MCP Cold Start: Recommendations Before the Data Arrives

A Marketing Cloud Personalization recommendation engine learns from behavior. View events, purchase events, cart events, search events. With enough volume across enough visitors over enough time, the platform develops a model of who likes what and what to surface. The trouble is that on day one, day seven, and often month two, the engine has none of those signals. The cold start problem is what separates implementations that look intelligent at launch from implementations that look broken until they have run long enough to seem intelligent retroactively.

Three cold-start scenarios show up on real engagements. Each requires a different fallback strategy.

Scenario 1: Cold catalog, warm site

An existing high-traffic site rolls out MCP for the first time. The catalog has thousands of items. Visitor traffic is healthy. But the platform has zero behavioral data because day one is day one.

The fallback that works:

• Bestsellers and Trending recipes seeded from prior data. If the e-commerce platform has historical purchase data, push the top sellers list as a starting bias for the Bestsellers recipe. The platform begins with informed defaults instead of empty buckets.
• Content-based similarity for item-detail strips. Similar Items uses item attributes (description, category, price tier, brand) to find matches. It does not need behavioral data. It produces meaningful recommendations from the moment the catalog is uploaded.
• Categorical placement for category pages. Visitor lands on a Beauty category. The strip uses Categorical recipe scoped to Beauty. Fallback to bestsellers in that category. No cross-category drift, no dependency on training data.

What does not work in this scenario: any recipe that requires affinity profiles to be developed (Personalized, View Also Viewed at scale). Defer those for the first four to eight weeks while traffic builds the data.

Scenario 2: Cold visitor, warm catalog

The platform has been live for months. The catalog is rich. Affinity models have learned. But this specific visitor is new. Either it is a first-time visitor or a returning visitor on a fresh cookie.

The fallback that works for the first session:

• Trending replaces Personalized for cold visitors. The visitor has no affinity. Showing them what is trending across the population is the best signal available. Once they have engaged with two or three items in the session, switch to affinity-driven recipes.
• Categorical relevance over personalization. Cold visitor lands on a category page. Show items popular in that category, not items personalized to a profile that does not exist yet.
• Item context over visitor context. On item-detail pages, recommendations should anchor on the item being viewed. Similar Items, Frequently Bought Together. Both work without visitor history because the anchor item provides the context.

The transition from cold to warm should be smooth. The recipe sequencing pattern handles this: try Personalized, fall back to Trending if no affinity exists. The visitor sees relevant content from the first page and progressively more personalized content as they engage.

Scenario 3: Cold both

Brand new site, brand new catalog, brand new audience. The hardest scenario, and the one that gets cited as evidence that "personalization does not work for early-stage businesses."

The cold-both case requires two acceptances. First, that the personalization layer is not going to add measurable lift in the first quarter. The platform is collecting baseline data. Second, that the implementation work done during the cold-both phase pays off in months three through twelve as data accumulates.

Pragmatic actions during cold-both:

• Configure the system. Validate sitemap. Stand up campaigns. All of the integration work that does not require traffic to validate. Get the catalog clean. Get the data layer firing correctly.
• Use content-based recipes everywhere. Similar Items, Categorical, Bestsellers seeded from any historical data the team can produce.
• Set realistic stakeholder expectations. Tell leadership that the system is in data-gathering mode for the first eight weeks. The case studies in the marketing material assumed warm-catalog conditions.
• Plan a data-richness milestone. Define explicitly when the system has enough data to switch on affinity-driven recipes. Tie it to a metric like "1000 unique visitors per category per week" or "500 weekly purchases per top-100 item."

Cold start for new categories on warm sites

A subtler cold start: a mature site adds a new product line. Old categories have years of data. The new category has none. Recipes that span the catalog produce odd outputs for visitors browsing the new category because the new items have no behavioral history.

The fix is category-scoped recipe sequencing. The recipe sequence on a new-category page tries Categorical first, scoped to the new category. If that returns weak signals (because the catalog is new), the next fallback is content-based Similar Items anchored on whatever the visitor has shown interest in elsewhere on the site. Last fallback is sitewide Bestsellers.

Done well, the new category looks active and curated even on day one. Done poorly, the category page recommendations show the same three new items in different orders for a month, and visitors lose interest.

Manual curation as a cold-start lever

Custom recipes accept a manually defined item list. During cold start, a hand-curated recommendation set often outperforms the platform's defaults. Three to five hundred well-chosen items, organized into themed lists (best for new visitors, best gift items, seasonal staples), give the merchandising team direct control while the platform learns.

The discipline is to set a sunset date on manual curation. Three months in, evaluate whether the manual list is still beating the algorithmic option. If yes, keep curating. If no, switch to the recipe and reclaim the merchandising time.

What success looks like during cold start

Realistic expectations for a cold-start phase:

• Recommendation strips look populated and relevant from day one (because of fallback design, not because the platform has learned).
• Click-through on recommendation strips runs slightly above non-personalized control (the lift is small but positive).
• Affinity profiles are developing, visible in the platform's analytics dashboards but not yet driving revenue.
• The eight-to-twelve-week mark shows the inflection. Affinity-driven recipes start outperforming fallbacks. Lift becomes measurable.

The cold-start phase is unavoidable and survivable with the right fallback design. The teams that struggle with MCP launches are usually the ones that turned on affinity-driven recipes from day one and watched the strip serve generic results for two months while the recommendation system was blamed. Sapota's Salesforce team designs the fallback chain explicitly during sitemap and recipe configuration on every new Personalization engagement, because it is the difference between a launch that builds momentum and a launch that loses stakeholder confidence in the first month.

---

Launching Marketing Cloud Personalization or restarting after a quiet period? Sapota's Salesforce team designs cold-start strategies and fallback sequencing on production engagements. Get in touch ->

See our full platform services for the stack we cover.