Token Economics: Why Heavy Web Pages Cost AI and the Planet More

Apple's homepage sends an AI crawler about 63,500 tokens of HTML. The actual content on the page is about 194 tokens. That is a 327 to 1 ratio: for every token of words a person reads, a machine has to work through 327 tokens of markup, scripts and boilerplate.

We call this token economics. It is one of the six pillars in the Siteiz score, and it matters for two reasons. Heavy pages are harder for AI to read in full. And every wasted byte costs electricity, on every fetch, by every crawler, forever.

What we measured on famous homepages

Siteiz fetches a page the way an AI crawler does: raw HTML, no JavaScript. It counts the tokens in that response, then counts the tokens of real content on the page. These are homepage scans from our published reports, as of October 2026.

Site	Raw HTML	Tokens to read the page	Tokens of actual content	Ratio
Apple	254 KB	~63,500	~194	327 to 1
Shopify	695 KB	~173,400	~749	231 to 1
Notion	231 KB	~57,700	~305	189 to 1
Klarna	925 KB	~230,400	~1,471	157 to 1
Linear	1,287 KB	~321,800	~2,313	139 to 1
HubSpot	644 KB	~160,900	~1,266	127 to 1
Microsoft	261 KB	~65,300	~1,010	65 to 1

These are well-built sites run by strong engineering teams. They look fast and clean in a browser. The weight is invisible to people because browsers are good at hiding it. Machines see all of it.

Machine efficiency: the same idea in plain words

Token economics sounds technical, so here is the plain version. Machine efficiency is how much effort it costs a machine to get your words out of your page.

A page with 750 words of useful content should not need 170,000 tokens of input to understand. When it does, three things happen:

• AI systems read less of you. Agents and retrieval systems work within per-page budgets. A page that blows the budget gets truncated or skipped, and the part that gets cut is often your actual content, buried below scripts and inline data.
• Every reader pays more. The same bloated HTML goes to every visitor, every search crawler and every AI crawler, again on every recrawl.
• The planet pays too. Bytes are not free. They are generated by servers, carried by networks and parsed by machines, and all of that runs on electricity.

Is a lighter website better for the planet?

Yes, in direction. Less data sent means less work for servers, networks and devices. That is the core idea behind sustainable web design, and the W3C now publishes Web Sustainability Guidelines built on it.

AI adds a new reader to that equation. A person loads your homepage now and then. AI crawlers fetch pages continuously, for training, for search indexes and for live answers. Each of those fetches downloads and parses the full HTML. Some systems then run a full browser to execute JavaScript, which costs far more compute than reading plain HTML.

We want to be precise about what we claim. Many AI pipelines strip markup before a model sees the page, so the model itself may not process all 300,000 tokens. And no one can honestly put a kilowatt-hour figure on one page fetch, because it depends on hardware, networks and energy mix. So Siteiz does not show a carbon number. What we can say with confidence: a page that sends less code for the same content costs less energy to serve and to read, every time, for every reader.

Good news for anyone who cares about both: the fixes that make a page lighter for the planet are the same fixes that make it easier for AI to read.

How to check your own page

You do not need a tool to get a first look. Fetch your page as a crawler would and compare the size with the words you can see:

# How many bytes does an AI crawler receive?
curl -sL -A "Mozilla/5.0 (compatible; GPTBot/1.0; +https://openai.com/gptbot)" \
  https://yourdomain.com/ | wc -c

# How much of that is inline script?
curl -sL https://yourdomain.com/ | grep -o "<script[^>]*>" | wc -l

If a page with a few hundred words returns several hundred kilobytes, most of what machines receive is not content. Open the source and scroll: large inline JSON blobs, repeated SVG icons and long inline style blocks are the usual suspects.

Best practices for lighter, AI-ready pages

These eight practices cut machine cost without changing what people see. Where they match the W3C Web Sustainability Guidelines, the guideline number is in brackets.

1. Put your content in the server HTML. Text that only appears after JavaScript forces crawlers to render the page in a browser, the most expensive way to read it, or to miss the content entirely. Server rendering or static generation fixes both. See does ChatGPT read JavaScript for how to test it.
2. Shrink hydration payloads. Frameworks often embed the full page state as inline JSON so the browser can rebuild the page. On many sites that blob is larger than the visible content. Send only the data the page needs.
3. Move inline SVG and CSS into cached files. An icon repeated forty times inline is sent forty times on every fetch. A cached sprite or stylesheet is sent once.
4. Remove unused code and dependencies (WSG 3.2, 3.12). Old tracking tags, abandoned A/B tests and unused libraries add weight on every page.
5. Load third parties on interaction (WSG 3.5). Chat widgets, video players and social embeds can wait until someone clicks. Crawlers never click, so they never pay for them.
6. Compress and cache (WSG 4.2, 4.3). Brotli or gzip cuts transfer size sharply. Note that compression reduces bytes on the wire but not the tokens a model reads, so it complements the steps above rather than replacing them.
7. Publish your key facts as structured data (WSG 3.8). A compact JSON-LD block gives machines your name, products and prices directly, without parsing the whole layout.
8. Offer a clean text version. An llms.txt file or Markdown versions of key pages give AI a lightweight path to your content. Read why Markdown matters for AI search for the reasoning.

After each change, measure again. The AI visibility audit walks through the full manual check, and the methodology explains how the efficiency pillar is scored.

Lighter is the rare fix where everyone wins

Most website trade-offs pit one goal against another. This one does not. A lighter page loads faster for people, is cheaper for AI systems to read in full and uses less energy every time it is served. If you care about AI search, about performance or about the planet, the work is the same.

Check your page's machine efficiency with a free Siteiz scan.