Designed an adaptive
search & recipe discovery chatbot
Good Food’s subscriber expectations for recipe discovery have changed.
I designed Cook Companion, an AI discovery chatbot, and built a working prototype of adaptive search, partnering with product, engineering and editorial.
Problem
My research highlighted a change in user habits
78% of customers say they use AI to find and discover recipes with 62% explicitly using AI chatbots — suggesting that new discovery behaviours were forming outside recipe platforms including Good Food.
A deeper problem
Chatbots aren’t just helping customers discover recipes; they’re supporting customers across the entire cooking journey. And only 23–28% say they cross-check AI suggestions against a recipe platform.
Before cooking: 78–83% use AI for recipe discovery and deciding what to cook.
During cooking: 55% use AI for problem-solving and recipe adaptation.
Key threats
How customers discover using chatbots
Most prompts included constraints, often ingredients, but varied considerably in intent and uncertainty.
Customers say chatbots often narrow their options too quickly and make assumptions about their needs. This raised the question of whether chatbots are the best response for every discovery prompt.
Hypothesis
Search remains an important content-discovery tool, and chatbots aren't always the best response to a discovery query.
Validating using real customer prompts
Using the exact customer prompts captured in the earlier research, I ran two studies: a text-only card sort followed by a study showing prototypes of all three responses — to understand whether customers expected Search, chatbot conversation or a chatbot AI summary — to best support each query.
Insights
The majority expect an adaptive experience while only 25–28% consistently preferred traditional Search.
01
Chatbots are largely expected when prompts are question based or open-ended
02
Constraint led queries favour a traditional search experience
03
When clear constraints are paired with uncertainty, it produces a more nuanced picture.
04
AI summary was nearly always expected for problem solving prompts
Opportunity
Enhancing search, not replacing it.
Alongside traditional search, we've designed a chatbot experience to support customers who need guidance both with building their search intent and evaluating options to make a decision.
This is a real opportunity for Good Food to provide an experience customers expect and value, which is adaptive and not a one-size-fits-all approach.
Technical
feasibility
Is query routing possible and can it reliably meet our needs?
I built a search router that isolates word classifications, identifies and scores intent signals which determine the likelihood of a follow-up question. Together these values determine which type of experience is best for the customer.
The logic behind the router is configurable, so as we progress this can be iterated and refined.
Solution
One entry point routes users towards the experience that best supports their mission. Queries signalling a need for guided support are routed to Cook Companion chatbot.
Conversation discovery
Conversational experiences aren’t for everyone, so we’re selective about which queries are routed to Cook Companion, while always giving users the freedom and control to return to Search.
Traditional search
Many queries point towards the existing search flow, but with a small but significant improvement.
Customers often struggle when deciding which recipe to choose from a long list. We've introduced a “Help me decide” button that gives users access directly to Cook Companion at the point where support is needed, helping narrow the results into a smaller set of relevant recommendations.
Chatbot
Working closely with engineers, we designed conversation patterns to guide the chatbot as it builds intent and narrows options, helping us create a consistent and reliable conversational system that works across different levels of discovery intent.
Roll out
During testing, we’ve seen high token usage. We’re focused on reducing the overall cost of supporting this feature while delivering the right customer experience.
Our goal for launch is for 80% of sessions to result in iterative success.
