Is the problem really the AI… or the data we feed it?

Ever since generative AI arrived, the same criticism keeps coming back. “The AI gets it wrong.” “The AI lacks relevance.” “Its recommendations are too generic.”

The conversation has focused almost entirely on the models. Which one is the smartest? The fastest? The most reliable? As if the quality of an answer depended solely on the machine producing it.

A LirenPrism measurement suggests shifting the gaze slightly. Not onto the models, nor the brands, but onto what feeds an AI engine's reasoning before it even begins to answer.

An AI always answers with what it knows

By default, an AI engine draws on two main sources. The knowledge it acquired during training, and — when it can — information retrieved from the web. In both cases, it works from a general view of the world.

But an executive doesn't ask an AI to recite general knowledge. They ask it to help them decide. And a decision is never better than the information that builds it. Your market isn't general. Your customers aren't. Neither is your competition. Between the world's average knowledge and a company's precise situation lies a gap the model alone cannot close.

It's an old principle in computing, and it hasn't aged. Garbage In, Garbage Out. If the inputs are incomplete, outdated or generic, the recommendations will be too. This isn't an isolated intuition. Recent work by McKinsey stresses that data readiness and quality have become one of the main barriers to scaling AI in the enterprise. Models aren't enough when the context they're given is thin. The same premise underpins so-called RAG architectures, which feed an AI information specific to an organisation: their effectiveness depends directly on the quality, freshness and reliability of the data retrieved.

Measuring what actually shapes an answer

Rather than asking an AI to propose a strategy directly, we started with a deliberately more modest step. Not “what decision do you recommend?”, but “what criteria do you consider important to answer this question?”.

The question analysed was intentionally simple and concrete: “What gifts for my 10-year-old daughter for Christmas with a budget of €100?”

Forty queries were run with the AI alone. Then forty more with web search enabled. Same question, same wording. Only the informational context changed.

Context doesn't just change the answer — it changes the criteria

The results, and this is a fact demonstrated by the measurement, show a clear shift in the criteria mobilised depending on whether the AI has web search or not.

Without web search, the price criterion already dominates heavily, present in 87.5% of answers. With web search, it stays dominant, at 95%. No surprise there: the budget was written into the question.

But around that price criterion, the rest of the reasoning transforms:

• Performance rises from 7.5% to 77.5% of answers.
• Practicality, from 20% to 70%.
• Quality, from 10% to 40%.
• Aesthetics, from 2.5% to 32.5%.
• And criteria that were simply absent before appear: safety, health, versatility.

In other words, for an identical question, the whole set of dimensions the AI deems worth considering reorganises itself the moment you change what it has in front of it.

We must be precise about what this observes, and what it doesn't. The study does not say the enriched criteria are “better”. It passes no judgement on the relevance of the answers. It shows one thing, and one thing only: the criteria an AI mobilises change when its informational context changes.

No longer a question of the answer, but of the reasoning

That shift is what deserves attention. As long as we compare final answers, we stay on the surface. But the moment we look at the criteria, we reach the reasoning itself.

And if the criteria change, then the final recommendations can change too, without the user being remotely aware of it. They read a structured, well-argued, convincing answer. They don't see that the very same answer could have organised itself around an entirely different set of criteria had the AI held different information.

An AI never decides better than the information it's given.

What this measurement suggests is that a large part of what we take for an AI engine's “intelligence” is in fact a function of its context. The model doesn't change from one query to the next. What changes is the material we give it to work with. This reading remains, at this stage, a working hypothesis: the measurement establishes the shift in criteria, it does not mechanically isolate every cause.

Another way of working with AIs

Today, many companies ask the big question directly. “What marketing strategy do you recommend?” “What product should I launch?” Perhaps we should start differently. Not with the recommendation, but with the context of the reasoning:

• Which criteria does the AI consider important for this question?
• Which does it ignore?
• What is it missing to reason otherwise?
• What market information should we feed it before it answers?

Interrogating the criteria first, then enriching the context, then only asking for a recommendation — this sequence inverts the way most professional uses of AI work today. It treats the answer not as a destination, but as the output of a reasoning process we can inspect, and act upon upstream.

The real question

For several years, the race focused on the models. Which is the best LLM? The smartest? The fastest? Legitimate questions, and they remain so. But as models progress and converge, raw performance stops being the deciding factor.

Another question then takes over. Is the problem really the AI… or the data on which it builds its answer?

Models will keep evolving. But one constant will likely stay true, because it predates the technology: better information generally produces a better decision. It was true before generative AI. It still is with it.

And what if, tomorrow, the real competitive edge no longer lay in choosing the best model, but in the ability to give it the best possible measurements before it even starts to reason?