Useful on visit one
No history required, because the ranking reads the session rather than a profile. The cold-start case is the normal case here rather than the degraded one.
Recommendations
Recommendations that work on a first visit.
Most engines need history before they say anything useful, so a new visitor gets your bestsellers. This reads the visit that is happening and ranks against it, which means the first session is served as well as the hundredth.
A first-time visitor, no history, no accountWhat can a recommendation engine say about this person?
Signals available before the first recommendation
0Purchase history7Signals this visit
Referrer, landing product, search terms, filters touched, products opened, time on each, and what was skipped. None of it needs a profile.
Useful on visit one, which is the visit most engines sit out.
Where you are
Your recommendation engine is quietest exactly when it matters most.
Collaborative filtering needs a history to filter against. On a visitor with no history it has nothing, so it falls back to what sold well last week, which is the same strip every anonymous visitor sees. That is most of your traffic on most storefronts.
The usual fix is to shorten the cold-start window by collecting more: a login prompt, a cookie, a profile stitched across devices. Each of those adds a consent surface, a regulatory question and an engineering commitment, and none of them help the person who is on your site right now for the first and possibly only time.
The information you need is already in front of you. Somebody who has opened three products under a price point and filtered twice has told you a great deal, and none of it required knowing who they are.
A recommendation engine that needs to know who somebody is cannot help on the visit where it would matter most.
The cold start, stated plainly
Most engines are quietest on the visit that decides whether there is a second one.
Two approaches to the same first-time visitor, and only one of them has anything to work with.
History-based engines
Need to know who somebody is before they help.
Collaborative filtering learns from what people like you bought. On a first visit there is no you to be like, so the fallback is a bestseller strip, which is the same strip everybody sees.
Silent exactly when the visitor is deciding whether to stay.
Session-based ranking
Reads the visit that is actually happening.
What was searched, filtered, opened and skipped in the last ninety seconds narrows what this person is deciding between, whether or not they have ever been here before.
No identity required, and nothing to warm up.
What it looks like
The same ranker, pointed at a slot.
Not a second model. The four factors that order a search result order a recommendation strip.
Base relevance34%
Persona fit31%
Brand affinity21%
Price band14%
No history, no cookie, no login. Ranked from this visit alone.
What happens
How a strip gets ordered
The same ranker as search, pointed at a slot rather than a result set.
- 01Signals
The visit emits
Views, searches, filters and cart actions in this session. Each one moves persona weight and attribute affinity. Nothing is stored against a person, because there is no person to store it against.
- 02Rank
The catalogue is ordered
The slot asks for n products and gets them ranked on base relevance, persona fit, attribute affinity and price band, the same four factors search uses.
- 03Record
Exposure is logged
What was shown is recorded against the decision id, so the strip can be evaluated later rather than assumed to have worked.
What you get
What you get
Any slot on any surface
Product detail, cart, category, home. A slot asks for products and gets them ordered; nothing about the placement is special-cased in the engine.
One ranker, not two
The strip and the search results are ordered by the same four factors. A merchandiser who understands one understands the other, and there is no second model to keep in sync.
Falls back to your own strip
If the engine is slow or unreachable the surface renders whatever you had. Degrading quietly is the default path, not an error state in front of a live store.
Before you commit
What recommendations will and will not do
What this doesRun with no cookie, no login and no cross-site identifier
Hold a share of traffic back on every surface, so a strip is evaluated rather than assumed
Start returning a useful order on the first page view, with no warm-up and no training window
Switch off per surface, without touching anything else on your storefront
What it does not doQuote you an uplift. The held-back arm is what produces that number, and it has not run on your traffic yet
Build a long-term profile of a shopper. It reads the session and forgets it, which is a design choice and also a limit
Recommend on an attribute your catalogue never recorded
Beat a well-tuned collaborative filter on a logged-in repeat customer with years of history. Different problem, and we will say so
Questions
The ones that decide it.
- How is this different from collaborative filtering?
- Collaborative filtering answers what people like this person bought, which needs a history and an identity. This answers what this visit implies, which needs neither. On a returning logged-in customer with years of history, a good collaborative filter has information we do not. On an anonymous first visit, it has none and we still work.
- Does it need a cookie or a login?
- No. It infers from views, searches, filters and cart actions inside the current session. Nothing is written against a person, there is no cross-site identifier, and there is no profile to export or delete because none is kept.
- Where can strips be placed?
- Any slot on any surface: product detail, cart, category and home are the common ones. The engine is asked for n ranked products and does not know or care what the placement is called.
- Is this a different model from your search ranking?
- No, and that is deliberate. It is the same ranker with the same four factors pointed at a slot instead of a result set. One thing to understand, one thing to audit, and no second model quietly drifting away from the first.
- What happens if it is unavailable?
- The slot renders the strip you had before it, and nothing else on the page changes. A recommendation surface that can fail visibly is worse than one that was never installed.
Make the shelf decidable, then recommend from it.
A strip can only order products your catalogue can tell apart. The audit answers that before anything is turned on.
Request a Decision AuditPriced after a pilot, against what the held-back arm actually shows.