ChatGPT water usage vs Google Search

We cannot reliably say how ChatGPT’s water usage compares with one Google Search. OpenAI has published a company-level average for a ChatGPT query, but without a detailed method. Google has published a detailed water estimate for a median Gemini Apps text prompt—not for a conventional Search query. No current, like-for-like public measurement covers both products.

Any neat ratio therefore substitutes assumptions for evidence. Search and chat can also do different jobs: retrieving a webpage is not the same task as generating a long answer, and modern search may itself include an AI summary. The honest conclusion is conditional. A simple search likely performs less generative work than a long ChatGPT task, but public water data does not support an exact “times more” claim.

What OpenAI has disclosed

In 2025, OpenAI chief executive Sam Altman wrote that the average ChatGPT query uses 0.000085 US gallons of water. That converts to roughly 0.32 millilitres. His post did not publish a supporting calculation or specify the mix of models, request lengths, locations, cooling systems or direct and indirect water included.

The figure is useful as an operator’s stated fleet average. It is not a meter reading for a particular request. Asking for a sentence edit, uploading a large document and running a web-research workflow can trigger different amounts of computation. An average can hide that distribution.

Earlier academic work modelled much higher water use under different infrastructure and accounting assumptions. That conflict cannot be resolved by choosing the number that best fits a headline. Our AI environmental impact FAQ explains why measured, modelled, average and scenario figures must keep their labels.

What Google has disclosed—and what it has not

Google reported that a median Gemini Apps text prompt in May 2025 consumed an estimated 0.26 ml of water. Its technical boundary includes active accelerators, idle capacity, CPU and memory, and facility overhead. Water is derived using Google’s fleet water-use efficiency.

That is stronger methodological detail than the ChatGPT statement, but it is the wrong denominator for this comparison. Gemini Apps is a generative assistant. It is not the ordinary Google Search system. The figure also describes a median prompt at one point in time, while Altman described an average query.

Google publishes environmental data for its data-centre fleet, which runs Search, Gemini, YouTube, Cloud and other services. Fleet totals and efficiency metrics do not reveal the water allocated to one modern Search. Dividing them by guessed query volume would mix products and create a result Google did not report.

Why water per request is hard to assign

Data centres may consume water directly through evaporative cooling. Electricity generation and semiconductor manufacture can use water too. Studies differ on whether they include these indirect stages. “Water use” may also refer to withdrawal or consumption, which are not interchangeable.

Shared infrastructure adds another allocation problem. Cooling systems serve racks and workloads together. Idle equipment remains ready for traffic. A per-request figure apportions part of those shared systems using a method; no droplet is physically reserved for a search box or chatbot window.

Allocation rules can also change as providers improve metering, so the publication date belongs beside the result.

Location and timing affect both volume and consequence. A hot afternoon can create different cooling demand from a cool night. Water from a stressed catchment has a different local impact from the same volume in a water-abundant region. What a data-centre PUE means helps with energy overhead, but PUE does not measure water or water stress.

Search and ChatGPT may not perform the same task

If you need a known website, a conventional search can return links without generating a long response token by token. If you need synthesis or rewriting, ChatGPT may complete work that otherwise takes several searches and page visits. Comparing one visible interaction on each side says little about task completion.

The product boundary is also blurred. Google Search may display an AI-generated overview. ChatGPT may search the web, retrieve several pages and call its model multiple times before producing one answer. Either interface can combine ranking, retrieval and generation.

A fair experiment would define the information need, require an acceptable result, count every backend step and retry, and measure both services using the same water boundary, geography and period. Commercial providers do not expose enough data for an independent public test of that design. Claims that one is exactly a certain multiple of the other go beyond the evidence.

Energy does not provide a clean shortcut

Some comparisons try to estimate water from electricity alone. Energy can be a starting point, but a single conversion factor assumes a cooling and electricity-water relationship that may not match either service. An efficient data centre can still make a water-versus-energy trade-off by using evaporative cooling to reduce electricity demand.

Old energy comparisons create another trap. The famous Google Search figure dates from 2009, while AI services and hardware are recent and changing. Our guide to choosing an AI assistant by energy use recommends rejecting comparisons that mix years, tasks or boundaries.

Carbon cannot settle the water question either. Lower-carbon electricity may be water-intensive, and low-water cooling may use more electricity. Report the metrics separately rather than collapsing them into one environmental winner.

When a comparison omits the conversion factor, cooling design or grid boundary, it is not reproducible enough to guide a product choice.

What users can reasonably do

Use conventional search when you want links, a current source or a known page. Use an assistant when generated synthesis adds value. Keep requests focused, avoid unnecessary long context and reserve research agents, images or repeated variants for tasks that justify the extra work.

When assessing provider claims, ask whether the number covers direct cooling, electricity supply, idle capacity and the whole response workflow. Preserve the date and whether it is a median or average. Look for local water-risk reporting alongside global totals.

Treechat estimates its own message water from estimated energy using a fixed assumption disclosed on its methodology page. It does not measure a cooling tower or claim a universal comparison with Search. Until providers publish comparable, product-specific measurements, “we do not know the ratio” is the most accurate answer—and a better basis for decisions than a confident but unsupported number.