In partnership with

In our previous article, we discussed how creating effective product feedback loops is essential for early-stage startups to iterate quickly and build products that people actually use. 

In the immortal words of Nikita Bier… 

If you were left wondering, “well how do I do that?”, your prayers have been answered. 

Today we're exploring how Liner has taken feedback mechanisms to the point where their LLM now outperforms Chat-GPT…

Liner AI is an AI-powered search engine and research tool that helps users find reliable information quickly, providing AI-driven summaries, source-backed answers, and the ability to highlight and collect information from various sources. 

Liner CEO, Luke Jinu Kim shared with us how his team designed specialized feedback channels that directly feed into their model training, helping them achieve remarkable accuracy rates that outperform even the biggest names in AI.

Luke thinks of this process in two (obvious) segments: collecting feedback, and using it. 

Setting up the traps

If you’ve heard the classic start-up stories on how scrappy start-ups sit alongside their users building their product, 8 times out of 10 it’s probably a B2B business. And that makes sense because you’re actively saving their employees time, or the business money. 

The story is very different for B2C players. Most individual customers bump in and out of new products all the time, and have no incentive to leave you reviews (unless you’ve done something wrong and then they’ll make time to leave you a review). 

As a B2C start-up, your approach to collecting feedback has to be a lot more thoughtful, says Luke. Specifically, focus on two things… 

#1 Feedback collection points

Assuming you’ve used practically any online product or app recently, you’ve probably been asked for feedback. Mostly asking for 5-star reviews or comments on the App Store, which works if you’re Google Maps, but not if you’re at the inception of your product journey, according to Luke.  

Rather than settling for simple thumbs-up/down buttons like Claude or Chat-GPT, Liner implemented several targeted feedback mechanisms:

Not just thumbs up: While Liner does include traditional like/dislike buttons for responses, they've designed these to capture specific feedback about reliability and accuracy rather than general satisfaction. Give your users the option to just click through, they’re not going to spend any time typing these things out. 

Getting specific: Perhaps their most innovative feature, Liner allows users to flag hallucinations for individual sentences within responses, providing granular data about exactly where models are failing. This allows a whole new level of granularity in the feedback they collect. 

#2 Track everything

Tracking user behaviour, whether it’s a click or hover, gives Liner extremely valuable insight into what their users are thinking. 

It also answers the question on how engaged are their users really with the content coming out of their platform. Do users only care about the summarised research, or actually want to look at the source material too? It matters for everything for Liner. 

Source visit tracking: Liner monitors when users click through to reference pages, which serves as an implicit signal about which sources users find valuable enough to verify or explore further. These sources are then prioritised for similar requests going forward. 

Bookmark analytics: The system tracks which specific sources users bookmark, providing clear signals about reference quality and relevance to their target academic and research users.

Content sharing metrics: By analyzing which answers users share with others, Liner gains insight into which responses have the highest perceived value and accuracy.

So I have all this data, now what? 

What makes Liner's approach particularly effective isn't just collecting this feedback—it's how they use it. 

As Luke explained, "All this user feedback goes back to fine-tuning and post-training models and actually makes the models performance better."

Some of this is automatic, and some of it is manual. Luke says they're actively trying to automate the conclusions of the data that they’re tracking into re-training and improving the model. 

This tight integration between user signals and model training creates a continuous improvement cycle that's particularly valuable in the academic and research contexts where Liner focuses. 

As Isaac Newton said, “compounding is the 8th wonder of the world”. (He definitely didn’t say that btw, but anyway) Liner’s in-depth approach to feedback and user behavioural data collection means that their product development compounds at a significantly faster rate compared to their competition. 

Add in the impact of automating that feedback-to-product cycle, and the results are even better. 

What that did for Liner 

The impact of this approach is impressive. While OpenAI's latest models achieve around 90% accuracy and Perplexity claims 93.7% for their Deep Research feature, Liner has reached over 95%—nearing 96% accuracy according to Luke.

This performance gap is particularly meaningful for Liner's target users—university students, graduate students, and researchers—who need reliable information with trustworthy sources. The company has seen strong adoption in academic settings, with Kim noting they've achieved approximately 10% penetration at Berkeley.

What this means for you 

These lessons from Liner can be adapted to any business, not just tech start-ups. 

First, increase the number of feedback touchpoints, reduce the amount of friction a user has when delivering you feedback, and don’t just ask for lame 5-star reviews, ask for more specific feedback based on where in the product journey they are. 

Second, track everything your customers are doing. Once you have these two sets of data, automating  that feedback-product loop will take you to the next level. 

If you enjoyed this issue, please share it with fellow founders, operators, & investors.

See you next week!

Rahul & Aryaman