# Token measurement methodology — Agent Delivery Network — ADN

← Back to home Token measurement methodology
 ADN compares how many tokens an AI model would process when reading a page as full HTML versus ADN's indexed markdown version. This page documents what we measure, how we calculate dashboard metrics, and what we do not claim.
 What we measure
 For each indexed page we count tokens in:
 1. Origin HTML — the HTML returned when ADN ingest fetches the public URL (including browser render when the static HTML is a client-rendered shell).
 2. ADN markdown — the markdown ADN stores and serves to agents for that path.
 Both counts use the [o200k_base](https://github.com/openai/tiktoken) tokenizer (the same family used by recent OpenAI models). Counts are stored on domain_pages at ingest time.
 Results (marketing benchmark)
 On 1 Nov 2026 we measured token counts on sites we serve (aggregate only — we do not publish per-path breakdowns):
 Metric Value
 HTML tokens (total) 1,109,491
 Markdown tokens (total) 29,600
 Fewer tokens (overall) [97.3%](/methodology#results)
 Median page [97.7%](/methodology#results) fewer
 Homepage (HTML) 45,172
 Homepage (markdown) 1,157
 Rounded headline figure: 97% fewer tokens on sites we serve (from 97.3% overall ).
 Download the same aggregate figures: [token-benchmark-2026-11-01.csv](/token-benchmark-2026-11-01.csv) (UTF-8 CSV; totals and percentages only).
 How we count tokens saved (dashboard)
 Per page: html_tokens − markdown_tokens (never negative).
 For a period (e.g. this calendar month, UTC): for each page, multiply that per-page saving by the number of times agents were served ADN markdown, structured XML, or JSON from the indexed path. We join request_logs on domain_id + normalised path, counting rows where serve_format is markdown , doclang , or json , filtered by created_at .
 Non-page paths (for example /opengraph-image ) are excluded. Trailing slashes are normalised before matching.
 When stored token counts are missing, we estimate using ≈4 characters per token (OpenAI rule of thumb) and label the metric as estimated until ingest refreshes the page.
 From tokens to energy and water (context only)
 Smaller inputs usually mean less compute, but ADN does not measure your AI provider's data centre . Actual energy and water depend on the model, hardware, region, and whether cost is reported for input, output, or hardware. The figures below are third-party estimates and medians — not ADN measurements.
 Google (21 Aug 2025)
 Google reports that the median Gemini Apps text prompt uses about [0.24 Wh](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) , [0.03 gCO₂e](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) , and [0.26 mL of water](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) ( May 2025 data ). Google states these figures are not independently verified.
 • Measuring the environmental impact of AI inference
 • arXiv:2508.15734
 Epoch AI (7 Feb 2025) — estimate
 Under Epoch's assumptions , a typical GPT-4o query is roughly [0.3 Wh](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use) ; a 10k-token input around [2.5 Wh](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use) ; a 100k-token input almost [40 Wh](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use) , with input cost growing faster than input length.
 • How much energy does ChatGPT use?
 Mistral (22 Jul 2025)
 Mistral reports [marginal inference impact including upstream hardware (excluding user devices)](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai) for a 400-token Le Chat response : [1.14 gCO₂e](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai) and [45 mL of water](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai) .
 • Our contribution to a global environmental standard for AI
 Li et al., "Making AI Less Thirsty"
 Li et al. estimate roughly [500 mL per 10–50 medium-length responses](https://arxiv.org/abs/2304.03271) (GPT-3 in Microsoft US data centres).
 • arXiv:2304.03271
 IEA, Energy and AI (2025)
 The IEA reports data centres used about [415 TWh in 2024](https://www.iea.org/reports/energy-and-ai/executive-summary) (~ [1.5%](https://www.iea.org/reports/energy-and-ai/executive-summary) of world electricity), projected to reach about [945 TWh by 2030](https://www.iea.org/reports/energy-and-ai/executive-summary) .
 • Executive summary
 Limits
 • Provider figures vary widely; most do not separate input from output cost.
 • ADN reduces input tokens only (HTML not sent when agents receive markdown).
 • We do not control or measure providers' infrastructure, cooling, or grid mix.
 • Relating token savings to Wh or litres requires assumptions we do not make in product metrics.
 What we don't claim
 • No carbon-neutral , net zero , or offset claims.
 • No litres of water saved or Wh saved stated as fact in ADN marketing or dashboard numbers.
 • We do not use standalone vague terms such as "eco-friendly", "green", or "sustainable AI" without this measurement basis.
 Update log
 • 1 Nov 2026 — Initial publication: aggregate marketing benchmark on sites we serve, o200k_base methodology, dashboard tokens-saved metric, aggregate CSV.
 Ask ADN