# Measurements [Why this site exists](why-this-site-exists.md) makes a claim about cost. This page is the evidence, measured on live pages of both sites so anyone can repeat it and disagree with the result. ## What one page costs an agent A **token** is how a language model measures text — roughly three or four characters — and it is what the model pays for, in money and in the limited space it has to think in. A model with a 200,000-token working memory can hold **two** of our documentation pages as the live site serves them, or about **seventy** of the same pages as this site serves them. | Page | Today, on `docs.seon.io` | On the new docs, as Markdown | Cheaper by | |---|---:|---:|---:| | Workflow API | 96,495 | 3,917 | 25× | | Orchestration integration | 60,281 | 4,551 | 13× | | eKYC | 118,868 | 1,820 | **65×** | | Proof of Address | 116,261 | 1,211 | **96×** | ### Is it the same content? The obvious objection is that any page can be made cheap by deleting things, so here is the honest answer, counting prose words with markup stripped from both sides: | Page | Words in the live article | Words here | | |---|---:|---:|---| | eKYC | 1,098 | 1,144 | 104% | | Proof of Address | 923 | 932 | 101% | | Workflow API | 2,384 | 1,658 | 70% | | Orchestration integration | 1,919 | 1,405 | 73% | **The two prose pages are at parity**, so their 65× and 96× are clean: the saving is markup, not substance. **The two API pages are not**, and the reason is a real gap rather than lost prose. Their Markdown contains `{% attributes %}` and `{% endpoint %}` tags, which expand into full reference tables when the page renders. An agent fetching the Markdown gets the tags; a person reading the page gets the tables. So 25× and 13× flatter us — those two rows compare a full page against a thinner one. Expanding the tags when the Markdown is written, rather than only when the HTML is, is the fix. {% callout type="note" title="How this was measured" %} Both sites were fetched with `curl`, following redirects, taking exactly what the server returns. Text was tokenised with `cl100k_base`. The two tables use different baselines on purpose, because they answer different questions. **Cheaper by** compares the whole response, since that is what an agent downloads and pays for. The word counts compare only the live page's article — its `
` element, without the navigation and sidebar — since that is what tests whether the substance survived. Counting the live site's navigation as content would have flattered these numbers rather than tested them. {% /callout %} ## What each site offers an agent Every row below was probed against both live sites. | Capability | `docs.seon.io` | This site | What it means for an agent | |---|---|---|---| | `/llms.txt` index | {% yes label="real, text/plain" /%} | {% yes /%} | How it discovers what exists at all. Without one, it crawls or guesses URLs. | | Per-section `llms.txt` | {% no /%} | {% yes /%} | Lets it pull just the Fraud API without ingesting the entire corpus first. | | `/llms-full.txt` for bulk ingestion | {% no label="404" /%} | {% yes label="219 KB" /%} | One request instead of hundreds when vectorising or working offline. | | Page as raw Markdown, `.md` | {% no label="404" /%} | {% yes /%} | The whole ballgame: 3,917 tokens instead of 96,495 for the same page. | | Page as raw Markdown, `/index.md` | {% no label="404" /%} | {% yes /%} | Agents guess differently. Serving both means neither has to know our conventions. | | Content negotiation on `Accept: text/markdown` | {% no label="returns HTML" /%} | {% no label="returns HTML" /%} | Would let an agent holding only the canonical link get Markdown with no path convention at all. | | `/robots.txt` | {% no label="404" /%} | {% yes /%} | The one standard channel for telling a crawler what not to ingest. | | Deprecated pages kept out of agent indexes | {% no /%} | {% yes /%} | Stops a coding assistant confidently recommending an integration path we no longer support. | | Copy-as-Markdown control in the page | {% no /%} | {% yes /%} | Lets a person hand the page to their own agent in one click, correctly. | | `` in the head | {% no /%} | {% yes /%} | An agent that landed on the HTML learns a cheaper representation exists, without guessing a path. | | Markdown served as `text/markdown` | {% no /%} | {% no label="octet-stream" /%} | Content-type-sensitive agents may skip a file they cannot identify. Ours needs a host-side fix. | {% callout type="tip" title="Credit where it is due" %} `docs.seon.io` already publishes a genuine `/llms.txt`, correctly served as `text/plain`, listing the whole knowledge base. Any claim that the current docs ignore agents entirely is wrong. What is missing is everything downstream of discovery: having found a page, an agent still has no way to fetch it as anything but a 96,000-token HTML document. {% /callout %} ## Gaps on the current documentation 1. **No Markdown representation.** `llms.txt` points an agent at pages it can only then read as full HTML. Discovery without cheap retrieval. 2. **No `llms-full.txt`.** Bulk ingestion means crawling every page at full HTML cost. 3. **`/robots.txt` returns 404.** Crawlers get no directives at all, so there is no way to steer them away from superseded pages. 4. **No way to mark a page superseded.** Legacy integration paths read to an agent exactly like current ones, so a coding assistant can confidently recommend an approach we no longer support. 5. **Hydration overhead.** Up to half of each page is a duplicate copy of its own content in JSON. ## Reproducing this ```bash # What the canonical URL actually returns curl -sSL https://docs.seon.io/api-reference/workflow-api | wc -c curl -sSL https://hackathon.seondf.net/api-reference/workflow-api.md | wc -c # What each site offers an agent curl -sSI https://docs.seon.io/llms.txt | head -1 curl -sSI https://docs.seon.io/llms-full.txt | head -1 curl -sSI https://hackathon.seondf.net/llms-full.txt | head -1 ``` Token counts used `cl100k_base` via `tiktoken`. Figures were taken on 2 September 2026; the live site changes, so re-run rather than trusting them.