Intent

A few weeks ago, my post was about how to personalize websites well, and whether that’ll become more the norm as AI continues to develop. One of our customers responded and asked: “How do you accurately assume the intention of a customer EVERY time?” Presumably this would be necessary to personalize well.

I’ve been slow to respond (sorry, Mike), and as I started to I realized it’d be a pretty good newsletter topic, so here we are :) 

To start: I’m not sure accurately predicting the customers’ intent matters. At least not every time.

Sort of hot-takey, I know, so let’s start with the context of what I discussed in the newsletter that prompted the question: 

The assumption inside Mike’s question is that the ability to predict intent is a prerequisite for this rapid store reconfiguration. That, to me, actually poses two questions: 

• How do we gather intent data and where does that data come from? 

• Do we actually need that data to do “better”  personalization?

The first is a more tactical question of putting intent data into play and the second is a more philosophical question about its role and, more broadly, the role of brand in terms of creating intent. 

On the tactical side, there are intent signals scattered everywhere. But some are stronger signals than others;some are easier to access; some are immediately available vs. lagging. Obviously, the more of them you can collect, the better. Across our customer base, the most popular signals (and some of these you might debate are not signals of intent, but more context clues) we see used are:

• Where someone came from—channel, source, campaign, etc. You generally want the site experience to “rhyme” with what they just clicked through from 

• The first few actions on the site—what collection was viewed first? Which products were added to cart? 

• Information gathered through the welcome discount pop-up—increasingly, you’re seeing stores ask a question or two before handing out that sweet 10% off code

• Information about prior purchases / other first party data—this is incredibly valuable but can be hard to operationalize. You need to get customers to log in or have a clever UTM strategy in text and email campaigns where you can “id” different segments. I anticipate a big push over the next few years to get more users logged in as they browse

The basic answer is: use what you have, when it makes sense. Personalization isn’t binary where you turn it ON/OFF for everyone. The combination of signals you have to predict intent are going to vary widely by session and by visitor—you’ll always have a “baseline” version of your site and then can personalize / customize from there, depending on how good a prediction you have. 

You may find, for example, that a particular signal allows for you to create a pretty narrow funnel that increases conversion rates, because that signal is highly predictive of a customer who only wants to buy one, specific thing. You may also find, for example, that another signal is highly predictive of intent to purchase (e.g., this person came from organic google search), but has far less fidelity when it comes to specific products.

How you—or, in the future, AI tools—choose to act on that information is pretty interesting.There will need to be (a) a reliable approach to gather first party data (strong UTM structuring; logging in users; post purchase surveys; UX that encourages users to declare their intent), and (b) a scaffolded approach for how personalized to make each journey. 

All of this relies on a core assumption—that personalization is about predicting what a customer wants and then serving that to them. Divination. But back to my hot take—I don’t think accurately predicting the customer’s “true intent” actually is a requirement for building a great personalization program. 

If our goal is to predict intent, that implies a brand’s job is to serve the customer what they want. While a brand does need to do that, a good brand’s job starts before that: It convinces a customer of what they want. I.e., A brand’s job is to influence intent.

When we think of how brands influence intent today, that usually happens through some combination of offsite marketing (advertising, influencers, email, etc) and onsite merchandising (be it the traditional product recommendation blocks, or the in-cart upsells, or the visual story being told through the images and home page).

Offsite marketing leans more into influencing intent, to be sure, and onsite merchandising leans more into capturing the intent that was created. You could even argue that onsite merchandising might be best viewed as “last-mile” optimizations: I got a customer interested enough in my content to visit my site; how do I get them to take one last step and get them to take the risk to give my product a try?

Those onsite efforts today are made in the broadest possible contexts. They rarely, if ever, consider trying to predict intent—yet they’re still typically valuable.

If you look at it this way, you don’t really need perfect accuracy in predicting what someone’s intent is to benefit from personalization. You can use the context you have to make a best guess, and try to direct them that way (the “scaffolded” approach mentioned above). And if you’re wrong, then what is the risk? Perhaps some customers will drop off, but a well-designed site experience will give them off-ramps so they can go find what they were actually looking for instead. 

Think about how brick and mortar stores are laid out. The Walgreens in my residential neighborhood has the toiletry and pharmacy sections up front, serving a classic drug store purpose. Near my office in FiDi, they have all the grab-and-go food and drink options up front because most traffic is coming in on lunch break. If I need to pick up allergy medicine there, I’ll still walk to the back and find it. This store has a “personalization” that is based on a broad assumption/context (people in FiDi are looking for snacks) and likely works, and I find what I’m looking for nonetheless. 

In that situation, the store can be completely ignorant of my intent, and still use personalization well based on context clues and assumptions. So perhaps that future—where each customer gets a completely generative experience—doesn’t actually need to worry about predicting intents all that accurately.

Obviously in ecommerce, you wouldn’t use the neighborhood as context—but there’s lots of similarly easily accessible data: Which consumer segment does the customer belong to? What content has that person interacted with? Which products does the customer know exist?

These are relatively easy to know and to act on. And make it feel way lower-stakes to get started. You don’t need to have every visitor figured out all at once.  And they’re far more likely to have clues around what other actions from the brand are creating influence that, when acted on, can push interest into intent. And it feels like that might be where the unlock is.