{{ p.title }}

The first visit is where most personalization gives up. There is no purchase history, no account and often no usable cookie, so the system falls back to bestsellers and the visitor’s first impression of the store is the page everyone else sees. Yet a first session is rarely empty. The page the visitor arrived on, the campaign that brought them, the category they opened and the first two products they compared all say something about what they are looking for. The question is whether the system is built to read it.

Why segments break down

Most recommendation systems describe customers with segments: groups defined by stable attributes such as demographics, lifetime value or past purchases. Segments are useful for planning and reporting. As an input to individual decisions, they have three weaknesses.

They need history, so a new visitor has either no segment or a default one. They are coarse, so a single bucket such as “high-value, 25–34” hides very different intentions. And they change slowly, so yesterday’s segment describes yesterday’s shopper. The same person buying a gift for a parent and buying something for themselves on the same evening is two different shoppers for the purpose of a recommendation, and a segment treats them as one.

Dimensions instead of buckets

We describe a shopper with a set of dimensions scored for the current session, which we call the User Barcode. Each dimension is a property of shopping intent that matters in the category, held as a value and a confidence. In fragrance, the dimensions include scent family, occasion, longevity, season, strength and price range. A shopper who has opened three warm, amber-led scents and filtered for evening wear might read as high warmth with strong confidence and an evening occasion, while longevity is still unknown.

Dimensions differ by category. Apparel needs fit and style, grocery needs dietary preferences and basket context, electronics needs budget and use case. Defining them is domain work done with the merchandising team, and it sets a ceiling on decision quality: a system can only read what its dimensions can express.

In our decision path, the input engine captures the session’s signals and the structure engine turns them into the barcode. The ranking engine then scores products against those dimensions. The barcode describes behavior in the session and carries no name, email address or other personal data.

What a first session carries

Before the first click, a system typically has the landing page, the referring campaign or search, the device, the time of day and a coarse region. None of these identify the visitor. Together they suggest reasonable starting values. A visitor arriving at 9 p.m. in December from a winter gift campaign on a fragrance store is more likely to be shopping for someone else, for an occasion, within a gift-sized budget. The starting barcode reflects that, with low confidence.

Each interaction then updates the values. Opening a product, comparing two, reading the notes, adding to cart or dismissing a suggestion each move specific dimensions. In our experience most of the useful signal arrives within the first few interactions, which is why a well-designed first session improves quickly.

Start wide, then narrow

The first decision for a new visitor should be deliberately varied. Where confidence on a dimension is low, the ranking shows options that span it, so the visitor’s response is informative. Showing three nearly identical warm scents to someone whose preference is unknown wastes the decision. Showing one warm, one fresh and one woody scent, all within the likely budget, teaches the system something whichever one the visitor opens.

As confidence rises, the ranking narrows toward the dimensions the visitor has confirmed. This is the exploration and exploitation trade-off familiar from recommender research, applied within a single session. Too much exploration feels random to the shopper; too little locks in a guess based on one click. The balance is tuned per category and per surface, and whether it works is something the holdout measures directly by comparing first sessions with the current experience.

The reason shown with each decision follows the same rule. Early in a session it names only what the system knows with confidence, such as the occasion, and it adds dimensions as they firm up. A reason that claims more than the system has learned would undermine trust in every reason that follows it.

What the system should not infer

Reading a session closely raises the question of how far inference should go. We limit dimensions to shopping intent in the category: what the person is looking for, for which occasion and at what price. The system does not infer health conditions, ethnicity, religion, sexual orientation or similar characteristics, even where browsing might suggest them, and no such inference can appear in a reason.

Pseudonymity is part of the same design. The barcode is attached to a pseudonymous identifier, and a business that wants to connect sessions to its own customer records holds the key to do so on its side. Behavioral data is stored in the region the customer chooses, the US, the EU or India, and removed within 30 days on request.

Trade-offs

Session-first personalization has limits. Some visits are too short to read: a single page view followed by an exit gives the system little beyond its starting values, and there the decision will be close to a sensible default. For returning customers, ignoring history entirely would waste useful information, so where a business passes its own consented signals, they can inform the starting values. And every decision depends on the dimensions, which means a category with poorly chosen dimensions will produce confident decisions about the wrong things.

None of this replaces measurement. Whether reading the session improves first visits is an empirical question, answered by comparing first-session behavior against the holdout on each surface.

The Decision Inspector walks through a worked example decision and shows each input with where it came from.

See the inputs behind a decision