Beyond Search: How Recipe Recommendation Algorithms Could Transform Food Discovery

Most food platforms focus on helping users find what they're already looking for. Type "chicken curry" into a search box, get a list of chicken curry recipes. But CKBK founder Matthew Cockerill is building something more ambitious: a system that helps users discover what they should cook next, even when they don't know what that is.

On the Levels Podcast, Matthew shared his vision for algorithmic food discovery that goes far beyond traditional search. His approach offers insights for any platform looking to move from reactive search to proactive recommendation systems.

The Current State of Food Discovery

Most recipe platforms operate like early Google - they're great if you know what you want, but limited when it comes to inspiration and discovery. Users get stuck in cooking ruts, repeatedly searching for the same familiar dishes.

"You will otherwise have the choice right now, if you use Google, you'll find some good content, but it's all in different formats. Some's buried in a blog here, some's behind a paywall there, some's ad supported over there. And that makes the experience of finding what your inspiration and then managing the recipes which you have cooked and what you want to cook next, all just to be a bit chaotic."

CKBK is building the technical infrastructure to solve this discovery problem through personalized recommendations based on user behavior, preferences, and cooking patterns.

Learning from Music and Video

Matthew frequently references Spotify's recommendation engine as the gold standard for content discovery. The parallel makes sense: both platforms need to balance user preferences with serendipitous discovery.

"Just like Spotify's home pages drives a vast amount of the listening is by just people saying, yeah, I'll listen to that."

But food recommendations present unique challenges that music platforms don't face. Seasonal ingredients, dietary restrictions, cooking skill levels, and available kitchen equipment all factor into whether a recipe recommendation will be useful or frustrating.

The Graph Database Foundation

CKBK's recommendation system is built on a graph database that captures relationships between recipes, ingredients, techniques, and authors across their entire cookbook library.

"We have all the content in a graph database, which has a whole bunch of algorithmic technologies which can be used."

This approach enables sophisticated connections that simple keyword matching can't achieve. The system can identify that two recipes are related not just because they share ingredients, but because they use similar cooking techniques or come from the same culinary tradition.

Personalization Without Overwhelm

The challenge with algorithmic recommendations is avoiding the "filter bubble" effect where users never discover anything outside their established preferences. Matthew has observed this problem firsthand with other platforms.

"I remember at some sites which are very driven by that sort of thing and you signed up and you said this and then you access one recipe and then it was very difficult to get it to escape from the fact that it just seemed, it thought I was obsessed with beans and I wanted to cook everything with 20 different types of beans."

CKBK's approach balances personalization with exploration, ensuring users don't get trapped in narrow recommendation loops.

The Recipe Roulette Concept

One feature Matthew is considering is a "roulette" button that provides truly random but intelligent suggestions from high-quality content.

"Some people have asked us, as well as recipe of the day, could there be a button, essentially a roulette button, and we could use whatever we wanted to kind of make it be a roulette from a interesting set, but they just like to be able to get something interesting and different."

This addresses a common user need: the desire for inspiration when they're tired of their usual cooking routine but don't know what specific alternative they want.

Context-Aware Recommendations

CKBK's system considers multiple contextual factors when making recommendations. Time of day, day of week, seasonal availability, and user location all influence what gets suggested.

"Different people, some people will do most of their cooking at the weekends, some people will do most of their cooking on weeknights. And so when you send notifications and what type of suggestions you give people, as you say, depending on if..."

This contextual awareness means the same user might receive different recommendations at 3 PM on a Sunday (when they have time for elaborate weekend cooking) versus 6 PM on a Tuesday (when they need something quick for dinner).

Cross-Book Discovery

One of CKBK's unique advantages is their ability to recommend across different cookbooks and culinary traditions. Traditional cookbook usage keeps readers within a single book, but the platform can surface connections across their entire library.

"For every recipe, there's a set of related recipes, which is like, okay, you found this one. But basically, across many different books, we've got some other things which combine many of the same ingredients or something like that."

This cross-pollination enables discovery that would be impossible with physical cookbooks or single-author websites.

The Amazon Choice Model

Matthew draws inspiration from Amazon's approach to handling overwhelming choice through curation signals.

"I always think it's interesting what Amazon does as well, which is that often you search for things and yeah, well, there are 20 different versions of this product, which what's a good simple option? And there's this sort of Amazon's choice or this is the most popular or whatever."

CKBK implements similar "safe choice" indicators based on popularity, user ratings, and editorial curation to help users navigate their extensive recipe database.

Behavioral Learning Systems

The platform tracks user behavior to improve recommendations over time. Recipes that users save, print, or spend extended time viewing get weighted more heavily in future suggestions.

"We do have, when you go through a recipe, someone was just asking me about this and I was like, we actually already have that. And they're like, I just never quite scrolled to the end of the recipe to realize it."

This behavioral data helps the system understand not just what users say they like, but what they actually cook and engage with in practice.

The Timing Challenge

Food recommendations need to account for temporal patterns in ways that entertainment content doesn't. People have different cooking patterns on weekdays versus weekends, different seasonal preferences, and different meal timing needs.

"Apple put this into app recommendations years ago so that if you tend to use Duolingo at 11 at night, it will tend to make that the recommended app when you go there."

CKBK is building similar timing intelligence to surface dinner recommendations at appropriate evening hours and weekend project recipes when users have more time available.

Future Vision: Proactive Discovery

Matthew's long-term vision goes beyond reactive recommendations to proactive discovery. Instead of waiting for users to open the app looking for ideas, the system would surface perfectly-timed suggestions through push notifications and widgets.

The goal is creating what he calls "serendipity" - those moments when the platform suggests exactly what you didn't know you wanted to cook, at exactly the right moment.

Technical Implementation Challenges

Building effective recommendation systems requires significant data science resources - something most early-stage startups lack. Matthew acknowledges this challenge while building toward the sophisticated systems used by Netflix and Spotify.

"Obviously with the sort of resources that have gone into Spotify and Netflix, that's kind of at the heart of what they do is algorithmic recommendation."

The key is starting simple and building complexity over time as user data and technical resources grow.

Key Takeaways

• Graph databases enable sophisticated connections: Capturing relationships between recipes, ingredients, and techniques unlocks advanced recommendation possibilities
• Context matters more than preferences: Time, season, and cooking patterns often predict user needs better than stated preferences
• Avoid recommendation tunnel vision: Systems must balance personalization with discovery to prevent users from getting trapped in narrow content loops
• Cross-content discovery creates unique value: Platforms with diverse content libraries can surface connections impossible with single-source materials
• Behavioral data trumps survey data: What users actually cook and engage with reveals more than what they say they like

Matthew's approach to algorithmic food discovery demonstrates how recommendation systems can evolve beyond simple similarity matching to create genuinely helpful cooking inspiration. The challenge lies in building systems sophisticated enough to understand the complex factors that make a recipe recommendation truly useful.

Listen to the full conversation with Matthew Cockerill on the Levels Podcast to dive deeper into his vision for the future of food discovery platforms.