were never going to get us there; conversion rate, transactions. Really putting the person back into personalisation.

So the first job was measuring Intent.

Our earliest models, in 2023, were built site by site. With Boosted Trees and rigid custom tracking, they worked, for a while. Performance degraded while the whole approach lacked scalability.

We replaced them with a single neural network trained across everybody, which taught us that shared intelligence matters (although site level context couldn’t be ignored). This evolved into a model with a continuous rolling snapshot of the website itself as a second input. We learned that Intent is the interaction between a user and a site, not a user in isolation.

Two input streams into a wide and deep network, simultaneously addressing the cold start problem while compounding the accuracy of a global model with the context of a single one.

The first reads the person: around 600 features covering clicks, navigation, mouse movement, dwell time, time since they last did X, events since they last saw Y.

The second reads the environment they’re in, roughly 800 features covering page structure, price distributions, traffic composition and how converting and non converting visitors behave on that specific site, refreshed every three hours. Context on top of context.

Intent became a decision layer on top of whatever else a brand ran and we still run that today. Session summaries push an end of visit report card into Klaviyo or Braze, carrying purchase confidence, which way it was travelling and how likely that person is to come back, so a follow up fires on where the visitor actually got to rather than on the crude fact they looked at a product.

GA4 gets predictive dimensions, so paid audiences stop being built on page views. Snowflake, GCP and Azure get the raw event level data, so analytics teams can finally answer which traffic was worth having rather than how much of it there was.

That last one mattered more than we expected. Around 10% of visitors sitting on checkout pages aren’t ready to buy. Pages had been standing in for people for twenty years and everybody knew it, but nobody could prove it.

We got very good at this. Too good, in a way, because we spent a year being the most interesting dashboard that nobody acted on. That was the problem. Despite shipping an Intent Performance Score (think your credit score) trying to be actionable, brands couldn’t understand “we need to work on building intent” as opposed to “we need to increase conversion rate”.

Our mantra was “stages not pages”. Strategically bold, customers liked the freshness of it. We’d built a product to change brand mindset rather than one that met them where they were. They were being asked in board meetings why product page performance was down, while we were asking them what stage their visitors were in.

We now say “pages then stages”. Meeting customers where they are at. Our framework still sits as this: stages, signals and trends. The problem underneath didn’t move though. Customers could see intent beautifully, but they still couldn’t do anything with it. They wanted action.

v2.0. Responding to intent

To act on intent we needed to give brands the ability to create experiences. That meant that we’d plug into the experience stack brands already had. Intent triggers for Bloomreach, Dynamic Yield, AB Tasty, Kameleoon, Optimizely and plenty more. But also, we had to build the delivery layer ourselves.

We built our own Campaign Manager with a visual builder, template library, all customisable without a developer, both on top of the page and within the page. For the first time, a brand could understand intent and respond to it in the same place, without the signal going stale on the way. Actionability solved.

Targeting an experience on Intent made it, on average, 18.9% more effective than running the same experience across all traffic or against a segment you built yourself. Purchase confidence went up 24% against generic targeting. Fewer people saw the thing. More of the right ones did.

Email capture is the clearest example. The incumbent is a pop-up on a timer: on landing, after three pages, at thirty seconds. 55% of shoppers tell us early pop-ups annoy them and 40% say they make them less likely to buy. Time the same pop-up to the moment someone starts signalling they are about to leave and sign-ups go up 40% on average. Oh, from better subscribers too. Bensons for Beds kept their existing Bloomreach experience and changed only when it fired: sign-ups went up six times over. Le Chameau went the other way and showed theirs to fewer people, cutting focused shoppers out altogether, and a 3% lift in sign-ups turned into 24% incremental revenue from that segment.

Social proof is the one that ought to change your mind about all of this. “X sold last week” in checkout lifts conversion 5% for high intent visitors. For low intent visitors, a much bigger group, it drags conversion down by 1%. Run it for everyone, as almost everyone does, and you book a modest net positive with no sign that the larger half of your audience is quietly worse off for it. Golfbreaks had avoided social proof for years for that reason. They found popularity cues worked on low intent browsers and hurt everyone else, showed it only there, and took 5% conversion out of that segment. A global luxury retailer did the inverse, swapping reassurance for urgency on high intent shoppers backtracking in the basket, and got 25%. Across the base, sequencing social proof to intent averages a 20% lift.

But that’s not to say that it worked all the time.

Brands would create an experience, and guess when to fire it; which intent would best suit that experience.

Ultimately, brands were still guessing when someone would see the experience. What says an email capture should target low intent visitors? Why not low intent who are likely to leave? Or low intent, likely to leave, on paid media only? On mobile? On a category page? In their second session?

Despite the context within the targeting mechanism being undeniably better than the original, it’s still guess work and not scalable. At the end of the day it’s still a manual rule.

v3.0. Intent decisioning

Agentic experience delivery takes that guesswork away. By the end of April 2026, we built the ability to serve experiences via agents to the right context layers, excluding those that do not create impact. We haven’t looked back since.

The thesis was simple; stop testing treatments against (the guessed) segments you built. Instead, hand the agent your experiences and the metric you care about, and let it discover the audiences and the moments that matter itself.

In essence, we built a contextual bandit which trains a model that predicts performance for a given context. It takes priors from places that have never happened and uses them somewhere else (exactly the same logic that lets the intent model work on a site it has never seen by the way).

So, per campaign, this means you get up to 500 contextual combinations, which resolves to somewhere between 60 and 120 live segments on a typical retail site, plus up to 75 moment based triggers, all re-scored every night against a rolling 90 day window. Nobody built any of them, instead, the agent picks its own groupings where impact was best seen and best felt.

There’s no single winner, which is the bit that breaks people’s mental model. Different experiences win for different audiences. That’s personalisation, right?

Take Diamonds Factory, for example. The incumbent was a blanket 25% off to every basket abandoner. The first pass in V2.0 looked good; an 11% conversion uplift (despite average order value falling) for those likely to abandon. For v3.0, they gave the agent five options instead including a 25% off, a 15% off, an FAQs message, a finance message and, explicitly, showing no experience. Different messages designed for an agent to allocate to individuals, at different times of their journey. Revenue per customer increased by a further 13.8%. And for 52% of customers, the agent determined that sometimes showing no experience at all was exactly the right call. That decision on its own was worth £1.6m in incremental annual revenue.

Across the base, 77.6% of agentic campaigns produce a positive result. A stark difference where a split test in v2.0 demanded one winner across all traffic; where an agentic campaign in v3.0 succeeds by finding context specific answers. 70% of customers built a campaign inside the first month. Platform visits doubled. Nobody had to be talked into it, which after three years of explaining intent to people was a novel experience.

It truly is personalisation, at scale, based on the context that really matters: user intent.

Did I see this coming?

No. Not remotely.

In the early days, we were running a script across four websites trying to prove that a model could tell who was ready to buy. I’d have told you the destination was better insight. Better insight turned out to be table stakes, and close to useless on its own. It’s no secret that high intent equals a high conversion rate.

What we did have right was the question. The Person in Personalisation argued that personalisation failed because it operated on website attributes rather than human ones, and that recommendations were the only type of personalisation that ever worked. Why? No human was writing the rules.

It took us three years and four versions to apply that sentence to everything else.

What we have now is a model that reads intent and makes the decision. Brands still write the strategy it works from, and they can now see what it’s learning and where their input would help.

So where does it go next? The maths of contextual bandits is well understood and broadly the same wherever you find it. It will be commoditised, and quickly. It’s the only way to do personalisation at scale. What differs, enormously, is what you let the model know about the person. Attributes tell you who walked in, but Intent tells you what they’re about to do. There is nothing more indicative of someone’s intent than their real time behaviour; and that’s what we track.