Google vs AI: Mobile Search Uncovers Hidden Energy Costs

⚡

Key Takeaways

1A Google search on mobile consumes five times more energy than an AI query, according to a new study.

2The hidden energy costs include mobile networks and programmatic advertising, often overlooked in comparisons.

3LLM models reduce screen time and avoid energy waste related to advertising bids.

💡Why it matters — Companies need to rethink their digital strategy to reduce their energy footprint, prioritizing AI for certain tasks.

The Myth of Energy-Intensive AI vs. Google

The widely held belief that artificial intelligence consumes more energy than Google searches is based on a biased comparison. In reality, the data tells a completely different story, session by session.

Imagine a common scenario: you use your smartphone to perform a Google search. You type in your query, get a list of results, click on the first link, wait for the page to load, but it’s not what you were looking for. You go back, try another link, and so on. In fifteen minutes, how much energy has your phone consumed? Few people ask this question, and therein lies the problem.

Since 2023, a prevailing idea has emerged: an AI query consumes ten times more energy than a Google search. This claim has infiltrated corporate social responsibility reports, regulatory debates, and strategic discussions. It is ubiquitous in conferences and AI maturity benchmarks.

However, this comparison is flawed. It only compares the server cost of a Google query with that of a response generated by a language model, without accounting for the energy consumed by your device, the network, and the background advertising systems.

In March 2026, Charles Duprat from ICOM'Provence published a working paper that finally asks the right question: how much energy does a user consume to obtain a complete answer to their information need, from the first search to the final response? This shift in perspective changes everything.

A Biased Comparison Due to Incomplete Data

In 2023, a claim spread: a query via a language model consumes ten times more energy than a Google search. This idea has taken root in discussions about corporate social responsibility, regulatory articles, and expert speeches.

The problem is that this comparison only considers the energy consumed by Google's data center to process a query (about 0.30 Wh) and that required to generate a response by a language model (between 0.24 and 0.34 Wh, according to Google, Epoch AI, and OpenAI). But this only represents part of the story.

A Google search does not directly provide you with information but rather a series of links to information sources. The energy required to navigate these pages, download their content, execute JavaScript scripts, and manage advertising bids is borne by your device, your mobile network, and an often-invisible advertising infrastructure. None of this is accounted for in data center consumption calculations.

The Real Numbers of Energy Consumption

Let’s take a concrete example: comparing two technical solutions using three different sources on a non-standalone 5G network, which is still the norm in France in 2026.

Energy Consumption of a Web Search Session:

Query processing by the server: 0.30 Wh
Network: 3 pages × 2.56 MB (HTTP Archive median 2025) × 0.14 kWh/GB = 1.08 Wh
Page rendering (device CPU/GPU): 0.60 Wh
Advertising load (30% of rendering, according to Khan et al. 2024): 0.18 Wh
Screen time (6 minutes × 2.5 W): 0.25 Wh

Energy Consumption of an Equivalent LLM Session:

Inference (standard model, no complex reasoning): 0.30 to 0.40 Wh
Network: text payload of 5 KB, negligible
Screen time (2.5 minutes): 0.10 Wh

Total: 0.40 to 0.50 Wh

The central ratio is 5.4 times in favor of the language model. Duprat validated this figure through a Monte Carlo analysis on 10,000 simulations with 9 variable parameters. No combination of values brings the consumption of the search below that of the language model. In the worst case for the language model, the ratio is 1.6 times.

The Mechanisms Behind This Inversion

First mechanism: the mobile network is the main culprit. In 2025, the median mobile page weighs 2.56 MB. On a 4G network, this represents an energy cost of 0.44 Wh for transmitting a single page, even before displaying anything. A response generated by a language model is a text of 2 to 10 KB, with a transmission ratio of about 500:1. The network is not a marginal cost but the dominant component of the energy footprint of a mobile search session.
Second mechanism: programmatic advertising is an invisible energy tax. When loading an ad-supported page, a bidding process opens in parallel. Dozens of demand platforms receive the request, but most fail, consuming CPU cycles unnecessarily. Khan et al. (2024) measured that integrated ad blockers reduce device energy consumption by 15 to 44% compared to normal browsing. This means that a significant portion of the energy consumed by your smartphone while browsing fuels the advertising ecosystem, not the content you are viewing. A language model completely avoids this infrastructure.
Third mechanism: faster completion reduces screen time. The CHI 2025 study by Spatharioti et al. showed that users of language models complete their synthesis tasks faster, with fewer queries. Less screen time means less energy consumption. The behavior of "pogo-sticking" — clicking, finding the page disappointing, going back to search results, starting over — creates an energy penalty that static models never capture. Each return to search results on mobile costs an additional 0.30 to 0.60 Wh. The language model structurally eliminates this pattern by providing a complete synthesis from the first exchange.

The Blind Spots to Consider

Three real limitations, not mere nuances of comfort, must be taken into account:

On fixed Wi-Fi, the advantage of the language model collapses. On a fixed network (0.006 kWh/GB), the transmission cost drops by 95%. The advantage of the language model falls to 1.5–2.5 times for complex tasks and reaches parity for simple queries. The inversion is a phenomenon specific to mobile.
On reasoning models, the logic reverses, sometimes drastically. Claude Opus in reflection mode, GPT-o3, Gemini Deep Think: these models generate extended reasoning chains. Jin et al. (2025) documented an average expansion of 4.4 times the output tokens in production, with extreme cases reaching up to 113 times. The crossover threshold with mobile search — the point at which the language model becomes more energy-intensive — is between 4 to 8 times.
The Jevons paradox is not negated by unit efficiency. ChatGPT surpassed 2 billion daily queries by the end of 2025. If this demand is new rather than substituting for web searches, total consumption increases regardless of the unit ratio. Efficiency per session says nothing about aggregated efficiency. These are two distinct questions, and both deserve serious answers.

Concrete Implications for Organizations

For those leading an AI deployment strategy or a digital social responsibility policy, here are three direct implications:

First: the choice of model is an energy decision, not just a performance decision. Using a reasoning model for standard synthesis tasks — which the majority of companies do by default because "it's the best model" — multiplies the footprint by a factor that no one accounts for in carbon balances. Intelligent routing by task complexity is not a technical luxury; it is a CSR consistency imperative.
Second: your mobile teams are doing web searches where a standard language model would be five times less energy-intensive. For synthesis tasks, monitoring, multi-source comparison — the substitution is measurable, immediate, and requires no additional investment.
Third: the web's advertising infrastructure is an externalized cost that your employees bear without seeing it. Auditing your organization's digital footprint without accounting for the 15 to 44% of device energy absorbed by programmatic advertising is a significant accounting blind spot.

Field Observations Over Three Years

Among the 200 AI projects I deployed in B2B companies between 2022 and 2025, a consistent pattern emerged during digital maturity audits: teams that shifted their research and synthesis tasks to a standard language model did not do so for ecological reasons. They did it because it is faster.

The collateral result is that they mechanically reduced their mobile digital footprint, without realizing it, without measuring it. Duprat now provides the framework to quantify it. And the figure is stark.

What This Changes Practically

Let’s return to the scene at the beginning. You are searching for complex information on mobile.

Scenario A — Google:

Four queries, seven pages loaded, fifteen minutes of screen time, background advertising bids. Estimated consumption: 2.41 Wh.

Scenario B — Standard Language Model:

One query, synthesized response in thirty seconds. Estimated consumption: 0.40 Wh.

You haven’t made an ecological gesture. You have simply asked your question in the right place.

The narrative "AI consumes ten times more than Google" is not only inaccurate. It protects an infrastructure — the mobile advertising web — whose real energy cost has never been properly accounted for because no one had an interest in doing so.

A modern web page is not a document. It is a software package that executes hundreds of operations for advertisers you will never see. You pay the cost in time. Your battery pays the cost in watts.

The language model wins this comparison because its opponents are extraordinarily inefficient. Not because it is virtuous. This is a difference that matters — for your purchasing decisions.