è����Ƶ

Editor's note: this article is adapted from a recent è����Ƶ Live session with Jack Simkins, Digital Product Manager at Golfbreaks.com.

A test that lifted time on site by 14% still spent its first week looking like a failure. Conversion rate was down. Under a traditional A/B testing programme, Golfbreaks.com would have pulled the experiment.

They didn't. And the reason why says a lot about what happens when a genuinely high-consideration purchase journey meets a testing programme built for one-session eCommerce.

is a golf tour operator based in Windsor, with offices in Copenhagen and Charleston, though the Made with è����Ƶ account focuses on the US and UK. It sends golfers on trips ranging from a single night in the UK to a week in Spain or Portugal, across a lot of different golfer segments. It's a lead generation business first. Visitors don't check out online in one sitting, they enquire, then a sales agent works out flights, transfers, accommodation, and course access, and gets them to a booking over the phone, sometimes weeks later.

That's not unusual for travel, where research and comparison typically happen across several separate visits and sites before a decision gets made. Optimising a single-session conversion rate for a purchase that actually plays out over weeks measures the wrong moment entirely.

A booking journey that can't be forced into one session

Jack has spent seven years at Golfbreaks.com, the last couple focused on conversion rate optimisation. "We've got quite a unique scenario whereby we're trying to encourage that inquiry," he said. "Particularly in travel, in the industry in general, it's quite an unusual thing to not be able to book entirely online."

A small portionof trips get booked online. But most go through a sales agent, because a golf trip has too many moving parts (courses, transfers, flights, accommodation, and group logistics) for most visitors to configure and commit to in one sitting.

Colin Spooner, Principal Value Consultant at Made with è����Ƶ, put his finger on why that matters: "It's not our traditional eCommerce brand where it's a pure purchase journey. But that almost plays into the hands of intent, where you need to think about that considered purchase and how to get people through the funnel before even thinking about the booking, weeks and months down the line."

Measuring what happens before conversion

"You are what you measure" is a phrase the Golfbreaks.com team has adopted internally. If a new visitor is unlikely to enquire on their first visit, optimising purely for enquiry rate on that visit measures the wrong thing.

So alongside enquiry rate, the team tracks bounce rate and time on site together (a new visitor who bounces immediately clearly hasn't been given a reason to stay), pages viewed per session as a depth-of-exploration signal, and, specifically, movement from low to building intent, the kind of signals behind Made with è����Ƶ's content prioritisation and messaging use cases. None of these are vanity metrics here. They're proxies for whether a visitor is progressing through a decision process that runs across several sessions, not on a single visit.

Before building any experience, the team asks these questions to get in their customers shoes:

• What is a brand-new visitor actually trying to work out?
• Who are Golfbreaks.com?
• Can we be trusted?
• Do you have to pay full price up front?
• Can you book online at all?

That last one is really important. Because Golfbreaks.com can't be booked entirely online, setting that expectation early avoids disappointment later in the funnel, right when a visitor is closest to converting. As Colin put it, getting that messaging right up front was "a huge realisation" for how the whole experience needed to be built.

What agentic campaigns change about testing

Golfbreaks.com's testing programme runs on Made with è����Ƶ's agentic campaigns. In standard A/B testing, you decide up front which segment sees which variant, based on a hypothesis about who will respond to what. Agentic campaigns invert that: you define the strategy (the moment you're trying to influence, and the goal, whether that's enquiries, conversions, or a secondary metric) and hand the agent your set of tactics. It tests them against real segments and works out which one performs best, for whom, and when, using the same intent signals that power the rest of the platform.

For Jack, a self-described non-developer, the practical benefit was speed. "The tool allows me to get these experiences up much faster," he said. "My concept-to-live process is significantly shorter... it means the agents have got time to learn."

But the deeper change is what gets removed. "No longer am I having to set up those individual segments, or serve experiences to segments that I think will benefit from them," he said. "It's in the hands of the agent to then work out what segments it would benefit from... It's a much wider net." A message built for low-intent visitors might also help a segment already building toward a decision. A manual test is only as good as the human guess behind who it's shown to. An agent testing against a hundred segments simultaneously doesn't have that blind spot.

The "do nothing" variant is a genuine conversion tool

One of the more counter-intuitive parts of Golfbreaks.com's setup is what Jack calls the "do nothing" variant. In standard A/B testing, every visitor sees a control or one of several variants. Agentic campaigns add an option where the visitor sees nothing added or changed at all.

"The do nothing essentially sits within those variants as a copy of the control," Jack explained. "It's a safety net because it prevents us from showing negative experiences to customers that don't need to see it."

"In most experience it's always about adding things onto your site," Colin observed. "Having a version where actually sometimes the best thing is leaving the customer alone to progress, or even suppressing things on site, is a nice alternative to what we've experienced over the last 10, 20 years in experimentation." As the data below shows, it's frequently the top performer, because some visitors don't need an intervention. They're already progressing on their own, or they arrived with enough context that added messaging just gets in the way.

When the agent's early data looks wrong

Here's where the seven-day lesson from the top of this article comes back in. Jack's team built two welcome-visit experiences using the same messaging, one for the homepage, one for a location page such as a product discovery landing point for someone who searched "golf breaks in England."

The homepage experience delivered a 14% uplift in time on site. But conversion rate showed a negative trend for the first seven days. "With traditional AB testing, I would have perhaps turned it off," Jack said. "I would have panicked when I saw negative 14%, and I would have said, this isn't working."

He didn't, because the agent was still learning. That's consistent with what independent testing research shows more broadly: . Once the agent had enough data, performance turned around.

Same messaging, two different visitors

The location-page test surfaced something else: identical messaging performed in opposite ways depending on where a visitor arrived. On the homepage, a "how to book guide" message performed best, evidence of a genuinely low-intent visitor who needs some hand-holding.

On the location page, the top performer was a trust-building "number one tour operator" message and the do-nothing variant. Jack's take is that a visitor who searched "golf breaks in England" already has affinity toward the destination, closer to a returning visit mindset than a cold product visit. They don't need the basics explained.

A message about Golfbreaks.com's customisable packages underperformed with brand-new homepage visitors. It's true and important, but it's the wrong message at the wrong moment, the same lesson behind Made with è����Ƶ's discounting use case: showing a message before a visitor is ready for it does more harm than good.

The surprise that only showed up in the data

Asked what surprised him most, Jack pointed to something that had been sitting in plain sight. An early "ready to plan?" message aimed at brand-new visitors looked like a reasonable nudge. But in the data, it actually came across as overbearing.

"If you think about a new user landing on the site and asking, are you ready to plan? It's probably a bit overbearing," Jack said. "At the time, when you're setting up those tests, it's like, right, I'm going to use the same messages for the homepage, same message for the location page. They're surely going to work." They didn't, for every segment.

Colin's read: "The amount of times we see customers who have a predefined view of what will work, and it's completely different. That point around being subjective comes to life when you start to see the way the agent starts to make decisions." That tracks with the broader shift in shopper expectations, , a generic message is no longer neutral, it actively reads as a miss.

When a message doesn't resonate with any segment, the fix is simple. Delete it, let the agent relearn, and add a new tactic later if needed. You don't need to manually re-segment.

Measure the journey, not just the moment

None of this is unique to golf holidays. Any purchase with a real consideration cycle shares the same shape, including B2B software, where . A first-time visitor is rarely the same as a returning, further-along one, and treating them identically wastes the message on the visitor least ready to act on it.

Three things carry over regardless of your industry. You don't need to be a developer to find early wins, a visual editor is enough to start. If your purchase journey has any real consideration cycle in it, top-of-funnel testing should measure more than conversion. And build a tactic sheet of messages at a global level, then let the data show which ones resonate with which visitors, rather than deciding that yourself up front.

If your own funnel has visitors who aren't ready to buy on visit one, the same logic applies. See the intent framework behind it. Or book a demo to see it against your own traffic.

If you've enjoyed this write up of our latest è����Ƶ Live session, why don't you join our

Why standard advice on eCommerce bounce rate might not be enough

In 2023, Google Analytics 4 replaced Universal Analytics (UA) as the default, shifting the definition of bounce rate entirely. Under UA, any single-page session counted as a bounce, regardless of how long the visitor stayed or what they did. Under GA4, a "bounce" is a session with no meaningful engagement in the first ten seconds (, 2023).

Most published advice on reducing eCommerce bounce rate, including almost everything ranking on the first page of Google right now, predates that change. The benchmarks cited, the comparisons drawn, the thresholds used to define a "good" or "bad" rate: much of it is calibrated to a metric that no longer exists in its original form. It's worth bearing in mind before treating a published figure as a reliable signal that something is broken.

There's a connection here to the intent signals argument. GA4 defines a bounce as a session with no meaningful engagement, which is itself the absence of any intent signal. However, the metric has quietly moved closer to what we're suggesting: that what matters is whether a visitor showed signs of engagement and intent, not simply whether they viewed more than one page.

The standard fixes aren't useless. A page that takes four seconds to load on mobile will lose shoppers. Navigation that buries products three levels deep creates friction. These are real problems. But they're also largely table stakes. Most mid-market eCommerce teams have addressed them, or at least know they need to.

What the generic checklist rarely asks is: why did this particular visitor leave this particular page? Speed explains some of it. Confusing navigation explains more. But there's a third explanation that gets far less attention: the visitor arrived with a specific intent, and the experience they landed on didn't reflect it.

What on-site intent signals actually are

è����Ƶ signals aren't abstract. They're specific, observable events already firing on your site every day, most of them visible in your analytics if you know where to look.

Consider what happens in the first thirty seconds of a session. A visitor lands on a product detail page. Do they scroll past the first image, or stop there? Do they interact with the size selector, or skip straight past it? Do they hover over the "Add to Basket" button without clicking? Do they navigate to a second product, return to the category page, or leave entirely?

Each of those micro-behaviours carries a signal. Taken individually, they don't mean anything. Taken together, they start to suggest something about intent: whether the visitor is browsing loosely, comparing seriously, or hitting a wall they can't get past.

Some of the most informative on-site signals include:

- Scroll depth: how far down the page a visitor gets before stopping or leaving

- Hover behaviour: where the cursor lingers without a click (interest that didn't convert to action)

- On-site search queries: what visitors type into the search bar, and crucially, what they do next

- Dwell time relative to site average: a visitor spending significantly longer on a PDP than average may be closer to buying than the raw bounce metric suggests

- Variant and size selection: engaging with product options is a meaningful buying signal, regardless of whether a purchase follows

- Back-navigation patterns: returning from a PDP to the same category page repeatedly often indicates comparison behaviour, not disinterest

None of these signals individually tells you what a visitor intends to do. But patterns across them might tell you something useful about why they're not finding what they came for, and whether the bounce that follows is a genuine commercial loss or an inevitable one.

For a deeper look at why behavioural signals tend to outperform the proxy metrics most teams rely on, Predictions Not Proxies, our blog post, is worth a read.

The signals worth watching by page type

One limitation of tracking bounce rate as a single site-wide number is that it flattens very different problems into one metric. A visitor who bounces from the homepage is probably experiencing something quite different from one who bounces from a product detail page after two minutes of engagement. The signals worth reading, and the interventions that might help, differ depending on where the bounce is happening.

Homepage

Homepage bounces are often about relevance at first impression. Was the visitor expecting something the page doesn't immediately surface? Traffic source matters here. A visitor arriving from a paid social ad promoting a specific sale and landing on a generic homepage is likely to read that as a mismatch before they've even scrolled.

You should, instead, consider time to first scroll, engagement with any navigation element, and click-through to any product page. A visitor who lands and never scrolls is usually gone for reasons that faster load times can't fully address.

Category pages

Bounce from a category page often points to a discovery problem. Either the product range isn't what the visitor expected, or the tools for narrowing it down (filters, sorting, on-site search) aren't doing their job.

You could monitor for filter use, scroll depth through the product grid, and whether visitors click through to multiple PDPs or just one (or none). A visitor who opens the filter panel but doesn't apply anything may be signalling that the available options don't map to what they had in mind.

Product detail pages

PDP bounce is the most commercially sensitive, because this is where visitors are closest to a decision and where intent signals tend to be richest. Image engagement, variant selection, and dwell time relative to site average can all suggest whether a visitor is actively evaluating or has already decided the product isn't right.

A visitor who spends three minutes on a PDP, selects a size, and then leaves is a very different prospect from one who bounced in under ten seconds. Treating both as equivalent in a site-wide bounce metric misses that distinction entirely, and probably points the subsequent analysis in the wrong direction.

Paid landing pages

For pages receiving meaningful paid traffic, the most important signal is often the simplest: does the message on the page match the ad that brought the visitor here? Post-click relevance is frequently the first place to look when bounce rate on paid traffic is elevated, before speed or UX enter the conversation.

Consider a fashion retailer seeing high PDP bounce from paid social. You may think audience mismatch or slow load times. But scroll depth data tells a different story. Visitors are reaching the size selector and stopping, not scrolling away. The real problem is out-of-stock variants being featured in the ad creative. Shoppers arrive, find their size unavailable, and leave. This is nothing to do with a UX fix, simply a case of reading the right signal, the right context.

Acting on intent signals before the bounce happens

Reading these signals is useful. Responding to them is where it gets more interesting.

The standard eCommerce pop-up is a useful counterexample. A blanket overlay triggered by exit intent, offering a discount to everyone regardless of what they've been doing, ignores every signal the visitor has sent. A visitor who spent four minutes on a PDP, selected a variant, and then paused receives the same intervention as one who arrived and left in eight seconds. So, really, you're not delivering a proper personalised experience that's appropriate to your customers' different needs.

A more considered approach starts with the signal, instead of the user navigating away. A visitor showing strong PDP engagement (extended dwell time, variant selection, multiple image views) is telling you something. The right response is probably social proof surfaced at that moment: stock scarcity, recent purchase activity, a well-timed review. Not a discount. A visitor who used the search bar, found no useful results, and is now leaving needs a different intervention entirely: a related product suggestion, or a prompt to browse a relevant category.

The interventions don't have to be complex to be more relevant. Start with your highest-bounce page type, identify which signals are already firing there, and ask whether your current exit triggers reflect any of them. That's a reasonable first step, and it costs nothing to consider.

If you're looking at how this kind of approach works for browse abandonment specifically, our browse abandonment use case walks through one way to think about it.

What a bounce means for your CRM, and why it matters

Bounce rate is typically discussed as a traffic or UX problem. It's less often discussed as a CRM problem. But there's an argument that it should be.

Every visitor who leaves without converting, subscribing, or taking any traceable action represents not just a lost session, but a lost contact. For a brand spending meaningfully on paid acquisition, that loss isn't only the missed immediate sale. It's the absence of any first-party data to re-engage with later. The CAC clock is ticking whether or not the visit converts.

The intent signals that might help reduce bounce in the moment are the same signals that could inform a more relevant recovery sequence when the bounce does happen. A visitor who engaged with a specific category, hovered on a product, and then left is a different re-engagement prospect from one who arrived on the homepage and bounced immediately. Where that behavioural data is captured, it can feed a browse abandonment email or SMS that speaks to what the visitor was actually looking at, not a generic "you left something behind" message.

Most browse abandonment recovery today operates on a relatively simple trigger. The visitor viewed a product and left. è����Ƶ signals could make that logic more nuanced, and the message that follows more relevant to where the visitor actually was in their decision.

How to read bounce rate differently

Bounce rate, as a metric, doesn't tell you very much on its own. It tells you someone left. It doesn't tell you why, which page type to prioritise, or whether the bounce represents a genuine commercial loss or a session that was never going to convert.

On-site intent signals can't answer all of those questions. But they might answer more of them than page speed tests and navigation audits typically do. For eCommerce teams who've already done the basics and are still looking for what to try next, it's worth examining what visitors are doing before they leave, not just that they left.

That shift in framing, from "how do we stop bounces?" to "what are bounces telling us?", might be where the more useful work sits.

If you want to see how Made With è����Ƶ reads on-site intent signals across eCommerce traffic, book a demo and we'll walk through it with your site.

ran a discount and watched conversions climb 11%. On paper, it looked like a win, right?

It wasn't. Average order value fell 6.7% in the same period, and once the discount depth was accounted for, the "win" was worth barely £8k a month.

So, the luxury jewellery brand tried something different. It handed the decision to our told it to optimise for revenue per customer instead of conversion rate, and gave it the option to do nothing at all. 

Our model decided that for 52% of customers, nothing was exactly the right call. That decision alone was worth £1.6m in incremental annual revenue.

‍Editor's note: This post is a write-up of an featuring Jo Homer, Director of CX at the parent company of Diamonds Factory. Jo walked through a real, live test of Made With è����Ƶ's agentic campaigns, hosted by Charley Bader, VP Strategy & Ops, and Colin Spooner, Principal Value Consultant, both at Made With è����Ƶ.

Diamonds Factory is a made-to-order luxury jewellery brand, one of the brands under Neve Jewels Group. Jo was upfront that this wasn't automation for automation's sake. It was a genuine test of what happens when a business gives an AI the ability to make real commercial decisions, and watches closely what it actually decides.

Agentic campaigns, most recent product launch, replace the old manual approach (build a segment, write a rule, set it live, hope the hypothesis was right). You stay in control of the strategy, the moment being targeted, and the goal being optimised for. But the moment-to-moment decision, based on who sees what and when, goes to an AI agent that keeps learning and reallocating traffic as it goes.

The discounting trap that feels like a win

The original brief was simple: catch basket abandonment before a customer leaves, offer 25% off, and recover the sale. It's the default move for most eCommerce brands, and doubly so in a vertical as competitive as jewellery, where discounting is normal.

The first-pass numbers looked good. Conversions were up 11%. However, the problem is what sat underneath it. Average order value dropped 6.7% over the same period. Diamonds Factory wasn't gaining new revenue so much as buying back sales it would likely have made anyway, at a lower margin.

As Jo Homer, Director of CX at Neve Jewels Group, put it:

"We were basically training the customer to wait for a deal, which in return was affecting margin."

Diamonds Factory fell into a similar trap many retailers fall into. A single rule applied to everyone optimises for the metric that's easiest to see (conversions) while diminishing the one that matters (margin).

For a luxury brand, the cost isn't only the discount itself. It's what the discount teaches the customer to expect next time. Train customers to wait for 25% off, and full price starts to look like a mistake they'd be silly to make.

It's a familiar story for anyone who has run basket abandonment campaigns before. A rule fires the same offer at every visitor who matches a condition, regardless of why they left.

Some were genuinely price-sensitive and needed the nudge. Others were already going to buy and simply hadn't checked out yet. Blanket rules can't tell the difference, so it pays the same discount for both.

Changing the question the model was answering

The shift that mattered wasn't really about AI. It was about changing what "success" meant.

Diamonds Factory moved from a single rule (25% off, applied to everyone) to an agentic campaign with four possible responses: a 25% discount, a 15% discount, a trust-building message, or no intervention at all. Crucially, the goal changed too, from conversion rate to revenue per customer.

That distinction is easy to state and hard to act on, because it means giving up control over exactly which lever gets pulled and letting the model decide, case by case, based on what it observes. As Jo describes it:

"The tool didn't change. The objective did. And that's the real unlock in some cases."

The model started with no historical or behavioural data. It had to learn the pattern from scratch, in real time. The result: revenue per customer rose from £127 in the control group to £145 in the agentic group. A 13.8% uplift.

"It didn't come from offering more," Jo said. "It was offering much smarter."

Jo frames the difference as hiring someone and handing them a script to read, versus hiring someone and trusting them to think on their feet. A script covers the cases you anticipated. Under the agentic model, the system sets its own rules based on what it actually observes in customer behaviour, and updates them continuously rather than waiting for the next quarterly review.

What our agentic model found without being told

Neither Jo nor anyone on her team instructed our AI model to look for a basket-value threshold. 

It found one anyway, identifying a specific value point below which doing nothing consistently maximised revenue, and above which a discount or message made more sense. 

It also separated customers by where they sat in the buying journey, not just by what was in their basket. Shoppers who looked like they were still comparing options, still in a research mindset rather than a checkout mindset, were shown trust messaging instead. So, things like craftsmanship, heritage, and guarantees.

A discount at that stage would have read as pushy. The model inferred that without being told what a "research mindset" should look like.

"The model basically developed a theory of buyer psychology without being given one," Jo said. "No human would take months, sometimes years, to get to that kind of analysis."

Device behaviour told a similarly specific story. On desktop, 79% of customers were offered some form of discount. On mobile, 96% were shown nothing. The model had, in effect, worked out that mobile sessions were browsing sessions rather than buying sessions, and treated them accordingly, without anyone telling it.

The value of restraint

Marketers are trained to intervene. Numbers dip, and the instinct in the room is always "what should we do about this," never "should we do anything at all." Jo made the point that in a hundred trading meetings, that second question rarely gets asked, let alone answered with "nothing."

In this case, the model decided that 52% of customers were best served by being left alone entirely. 

"The highest value of action really was the restraint," Jo said. "The agent knew when to get out of the way."

That restraint was worth £1.6m in annual incremental revenue. Not because doing nothing is inherently valuable, but because for those customers, any intervention would have either converted someone who was already going to buy (at a needlessly reduced margin) or interrupted someone who wasn't ready to be nudged.

This figure is the outcome of testing across three separate markets, over more than a hundred individual tests. 

Turning Made With è����Ƶ into a trading lever

Beyond the topline number, the agentic model isn't a fixed campaign. It's something Diamonds Factory can turn on and off in step with its trading calendar.

Around Black Friday, for example, the team pauses the agent so customers already in a high-intent buying mode aren't shown a discount they didn't need. When trading returns to business as usual in January, the campaign flips back on, and the model picks up its learning from where it left off.

"It's a lever we can control," Jo said, describing how the model compounds its learning across each on/off cycle rather than resetting.

That reframes the agent from a one-off experiment into infrastructure. ����’s something that learns continuously and can be dialled up or down around known peaks and troughs. Jo's team started seeing patterns emerge after three to four weeks, and reached genuine confidence in the results within six weeks to two months, consistent across all three markets tested.

What setting up agentic campaigns looks like

, Principal Value Consultant at Made With è����Ƶ, walked through the mechanics live during the session, because "agentic" can sound more abstract than it actually is.

Setting up a campaign starts with choosing the moment it targets (new customer arrival, , browse or basket abandonment), then the goal it optimises for, revenue per customer, conversions, or something custom like a credit application.

From there, you build the toolkit of possible responses: a discount at whatever depth makes sense, a trust message, a finance nudge for customers who might be mid-month on cash flow, and, deliberately, a do-nothing option.

The agent then goes into a learning phase, typically two to four weeks, working out which response suits which individual, before it starts compounding that performance over time.

What this doesn't mean

The point that came out from the session is that a single discount applied to everyone, regardless of where they are in the journey, is a blunt instrument in a world that calls for something more precise.

It also doesn't mean every brand will find that 52% of its customers are best left alone. That figure is specific to Diamonds Factory's price points and customer base, and a mass-market retailer with a lower average order value would likely see a different split.

What should generalise is the method, not the exact number: test more than one response, optimise for a metric that reflects margin as well as conversion, and make "do nothing" a genuine option rather than an assumption nobody checks.

In conclusion: From what rule should we apply, to what does this customer need

The shift Jo describes isn't really technical, even though the mechanism is AI. It's a change in the question being asked. Diamonds Factory moved from "what rule should we apply" and "what do we think will work" to "what does this specific customer need right now."

Sometimes, as this blog post shows, the right answer to that question is nothing at all. That's a harder thing for a marketing team to sit with than a 25% discount, but it's the one that protected both the brand's margin and its positioning as a luxury retailer, while still growing revenue per customer by double digits.

If you're running basket abandonment campaigns on a single blanket rule, the question worth asking isn't whether your discount converts. It's whether it's the right decision for the customer in front of you, or just the easiest one to set and forget.

Not every abandoned basket is a fire to put out. Some are customers who are always going to come back, and every discount you throw at them is a margin you don't need to give away.

Curious what an agentic approach could do in your own basket abandonment process? with Made With è����Ƶ to see it in action.

If you liked this article and want to join our next webinar, follow the link to join our next session.