Agentic AI has shown dazzling promise, but it also raises serious concerns. And some of the most pressing ones aren't about reasoning at all.
A new analysis of 250 real-world queries across common retail tasks, such as product pricing, availability, ratings, shipping and specifications, reveals systemic inefficiency at the heart of web-based AI agents. On average, 97.9% of the data retrieved by agents from live web pages is irrelevant to the query being answered. Specifically, the average page ingested ran nearly 9,000 characters, while the average answer was just 32 characters, resulting in a noise-to-signal ratio of 278:1. Price queries were the most extreme outlier, with noise rates approaching 99.5%.
That's not a rounding error. That's a structural problem.
There Are Two Problems That Compound Each Other
The consequences of this data inefficiency show up in two ways, and they reinforce each other.
Accuracy takes a hit. Large language models are sensitive to the quality of context. When the context window is long, the signal is sparse, and multiple plausible-but-wrong answers compete for attention, models are more likely to pick the wrong one.
Retail pages are a near-perfect illustration: a single product page often contains the current price, a sale price, bundle pricing, prices for related configurations, and historical price references buried in user reviews. The model isn't failing because it's incapable; it's failing because it's working with a mess.
Costs scale with noise. At roughly four characters per token, a 9,000-character page translates to approximately 2,200 tokens per query. Multiply that across thousands of daily queries and multi-step retrieval pipelines where an agent might hit several pages before arriving at an answer, and the token math becomes alarming.
The majority of token spending in a typical agentic system may be attributable to data that never contributes to a correct answer. That's wasteful, of course, and for organizations running AI agents at scale, it's a meaningful cost driver.
It's Not Just a Retail Problem
While this specific dataset centers on retail, it reveals an underlying dynamic that affects virtually every industry that relies on live web data: the web was built for humans, not machines.
Human-readable web pages are packed with navigation menus, ads, related products, legal disclaimers, user reviews and various layout quirks. An AI agent retrieving a page to answer a narrow question doesn't benefit from any of that, but it still has to process all of it.
The same dynamic can just as easily show up in finance (extracting one metric from an earnings report), travel (finding availability from a booking page), real estate (pulling property details from a listing site), or healthcare (identifying specific data from a provider directory).
In every case, the agent is doing more work than necessary to retrieve a very small but specific signal.
The Missing Component? Retrieval Architecture.
There's a common assumption in the AI community that improving agent performance means better models, longer context windows, or more sophisticated reasoning chains. But the data here suggests otherwise.
If 97.9% of the input is irrelevant before the model ever begins reasoning, no model improvement will fully compensate for that. With this the case, the bottleneck isn't reasoning. It's retrieval.
This reframes where engineering attention should be focused for teams building production-grade agentic systems. Signal-to-noise ratio deserves to be treated as a first-class infrastructure metric, alongside latency and accuracy. That shift comes with three practical implications:
1. Prefer structured data over raw HTML. If the answer exists as a discrete field (a price, a rating or a ship date), it should be retrieved as a field, not extracted from a wall of prose. Structured data pipelines deliver more reliable inputs and dramatically lower token overhead.
2. Pull precisely scoped data, not full documents. Agents don't need entire pages. They need the specific slice of information relevant to the query. Retrieval systems should be designed to scope requests narrowly rather than broadly.
3. Evaluate the data layer, not just the model layer. When assessing AI agent infrastructure, the question shouldn't only be which model performs best on the given task. You should also be asking how much of what the model receives is actually useful.
Purpose-built web intelligence APIs designed to return structured fields rather than raw page dumps are increasingly where this gap gets addressed.
Better Data Beats a Better Model Every Time
The good news? This is a solvable problem. It doesn't require a better model or a bigger context window. It requires better data hygiene earlier in the pipeline.
For developers building or scaling AI agents, the most impactful architectural decision may not be which LLM to use; it may be how and what data is fed into it. Treating retrieval as an engineering discipline, with the same rigor applied to accuracy and latency, is what separates agents that work in controlled environments from those that hold up in the real world.