Nimble’s Research Reveals AI Retail Agents Waste 99% of What They Read
Analysis of 250 real retail queries finds a hidden cost and accuracy crisis in agentic AI as agents read an average of 278 characters for every one character that actually matters
NEW YORK, May 27, 2026 (GLOBE NEWSWIRE) -- Nimble, the only real-time web search platform trusted by hundreds of enterprises, today published research revealing a fundamental inefficiency at the heart of how retail AI agents interact with the web: for every character of useful information an AI agent extracts from a retail webpage, it processes 278 characters that contribute nothing to the answer, a noise rate of more than 99%.
This finding, drawn from an analysis of 250 live retail queries across Amazon, Walmart, Best Buy, Target and other major U.S. retailers, exposes what Nimble calls the signal-to-noise problem, and it has direct consequences for both the accuracy and cost of production AI systems.
"These numbers aren't a rounding error; they indicate a structural problem," said Nimble CEO and co-founder Uri Knorovich. "The web was built for humans, not machines. When AI agents retrieve full pages the way a browser does, they're reading an encyclopedia to answer a single trivia question. They then repeat that process at scale, in real time, for every query."
The Agentic AI Numbers Are Stark
Across all 250 queries, spanning product prices, ratings, availability, discounts, shipping policies, and product descriptions, the average webpage retrieved by an AI agent contained 8,795 characters, and the average answer was 31.7 characters long.
In aggregate, fewer than 0.4% of all characters processed across the dataset were relevant to the question.
For price queries, the problem is extreme. A typical Amazon product page exceeds 9,000 characters. The correct answer, a dollar amount like "$174" or "$100," is four to six characters. That's a 99.48% noise rate. In the five most data-inefficient queries in the dataset, agents retrieved pages of up to 45,000 characters to extract a price that fit in a text message.
The Signal-To-Noise Problem Matters Beyond Efficiency
The research makes clear that this is more than a waste problem; it's an accuracy problem.
For example, a typical Amazon product page for a consumer electronics item doesn't just list a single price. It contains the buy-box price, a struck-through "list price," variant prices for different configurations, bundle pricing, sponsored product prices, and prices mentioned in customer reviews. An AI agent that is handed 30,000 characters of that content and asked to identify a specific product's price must reason across dozens of plausible-looking but incorrect answers.
Research on large language model behavior has consistently found that accuracy degrades as context grows and as the density of relevant-seeming but incorrect information increases. The structural conditions measured in Nimble's dataset are precisely those conditions.
"The model isn't the bottleneck," said Knorovich. "If 97.9% of your input is irrelevant, no model can fully compensate for that. So the problem is retrieval, not reasoning."
A Cost Problem That Scales
LLM APIs are priced per token, approximately one token per four characters. At an average noise rate of 97.9%, an agent running 1,000 queries per day is paying to process roughly 8.5 million characters of useless content daily.
At current frontier model pricing, the token cost attributable to noise alone can reach hundreds to thousands of dollars per month, scaling linearly with query volume. In the full 250-query dataset, the overwhelming majority of total token spend was tied to content that never contributed to any answer.
These findings point to a clear shift in how production AI systems should be designed. Agents that retrieve structured data directly, rather than parsing answers out of raw webpages, process a fraction of the content at a fraction of the cost, while also reducing the surface area for accuracy errors.
As enterprises rethink how AI agents retrieve and process information, understanding the cost of noisy web data becomes increasingly important. The full dataset and methodology are available in this blog post.
About Nimble
Nimble is the enterprise platform for real-time agentic web search that turns the public web into trusted, decision-grade data for AI systems and business-critical workflows. As organizations scale AI, many models fail in production due to incomplete, outdated, or unverifiable data. Nimble closes this gap by delivering highly accurate, task-specific web data in seconds, designed for enterprise reliability at scale.
Nimble is trusted by banks conducting real-time due diligence, retailers making dynamic pricing decisions, and analysts performing time-sensitive market research, without the break/fix burden of traditional web data approaches. Headquartered in New York, the company is backed by leading global investors, including Norwest, Databricks Ventures, Target Global, Square Peg, Hetz Ventures, Slow Ventures, R-Squared Ventures, J-Ventures and InvestInData.
For more information, visit https://www.nimbleway.com/.
For Media Inquiries Contact:
BOCA Marketing Agency
nimble@bocamarketing.com
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
