Trust is Harder Than I Thought
I have been working on a new feature for Meeako. More on that in the next post.
While testing it, I noticed Meeako AI did relatively well when answering questions about apps in its library. But when I asked about apps or data not in the library, the responses weren't grounded.
That seemed straightforward to fix. It wasn't.
Meeako's domain (apps, user flows, product strategies) is well documented on the internet. Most large language models have been trained on it. So when I mention a well known app by name, even with instructions to stay grounded in Meeako's library, the model slips into reciting what it already knows and presents it as if it came from Meeako.
Here's what that looks like.
What metrics does Duolingo track for its onboarding?
I asked about Duolingo's onboarding metrics. Meeako walked me through "initial signup, language selection, placement/level assessment, and the first lesson experience." Confident. Detailed. None of it came from Meeako's database.
What are Instagram's key activation metrics?
I asked about Instagram's key activation metrics. Instagram isn't in Meeako's library, so Meeako correctly said it couldn't analyze it. But then it listed "feed-based follower models" and "engagement loops around discovery or creation." Generic PM knowledge. Not grounded in anything Meeako actually knows.
What metrics matter for Uber's ride-booking flow?
The sharpest one. I asked about Uber's ride-booking flow. Meeako said it needed to "narrow down which flow in the library we're looking at", implying it was going to check. Then immediately described Uber's flow from memory. "Initial search/request, driver matching and acceptance, the wait for pickup, and the ride itself."
It didn't check. It just answered.
What are the best metrics to track for a paywall?
Not all bad. When I asked "What are the best metrics to track for a paywall?". No specific app named. Meeako stayed in its lane. It said it couldn't help instrument a paywall but could show how paywalls are built in apps already in the library. Then asked what type of app I had in mind.
Why this matters
In post Meeako Needs Guardrails, I wrote that Trust is one of Meeako's four guiding principles. Every claim should be grounded. Honest even when data is thin.
Right now that principle isn't holding. The failure isn't obvious. The answers sound right. They're confident. They're formatted well. You'd have to already know the data to catch that it wasn't sourced from Meeako.