Your bloated HTML has an energy bill
Cleaner pages mean fewer bytes to transfer and fewer tokens to process. Content negotiation is a small but practical way to make agent traffic less wasteful.
There is no such thing as a free token.
Every page has to move across a network, get parsed, and often pass through a model. When the useful article is wrapped in navigation, scripts, cookie controls, and layout markup, all of that machinery comes along for the ride.
The climate impact of AI is a large systems problem. Your website will not solve it. But it does control one small input: how much irrelevant material an agent has to process before it reaches the point.
The math on wasted tokens
A heavily built blog post can weigh hundreds of kilobytes as HTML while its core copy fits in a few kilobytes of Markdown. That is not just a transfer-size difference. Markup becomes tokens too.
Converting HTML to Markdown reduces token consumption by an average of 60–88% depending on the page type. For tables, HTML can be 3–5x more token-heavy than Markdown. For full pages with nav, scripts, and footers, the ratio is often worse.
Tokens per joule is becoming a useful efficiency metric. Research from EuroMLSys measures how much model output a system produces for the power it consumes. The inverse matters too: tokens that add no useful context still require work.
We should be careful with the math here. A 95% reduction in input tokens does not guarantee a 95% reduction in total energy; model architecture, batching, hardware, output length, and data-center efficiency all matter. It does mean the system has less irrelevant input to move and process. That is still worth doing.
Data centers are not a rounding error
U.S. data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of total U.S. electricity consumption. That figure is projected to grow to 426 TWh by 2030. The IEA estimates AI-related data center emissions will reach 1% of global CO₂ by 2030 in its central scenario.
To put the water footprint in context: by 2030, data centers are projected to drain 731 to 1,125 million cubic meters of water per year for cooling. That's the annual household water use of 6 to 10 million Americans.
Bloated web pages are not the main cause of those numbers. Hardware, model design, utilization, grid carbon intensity, and cooling infrastructure matter far more. Token efficiency is simply one lever a website team can pull without waiting for the rest of the industry.
Agentic workflows multiply the effect
An agent researching a topic rarely stops at one page. It fetches a result, extracts the useful parts, follows a link, and does it again. The waste repeats at every step.
Our audit crawler visits up to 10 pages per site. When a site sends the full browser document to a Markdown request, the crawler has to move and inspect substantially more material than it would with a purpose-built response. The exact ratio changes by site. The pattern is remarkably consistent.
At web scale, small inefficiencies stop being small. We do not need a heroic estimate of total agent traffic to know that sending less irrelevant data is better than sending more.
Efficiency and sustainability are the same argument
This conversation usually starts with cost and performance: fewer tokens, faster responses, lower bills. Fair enough. Efficiency at infrastructure scale is also an environmental argument.
The web has spent 30 years optimizing for human browsers. Lazy-loading images, minifying scripts, caching static assets: these practices exist because sending unnecessary bytes has a cost. Content negotiation for AI agents is the same principle applied to a new kind of reader.
Agents are becoming another audience for the web. Give them the representation they need, not the entire interface we built for a browser.
Serving clean Markdown will not make AI sustainable by itself. It is simply the less wasteful option—and one we can ship today.
Preview one of your pages to see how much of its response is useful content and how much is along for the ride.