You paste your metrics into ChatGPT, ask for analysis, and get back something that reads like a first-year MBA textbook. The instinct is to blame the model. But the problem is almost always the input.
A recent analysis from Opinionated Intelligence lays out a framework that matches what most heavy AI users eventually figure out through trial and error: LLMs produce generic output when they can't tell your situation apart from thousands of similar ones in their training data. The fix isn't better prompts. It's better context, and a specific kind of context.
The Same Numbers Tell Opposite Stories
Consider a product with 5 million monthly active users and 80% daily engagement. Sounds healthy. Now add one detail: average session length is 30 seconds.
Is that bad? If it's a social media app, yes. If it's Venmo, where a 30-second session means "I paid my friend and left," it's exactly what you want. The product category completely flips the interpretation, but most people never think to include it when asking an AI for analysis.
Without that context, the model picks the most statistically common interpretation from its training data. For short sessions, that's usually "low engagement." It presents this assumption as a conclusion and moves on. You get a confident-sounding answer that happens to be wrong for your specific case.
Orthogonal Context Beats Deeper Data
The core insight is worth internalizing: breadth of context matters more than depth on any single axis.
Feeding an AI five different engagement metrics (DAU/MAU, session length, retention rate, actions per session, time on page) gives it correlated data points that all describe the same dimension. It's like giving someone five slightly different photos of the same side of a building and asking them to describe the whole structure.
What actually works is providing independent data points across different dimensions:
- Scale: users, revenue, team size
- Engagement: session patterns, feature usage
- Category: what the product actually is and who it's for
- External factors: subsidies, promotions, partnerships inflating numbers
Each new dimension eliminates competing interpretations. With enough orthogonal context, only one coherent story remains, and that's the one the AI will tell you.
A Diagnostic Signal, Not a Failure
This reframes generic AI output from a frustration into useful feedback. When Claude or ChatGPT gives you something bland, it's telling you that your input didn't contain enough information to distinguish your situation from the general case. The generic response is a signal that you need to add a different type of context, not more of the same type.
Five practical rules emerge from this:
- Include metrics from at least three independent dimensions
- Always specify your product category and target user
- Name any external factors distorting your numbers (promos, subsidies, seasonal spikes)
- Call out anything unusual about your situation explicitly
- Treat vague output as a prompt to add more orthogonal detail
This isn't just theory. Anyone who has spent real time with AI analysis tools has noticed the pattern: a simple question with rich, varied context consistently outperforms a sophisticated question backed by thin data. Now there's a name for why.