Personalisation or customisation: who holds the controls?

You sell roasted coffee online and want each customer to see something different. There are two ways. One: the system reads purchase history and picks a suggested bag. Two: the customer picks the roast, the grind and the blend. People call both "personalisation", yet they need different data, cost different money and fail in different ways. This post separates them with one question: who decides?

What separates personalisation from customisation?

The person who decides. In Arora and colleagues' (2008) definition, personalisation is when the firm decides which marketing mix suits an individual, usually from data collected earlier. Customisation is when the customer proactively specifies one or more elements of their own mix (Arora et al., 2008). Their example of the first is Amazon recommending books and music; of the second, Dell letting a buyer configure a computer before ordering (Arora et al., 2008). Both are one-to-one marketing, an extreme form of segmentation in which the target segment has size one (Arora et al., 2008).

The line is not always kept. The authors themselves note that the terms are often used interchangeably (Arora et al., 2008). Ansari and Mela's (2003) paper carries customisation in its title, yet it optimises the content of an email for each user from clickstream data, meaning records of what users click (Ansari and Mela, 2003). The firm is the decision-maker, so by the definition above that is personalisation, and Arora and colleagues file it under personalisation too (Arora et al., 2008).

Personalisation starts with collecting data and closes with a feedback loop that improves the algorithm. Customisation starts with building an interface for the customer and closes with producing exactly what was chosen. These five stages are the library's sketch, not quoted from any page of the three papers; see Going deeper.

What does system-led personalisation look like at Netflix?

Gomez-Uribe and Hunt (2015), both Netflix employees, write that the recommender system is not one algorithm but a collection of them serving different use cases. The personalised video ranker (PVR) orders the whole catalogue for each member profile, so the same genre row can show completely different titles to two people. It is not purely personal: PVR works better when personal signals are blended with "a pretty healthy dose" of unpersonalised popularity (Gomez-Uribe and Hunt, 2015).

Why let the system lead? The authors say people are surprisingly bad at choosing among many options and drift to "none of the above". Consumer research suggests a typical member loses interest after perhaps 60 to 90 seconds, having looked at 10 to 20 titles. They also write that what customers ask for (as much choice as possible) and what works (a few compelling choices, simply presented) are very different; they use that point to argue for A/B testing over what customers say (Gomez-Uribe and Hunt, 2015).

The business figures are all the company's own, with no method shown. The recommender system influences choice for about 80 per cent of hours streamed; the other 20 per cent comes from search. Monthly churn is in the low single digits, and over years personalisation and recommendations have cut it by "several percentage points"; the company thinks their combined effect saves it more than $1bn a year (Gomez-Uribe and Hunt, 2015). I use the paper to see how the approach works, not to prove that it pays. See also the Netflix case.

Where does Share a Coke sit?

Coca-Cola's history page, dated 25 September 2014, says the idea was chosen in Australia in 2011: the Coke branding on bottles and cans was swapped for the 150 most popular first names in the country. The company says it sold more than 250 million named bottles and cans that summer, in a country of just under 23 million people, and that the campaign later reached more than 70 countries (The Coca-Cola Company, 2014).

On my reading, set against Arora and colleagues' frame, a named bottle on the shelf falls in between. It is not personalisation: Coca-Cola chose names by national popularity, and the page mentions no data on individual buyers (The Coca-Cola Company, 2014). It is not customisation: the buyer specifies nothing and simply hunts for their own name. The nearest label is what Arora and colleagues call product proliferation, many variants for the customer to pick from, with no one-to-one dialogue (Arora et al., 2008).

The people who ran the campaign name its limit. They knew 150 names would reach only about 42 per cent of the population, even with generic names such as Mum, Dad and Mate, so they shifted the invitation towards giving a Coke to someone else. The parts nearer to customisation came elsewhere: travelling kiosks in shopping centres where people could customise a can, a second release in which the public voted for the next names, and an online shop in Germany for personalised bottles delivered to the door (The Coca-Cola Company, 2014).

The page has no pre-campaign sales, so "more than 250 million" cannot isolate what the printed name contributed. One interviewee also says the campaign was seeded through influencers and that they spent money on television, and that they would probably spend a fraction of that if they did it again (The Coca-Cola Company, 2014). See also the Coca-Cola case. My takeaway: a "personalised" label on a pack hides a design choice. A printed name needs no individual data, but it only reaches people who happen to be on the list.

Which should a small coffee roaster pick?

The five questions below are my synthesis of risks Arora and colleagues raise. The example is hypothetical: the firm and every number are my invention.

A hypothetical example. "Hat Gio Coffee", a roaster in Can Tho selling through its website, has 800 customers who have ordered at least twice. The owner weighs three options. A, personalisation: suggest a bag from past orders. B, customisation: the customer chooses roast (3 levels), grind (3 levels) and blend (4 types). C, hybrid: pre-fill choices from past orders and let the customer edit them.

These rows come from the risks Arora and colleagues name: a wrong guess, choice overload, and customers who design a product and then find they prefer the standard one (Arora et al., 2008). The authors themselves say that when to choose personalisation over customisation still needs more research (Arora et al., 2008).

Now apply it to Hat Gio. Option B multiplies out to 3 × 3 × 4 = 36 combinations. If grinding happens at packing, stock holds only 3 × 4 = 12 kinds of bean, but a screen with 36 choices can still overwhelm a newcomer. Option A on 800 customers who have ordered at least twice is thin data, and Arora and colleagues say a simple model such as RFM (date of last purchase, number of purchases, total spend) may be the best way to go when data are limited (Arora et al., 2008). I lean to C: use RFM to pre-select a bag for returning customers, then let them change one of three options. That is my inference from the hybrid the authors raise (Arora et al., 2008); the post has no measurement showing C beats A or B. It suits a roaster with a few hundred repeat customers and roasting to order. It does not suit one whose customers mostly buy once: with no history to pre-fill, offer three fixed bags instead.

Going deeper: what I read and what I did not

I read Arora and colleagues in full from a PDF a co-author posted publicly, and Gomez-Uribe and Hunt in full from another public PDF. Both are automatic text conversions. I read Coca-Cola's page in full. For Ansari and Mela and for Tam and Ho I read only the abstracts. From Tam and Ho (2006) this post takes one thing: users are receptive to personalised content and find it useful as a decision aid, in a laboratory experiment and a field study (Tam and Ho, 2006).

Three cautions. First, the five stages in the figure are the library's sketch. I do not find them in Arora and colleagues' full text; they name three steps (collecting data, turning it into insight, operationalising it), citing Black and Thomas (2004), which I have not read. Second, the 7.6 per cent from Rossi and colleagues and the 62 per cent rise in expected click-throughs from Ansari and Mela are numbers Arora and colleagues cite second-hand, and 62 per cent is a model's potential, not a realised result (Ansari and Mela, 2003). Third, Arora and colleagues wrote a review in 2008, so its examples are dated.

Netflix and Coca-Cola are both self-reports. The $1bn is a Netflix employee's estimate, and Coca-Cola's 250 million has no baseline. They are enough to describe the practice, not to say which approach wins.