Experiment summary
- Orange tested showing a €10 monthly Livebox Fibre promotion only to visitors a predictive model flagged as hesitant.
- A three way split compared no promo code, a promo code for everyone, and the AI targeted code.
- Kameleoon reports promo code use fell from 68% to 25% of purchases, with conversion up 11.6% against no code and 7.6% against the universal offer.
- The case omits underlying conversion rates, sample size, confidence, test dates and any revenue or margin impact.
- Kameleoon sells the platform and published the case, so the result has no independent analysis.
Orange tested a blanket €10 offer against an AI that targeted only hesitant visitors.
A discount can reach someone wavering who needs the nudge. The other was reaching for a card and just received a gift.
Orange, the French telecom operator, faced that exact split on its Livebox Fibre journey, the sales path for its home fibre broadband. A €10 monthly promotion sat ready to use. Showing it to everyone meant spending the incentive on shoppers who would have bought anyway. Hiding it meant losing some hesitant visitors. Marketers call the first problem discount leakage, which is margin handed to people who did not need it.
So Orange tried a third route. It trained a machine learning system to spot hesitation in real time, then tested whether the discount worked better when only those visitors saw it.
Conversion was half the problem
Orange and its lower cost Sosh brand receive 23 million monthly visits each, according to a case study published by experimentation vendor Kameleoon. Kameleoon sells software that lets companies run A/B tests and personalise their websites. Mobile brings in 53% of Orange traffic and 68% of Sosh traffic.
Volumes like that turn small conversion changes into serious money. They also make a blunt promotion expensive. A universal discount can lift orders while quietly lowering the value of orders that would have happened anyway. Zalando ran a related test when it let AI nudge its prices and sales did not flinch.
Orange therefore needed a narrower answer. Could it lift Livebox Fibre conversion while keeping the €10 monthly offer away from most buyers?
The company had plenty of testing experience to build on. Kameleoon says Orange set up a dedicated conversion rate optimisation (CRO) centre in 2017 and uses Kameleoon for testing and personalisation. CRO is the discipline of testing changes to a site so more visitors buy. Contentsquare, a behavioural analytics platform, supports that work. Google Analytics 4 and Qualtrics, a customer feedback platform, also appear in the reported stack.
Orange trained the model before showing the offer
Kameleoon says its predictive system watched behaviour on Orange’s site for several weeks. Once calibrated, it sorted visitors into three broad groups by their predicted likelihood of converting. A growth team could audit the commercial logic with ease. Keep the offer away from visitors unlikely to buy, avoid discounting those likely to buy anyway, and aim it at the uncertain middle.
This is where the test becomes more useful than a standard personalisation claim. Orange compared targeted discounting with two other options in a three way split.
- The control group saw no promo code.
- The first variation showed the promo code to every visitor.
- The second variation showed it only to visitors the model identified as hesitant.
The design tested both sides of the decision. The control measured whether an offer helped at all. The universal offer arm stood in for the common alternative. The AI arm tested whether prediction could point the cost of the promotion at people more likely to change their behaviour.
Targeting changed who received the discount
Kameleoon reports that 68% of purchases used a promo code when the offer showed by default. Under AI discount targeting, promo code use fell to 25% of purchases.
The same vendor published case says the targeted version lifted conversion by 11.6% against the no code control and by 7.6% against the universal offer variation. Those are relative lifts as Kameleoon presents them. A relative lift of 11.6% would take a 2.00% conversion rate to about 2.23%, which is only an illustration, since the source leaves out the underlying conversion rates, sample size, statistical confidence, test dates and any revenue or margin impact.
Those gaps limit what can be concluded. The test supports a reported improvement in conversion and a sharp drop in discount use. An outside reader still cannot calculate incremental profit, or judge whether the three groups were balanced across devices, acquisition channels and customer types.
A sourcing limit sits alongside those gaps. Kameleoon supplied the AI personalisation and experimentation platform and published the case. Orange executive Véronique Lannuzel, whom the vendor identifies as head of its Design Thinking and CRO expertise centres, is quoted endorsing the platform. The figures arrive without independent analysis or a full experiment readout.
The copyable lever is the comparison group
A smaller company can borrow the operating lesson without rebuilding Orange’s model. Stop judging a discount only against no discount.
Add a universal offer arm and track two outcomes together. One is the conversion lift. The other is offer usage among completed purchases. Suppose targeted incentives convert better than the control but only match a blanket offer. The targeting may still create value by shrinking the number of discounted orders.
The result therefore needs to be assessed on more than conversion alone. A growth team needs contribution margin, which is revenue minus the direct costs of the sale including the discount, or discount cost or revenue per visitor beside it. Orange’s published case stops short of those numbers, even though the experiment exists to protect margin.
Give the targeting rule its audit. Teams can compare a simple behavioural segment with an AI score before adding technical complexity. They can also hold back a slice of the eligible audience so the model’s effect stays measurable after rollout.
The next test should chase profit
Kameleoon says this was Orange’s first AI driven personalisation experiment and that the company plans to extend the approach to product recommendations and its Orange and MySosh apps.
The more revealing follow up would stay with the original problem. Vary the value of the incentive, keep the three way comparison and measure incremental contribution margin alongside conversion. A €10 offer may be enough for one hesitant visitor and too generous for another.
Judge an AI growth system by the decision it changes. Predicting intent is only the start. In Orange’s case, the reported result is strongest where the experiment asks the model to decide who should go without the promotion.
| Experiment detail | What the reporting shows |
|---|---|
| Change | A predictive model sorted visitors by likelihood to convert and showed the €10 offer only to hesitant ones, tested against no offer and a universal offer. |
| Reported result | Kameleoon reports conversion up 11.6% against no code and 7.6% against the universal offer, with promo-code use falling from 68% to 25% of purchases. |
| Evidence | Vendor-published Kameleoon case study quoting a named Orange executive. No independent analysis. |
| Limits | Kameleoon-published result; no independent analysis or profit data. |
FAQ
How did Orange run the test?
It split Livebox Fibre visitors three ways. A control saw no promo code, one variation showed the €10 monthly offer to every visitor, and another showed it only to visitors a predictive model identified as hesitant.
How large were the reported gains?
Kameleoon reports conversion up 11.6% against the no code control and 7.6% against the universal offer, and promo code use falling from 68% to 25% of purchases.
Does the case prove the targeting was profitable?
No. It does not publish underlying conversion rates, sample size, statistical confidence, test dates, revenue or margin, so incremental profit cannot be calculated from it.
How can a smaller company copy the idea?
Test a targeted discount against both no offer and a blanket offer, track conversion lift and offer usage together, and judge the result on contribution margin or discount cost, beyond conversion alone.
Sources
- Kameleoon case study on Orange’s AI predictive targeting
- Orange company website
- Véronique Lannuzel on LinkedIn
Compare the evidence and operating lessons from more AI experiments in our Growth Signal Index, or browse benchmark tooling in the Industry Contents Lab.
Read next
What AI shopping systems still get wrong Our analysis explores why understanding shopper intent remains difficult even when AI improves product discovery.
For another approach to choosing who sees an offer, read Very’s test of AI-selected shopper experiences.
