# Why Onto · The compatibility layer for the agent web > Three layers, one engine. Read API for AI developers, Serve SDK for site owners, and Act on the way. The infrastructure the agent web needs. **Source:** /why-onto **Extracted:** 2026-05-20T20:59:17.236Z --- Why Onto ## Anyone can clean up HTML. _Only one company is building the full compatibility layer._ The agent web has three problems. AI developers can't read the web cleanly. Site owners can't serve themselves to agents reliably. And agents can't act through any UI without breaking when a button moves two pixels. Read and Serve are already shipping. **Act** is the destination — and it's the layer no one else is building, because it requires the cleaning engine, the accuracy scoring, and the protocol-level work Read and Serve already proved out. Three layers. One engine. One destination. _That's the compatibility layer for the agent web._ 00 — The full stack ### Three layers. _One engine._ One destination. Read is the wedge. Serve is the install base. Act is what makes the agent web actually programmable. Each layer is built on the same cleaning engine, the same accuracy scoring, the same edge runtime. The layers compound — Read and Serve aren't products separate from Act, they're the engineering Act requires to ship. Layer 01 · Read #### AI developers call an API. Drop-in cleaner for any web HTML. Returns Markdown an agent can actually use, plus an AIO score that flags hallucination risk before the model sees the data. A Firecrawl alternative with accuracy scoring and MCP support — the wedge anyone building an agent already needs, served through the same API or directly via Claude Code, Cursor, and any MCP client. Live Layer 02 · Serve #### Site owners install one line. AI crawlers get clean Markdown at the edge. Humans get the full HTML, unchanged. Our own AIO score went 50 → 80. Gemini stopped hallucinating our scoring formula. Claude reads us in Markdown and explicitly confirms the MIME type — we ran the test live and Claude said: _“Onto literally serves clean Markdown to AI agents instead of raw HTML. They practice what they preach.”_ Alpha Layer 03 · Act #### Agents act through the same engine.the destination The vision. When agents can read, serve, and act through one engine, the agent web stops being a research surface and becomes a transactional one. Today's web wasn't built for agents clicking buttons through semantic intent — it was built for humans clicking buttons through pixels. Act changes that. Site owners opt in to make their product agent-interactable. Agents navigate any UI without breaking when the design changes. Semantic intent over CSS selectors. Same cleaning engine that powers Read and Serve, applied to action — because the engine that extracts structure for reading is the same engine an agent needs to understand a page well enough to act on it. This is the layer no one else is building. Browserbase and Playwright chase pixel coordinates and CSS selectors. They break when the UI moves. Onto doesn't, because Read and Serve already built the semantic foundation Act needs. 2027 01 — The widening gap ### Optimised for humans. _Invisible to machines._ Modern web applications ship hundreds of KBs of structural noise. Inside that payload, only a few kilobytes of business- critical content actually matter for AI agents. When crawlers ingest these pages, their token budgets burn on layout divs and utility classes. The semantic signal drowns in visual noise. The same problem hits AI developers from the other side: every agent calling the web pays 10 to 100× more in tokens than necessary, on every request. Today, optimising for humans is no longer the same as optimising for visibility. Two distinct audiences, two distinct formats — and most sites are still serving only one of them. Signal detection ~2% Noise ratio Extreme 95%+ Critical of typical page bytes are invisible to AI understanding 10–100× Achievable payload reduction possible with clean Markdown extraction 3× Trending increase in AI agent traffic to content sites in the past year $0 Null what most sites invest in AI optimization today 02 — Design philosophy ### Ontology: the study of _being_, abstracted from appearance. Onto borrows its name from ontology. Just as ontology asks what something fundamentally is rather than how it looks, Onto separates what your website is from how it appears. > The same content, faithfully served in two formats. We don't create a shadow web for bots. We faithfully represent the underlying substance — HTML for humans, Markdown for agents, action protocols for autonomous interaction. Onto design philosophy 03 — What's at risk ### What happens without _AI optimisation._ Four concrete failure modes — three for site owners, one for AI developers reading the web inside their product. All four are already happening across the web today. Risk · 01 #### Hallucinated facts — attributed to your brand AI models fill gaps with plausible fiction. Wrong prices, phantom features — all presented as official info. You don't find out until a deal falls through. Fail state"Product costs $29/mo" → actual: $49/mo Risk · 02 #### Invisible in AI search results AI search engines can't extract your value prop from React noise. Your competitors with cleaner markup get cited. You don't. This is happening right now. Fail statePerplexity cites competitor's clean docs, skips your heavy site Risk · 03 #### Token cost inflation for customers Every API call burns 10–100× more tokens than necessary. Your customers pay more for worse AI integrations of your content. You become the expensive option. Fail state596 KB HTML → 148K tokens per query → $1.48/call Risk · 04 #### Burning your customers' token budgets If you're an AI developer reading the web in your product, every call to a noisy site costs 10–100× more tokens than it should. Your customers feel the price. Your unit economics suffer. Compatibility layers like Onto's Read API exist because the cost gravity of dirty data is real and unavoidable. Fail stateAgent pipeline: $2.30 / query vs $0.04 / query with clean input For AI developers #### Building an agent? Cut your token bill by an order of magnitude. One API call. Clean Markdown plus accuracy score returned. [Read the API docs](https://docs.buildonto.dev/api/read) For site owners #### Don't wait for the shift. See where your site stands right now. It takes 10 seconds. [Get your score](/scanner) [Read the technical specs](/how-it-works)