"We aren’t ChatGPT." This simple realization led TheFork to scrap its AI chatbot, replacing the conversational interface with a high-performance "single-shot" search tool. The gamble paid off: by prioritizing direct results over small talk, the team has gathered encouraging positive signals regarding engagement and conversion for this new feature. Here is a look behind the scenes of a strategic pivot where user experience and business efficiency won out over the hype.
They say "curiosity killed the cat." At TheFork, it wasn’t curiosity, but a calculated strategic choice that pushed the team to "kill the chat." We aren't talking about a feline, of course, but a GenAI-powered chatbot that never saw the light of day because the team decided to change tack.
When people talk about GenAI, one image immediately comes to mind: the interface of ChatGPT, Gemini, or Mistral’s "Le Chat." Their common denominator? A conversational window - a chatbot that lets you talk freely with a Large Language Model (LLM).
The decision by the restaurant reservation giant to ditch this format is surprising. Why not just follow a recipe that has already proven successful elsewhere? Let’s look at a pivot that goes against the grain, proving that in the world of AI, you sometimes have to sacrifice the spectacular for the useful.
Putting AI on the menu
The story begins in May 2024 during TheFork’s annual hackathon, a week where teams prototype innovative projects to present to the company. Naturally, AI was on everyone’s lips, with one goal: put this unavoidable technology to work for the business. After seven intense days, a prototype for an AI-powered restaurant search took the top prize: AskTheFork.
The feature checked every box. It was technologically fresh and solved a long-standing "pain point." Previously, search at TheFork relied on a manual tagging system. Every tag for every restaurant was entered by hand by sales reps or the restaurateurs themselves.
According to Laura Sohier, Group PM in charge of Restaurant Discovery, this was far from ideal. A system where the manual nature of the process often led to inaccuracies, like a restaurant being tagged "vegetarian" simply because they served a side of fries.
This is where AskTheFork - and AI - stepped in. The idea was to use the interpretive power of LLMs to scan everything from menus to User Generated Content (UGC) to find the perfect match for a user's intent.
"Until now, it was impossible to search for the 'best truffle pasta'; you had to pick 'pasta' or 'truffles' and then dig through menus one by one," Laura explains. "Now, we can not only combine tags but search for abstract concepts like a 'romantic dinner.' AskTheFork combines photos and reviews to tell the user, based on AI scoring, exactly why a restaurant is the perfect fit for their date night."
While the feature’s technical potential was a given, shipping it to users was a different story. A chatbot seemed like the obvious frontrunner, but the team soon hit a fundamental snag: is a conversation really the most efficient way to book a table?
Stop gambling on AI. Download our AI Product Canvas, a comprehensive guide designed to help you tackle the critical questions before launching your next AI initiative.
“We Aren’t ChatGPT”
Initially, the "Forkies" (as employees are known internally) assumed a chatbot was the way to go. For Alberto Conti, Head of Design, it seemed like the natural choice: "Users are becoming more comfortable with GenAI. Our first instinct was to offer an experience close to what they already knew." The first version was a conversational interface with a UX very similar to ChatGPT to encourage adoption.
But just as you shouldn't put the cart before the horse, you shouldn't put the UI before the UX. Through user testing, the team had an epiphany. "The further we went, the less we saw the feature as a personal assistant chatting with users to figure out what they wanted for dinner. At its core, TheFork’s value is helping people book the right restaurant. We aren’t ChatGPT; let’s leave the conversation to them."
This shift was a win-win: it better served the user and simplified the technical side for Group Product Manager Quentin Ménard’s team. "The iterative nature of a conversation, saying you want a Margherita pizza and then changing your mind to tapas, is incredibly hard to build. You move from simple LLM workflows to 'agentic' logic, which requires significantly more development effort," he explains.
The result? A single-shot search where the user gets a list of highly relevant results instantly. "We’ve shown we can handle almost any query. Of course, there are guardrails; don't look for clothes on AskTheFork, you won't find any!" Quentin jokes. "But 95% of searches yield results. Where we used to only satisfy simple queries, we can now answer complex, multi-criteria human needs."
Of course, it’s not perfect. LLMs are fickle. Capucine Tiberghien, Product Manager, notes: "A huge part of the project is prompt engineering. For example, because AI works on semantic proximity, negation is hard. If you say you want 'anything but sushi,' the AI will probably suggest... sushi. But AI is a process of eternal refinement. You have to focus on what has the most impact; you can't spend all your bandwidth fixing prompts for edge cases that only affect a tiny fraction of users. In this case, tests showed the problem only affected a minimal number of users and AskTheFork was perfectly viable as is."
Opening the AI "black box"
Functional performance is one thing, but how do you package it for the user? The first step was protecting the core business. "Keeping AskTheFork as a standalone experience ensures we don't jeopardize the classic search," says Alberto Conti. "It’s the right way to handle experimental features: build trust before fully integrating it into the ecosystem. This allows us to observe, measure impact, refine, and then decide whether to continue investing or not."
Once they ensured they weren't putting all their eggs in one basket, they had to tackle the "black box" problem. LLMs are non-deterministic; the same query might not yield the exact same result twice. This can erode user trust.
Simran Singh, a consultant from Thiga and an AI Design specialist, explains: "Because GenAI is probabilistic, it won't always behave how you want. It’s vital to guide the user and explain why they are seeing a result. During MVP testing, every user asked: 'Why was this suggested to me?'"
The team listened. Now, results display a specific photo or a review snippet that matches the query. If you search for the "best pizza in Paris," you’ll see a photo of a Neapolitan Margherita or a customer review raving about the pizzaiolo.
For Simran, this isn't just a "nice-to-have." "Design is often underestimated in AI products. We see AI PMs everywhere, but AI Product Designers are rare. But design is what drives adoption. It’s not just about the interface; it’s about explainability and integration. The best AI feature in the world won’t be used if the interface isn't right."
The secret sauce
The numbers speak for themselves. In any case, those obtained by TheFork's teams are telling. With a conversion rate higher (5% more) than the overall app average, AskTheFork proved its results are more relevant to users - a true victory for the stakeholders.
There remains one last point of vigilance: an adoption rate that is still limited. This doesn't delight Laura Sohier, but it doesn't surprise her either. "The feature has just been launched and it introduces a fundamental change in how people search for a restaurant. You don't transform such deeply ingrained user behavior overnight," the Group PM reminds us. Especially since classic search remains the natural entry point: "Users don't open TheFork to test an AI feature, but primarily to find a restaurant."
However, the essential point is this: "The functionality has fully fulfilled its role by demonstrating the potential of artificial intelligence in our ecosystem." The next milestone? "Merging the best of both worlds. Our goal for this year is to unify the experiences to make them more coherent and answer a single need: helping every user find the restaurant that suits them."
What will this new search bar look like? Will it be covered in purple stars and "AI" branding? Alberto Conti is skeptical. "AI is trendy, and many companies use it as a marketing gimmick without asking what it actually does for the user. But users don't care if there's AI under the hood as long as the result is good. I don't really believe in 'AI branding'; it should be a technology that runs in the background."
In short: a secret ingredient. In the meantime, if this article has made you hungry and you’re not sure where to go... you know what to do.
Great tech is only half the battle. How does your AI actually land with your customers? Download our framework to evaluate the AI user experience and turn raw technical data into real-world business KPIs.
.png)
