However, there are a couple of problems with it. It's very expensive and very heavy because it captures all DOM changes, so it slows down the experience of the end user. For these reasons you are forced to do heavy sampling, for example record only 10% of the sessions. Apart from that, you need to spend a lot of time watching session replays, and sometimes you cannot find the session you are interested in as it might not be included in the recorded sample.
This was basically the trigger for the idea. I started working on building my own tool. The idea was to make it as simple and lightweight as possible, and focus on clarity rather than adding a lot of bloat.
Also worth mentioning, I have noticed that FullStory and other tools recently started adding AI session summary features, which means there is clearly a need. However, they cannot really do visitor-level summaries across sessions or generate broader insights, because due to the heavy script overhead you end up sampling only a small percentage of traffic. Technically you could record all sessions, but it would be extremely expensive and would slow down the experience for everyone. The difference with my approach is that because the tracker is lightweight, you can record all sessions and still generate visitor summaries and insights across the full dataset, at a much lower cost.
So the idea is simple: capture signals from user browsing, such as pageviews, clicks, scrolling, and metadata from the request headers. Record all these events in a session. Then schedule a job every X minutes to send the events to an LLM and get a summary of the session. Then move a level up and generate a visitor profile summary across all sessions of that visitor. Basically it's like a living memory layer of your customers. Finally, use the session summaries and visitor profile summaries as input and let the LLM identify patterns and generate insights, things you might not know about how your visitors behave, things you can act upon.
The first decision was what to use for recording customer behavior. Initially I tried to use rrweb, but I noticed that it slows things down a lot as well. So I ended up building my own tracker that captures just enough signals for generating a summary, but at the same time keeps it super lightweight. For example, on each pageview I record a snapshot of the DOM to give context to the LLM about what the visitor is browsing, but I'm not recording all DOM changes like FullStory.
The next thing was how to generate visitor summaries and insights. Should I send all raw events or just session summaries and visitor summaries? I ended up sending only the summaries as I was happy with the output, and this way I keep it much more cost effective in terms of LLM input tokens.
So far I've tested the tool myself by injecting the script with a Chrome extension and browsing around various websites. The summaries I see so far make a lot of sense. Now I'm looking to test it in the wild with data from real visitors, so I'm looking for people who are happy to embed it in their website and see how valuable the summaries and insights are. Of course I will collect feedback and improve the tool.
On the landing page you can click on Explore Live Demo and you will get an idea of how the tool works.
Thanks.