Citizen Answer Engine Optimization

Ontario's public services are invisible
to the AI citizens ask first.

We scanned ten Ontario government and health websites for basic AI-agent readiness. Not one cleared Level 1 of 5. Six scored zero, including the province's own health agency, its digital front door, and the education portal.

Summarize with AI: ChatGPT Claude Perplexity Google AI Grok
0/10
cleared Level 1
6
scored Level 0
0/8
any agent capability
21
signals scanned
52%
of Canadians use AI for health
The thesis

Citizens now ask AI first, and the government doesn't get to opt out of being the answer. It only gets to choose whether it's the source. Will Falk argued that every public institution must make its information machine-retrievable. This report is the field test of whether Ontario has. It hasn't.

Prepared by Mahmoud Halat, Space & Story July 2026 Companion field test to A Complement, Not a Substitute (CSA, Will Falk)
The Argument
The front door moved
01

The front door to government moved. Nobody told the buildings.

Fifty-two percent of Canadians now use AI for health information, and only twenty-seven percent trust it. Two hundred and thirty million people a week ask ChatGPT a health question. When a citizen asks an AI how to appeal an eviction, whether they qualify for a benefit, or what a drug does, the AI answers from whatever it can read on the open web.

The authoritative answer usually lives on a government website. But if that page is unreadable to machines, the AI can't retrieve it or cite it, and fills the gap with something worse: an outdated cache, a foreign source, or a confident guess. The government doesn't get to opt out of being the answer. It only gets to choose whether it's the source.

A large share of what people ask AI is bureaucratic, not clinical. Am I eligible? Which form? Where do I go? What changed this year? These are the questions government websites are supposed to answer, and increasingly the ones people route around them to ask a chatbot instead. In Canada, 9% have already used AI to manage finances or file taxes.

52%
used AI for health information; 48% for treatment advice (CMA, 2026)
27%
trust AI to give accurate health information
230M
people/week ask ChatGPT a health question (OpenAI, 2026)
~6%
of health questions are pure navigation: benefits, forms, paperwork
My GenAI is the front door to my health system.Will Falk, "Patients are doing it for themselves"

Will's CSA paper argues that in Canadian healthcare, inaction is now the greater risk. His companion essay makes the corollary explicit: "Every public institution has the counterpart obligation: make your information RAG-ready. Structured, citable, machine-retrievable." He made the normative case. So we tested Ontario's front door.

The Scorecard
The report card
02

Six levels of readiness. Ontario reaches rung one, at best.

We ran the full scan, every one of the 21 signals the tool checks across five categories, and placed each site on its 0-to-5 capability ladder. Each rung adds a concrete capability, and the requirements are the tool's own. Read bottom-up, it tells a capability story: discoverable, then declares its terms, then readable by agents, then callable by them, and finally able to identify them.

5Agent-NativeAdds agent identity and auth (Auth.md, A2A Agent Card, OAuth metadata). Agents can securely register, authenticate, and transact.Reached by
This very report ↓
4Agent-IntegratedAdds the capability layer (API Catalog, MCP Server Card, Agent Skills, WebMCP). An agent can discover and call your tools, not just read them.No Ontario site
3Agent-ReadableAdds Markdown content negotiation. An agent retrieves clean, structured content to ground its answer in. The RAG layer.No Ontario site
2Bot-AwareAdds AI-bot rules and Content Signals. The site declares how AI may use its content.No Ontario site
1Basic Web Presencerobots.txt and a sitemap. A classic search crawler can find it.4 sites + benchmark
Public Health ON · UHN · SickKids · Ontario.ca · Canada.ca
0Not ReadyMissing even the basic crawlable presence.6 sites
Ontario Health · Health811 · Tribunals · OnLand · Curriculum · Open Data
SiteLevel Discoverability
of 4
Content
of 1
Bot control
of 2
Capabilities
MCP·Skills·API·Auth (of 8)
Ontario Healthontariohealth.ca0 · Not Ready0/40/10/20/8
Health811health811.ontario.ca0 · Not Ready0/40/10/20/8
Tribunals Ontariotribunalsontario.ca0 · Not Ready1/40/10/20/8
OnLand, land registryonland.ca0 · Not Ready0/40/10/20/8
Ontario Curriculumdcp.edu.gov.on.ca/en0 · Not Ready0/40/10/20/8
Ontario Open Datadata.ontario.ca0 · Not Ready1/40/11/20/8
Public Health Ontariopublichealthontario.ca1 · Basic2/40/11/20/8
University Health Networkuhn.ca1 · Basic2/40/11/20/8
SickKidssickkids.ca1 · Basic2/40/11/20/8
Ontario.ca, main portalontario.ca · 43/1001 · Basic2/40/11/20/8
Canada.cafederal benchmark1 · Basic2/40/11/20/8
full partial none · Full scan, all 21 signals · Canada.ca is a federal benchmark, not in the Ontario ten · commerce signals n/a for information sites

What the scorecard says out loud

The gap is will, not technology. The scanner's own site scores Level 5. Getting there is configuration and content work, not a moonshot or a procurement mega-project. That is exactly why a report card is the right instrument: it turns an invisible gap into a number a deputy minister can be asked about.

This report leads by example Level 5 · Agent-Native

Point an AI agent at this report and it can read it, cite it, and call it directly.

Rather than scold from a Level 0 page, it is built to the exact standard it measures. Everything Ontario's sites are missing, this page ships live. Click any capability to check it for yourself, then see the scorecard for the contrast.

Also embedded in the page: schema.org structured data and WebMCP in-browser tools. Scan it yourself ↗

What the scan measures: 21 signals across 5 categories

Nothing is hidden. The first three categories are what a content site can satisfy today; the capability layer is where "can an agent act on this service?" is decided, and where Ontario is uniformly empty.

Capabilities: MCP, Skills, Auth0 of 8 · every site

  • MCP Server Card: advertise an MCP server's tools
  • Agent Skills: publish installable workflows
  • WebMCP: in-page tools an agent can call
  • API Catalog: list your APIs (RFC 9727)
  • OAuth discovery + Protected Resource: how to authenticate (RFC 8414 / 9728)
  • Auth.md: agent registration instructions
  • A2A Agent Card: agent-to-agent identity

Discoverability0–2 of 4

  • robots.txt: crawl rules (RFC 9309)
  • sitemap: a machine list of your URLs
  • Link headers: HTTP pointers to resources (RFC 8288)
  • DNS-AID: DNS-based agent discovery

Bot access control0–1 of 2

  • AI bot rules: directives for GPTBot, ClaudeBot
  • Content Signals: declare ai-train / ai-input / search use (contentsignals.org)
  • Web Bot Auth: cryptographic bot verification (optional)

Content0 of 1

  • Markdown negotiation: serve clean text when an agent sends Accept: text/markdown

Commercen/a for info sites

  • x402 · MPP · UCP · ACP · AP2: agent payment and checkout protocols. Optional, and not relevant to an information site.

Method: Cloudflare's isitagentready.com API, full scan across all 21 signals in 5 categories, homepage-level, 2026-07-21. The 0–5 levels and their requirements are the tool's own, reconstructed from its published gating logic. Web Bot Auth, DNS-AID and the commerce protocols are tracked but do not gate a level. The tool grades technical delivery, not content quality, and several signals are Cloudflare-authored, so we show the full per-category breakdown rather than a single number. The pattern holds on the vendor-neutral basics too.

What's at Stake
Why a blank column is dangerous
03

When the real source is unreadable, the machine answers anyway, and it's often wrong.

A missing sitemap sounds like a technicality. The consequence isn't. An AI asked a government question will produce an answer whether or not it could reach the government. The only variable is quality.

Will's paper names the deeper stake: sovereignty. If Canadian authoritative sources are invisible, AI grounds Canadians' answers in whatever it can read, which skews American and commercial. As he puts it, "we can no longer rely on a search of US assets like the CDC to give us trustworthy, Canadian-relevant answers." Agent-readiness is how a Canadian source stays in the answer. It's a safety issue, an equity issue, and a digital-sovereignty issue at once.

We've run this play before
04

This is the web-accessibility fight, for a new kind of reader.

Twenty years ago, the web was unusable for people with disabilities. We didn't leave it to goodwill. A standard was written (WCAG), a law adopted it (Ontario's Accessibility for Ontarians with Disabilities Act, the AODA, required WCAG 2.0 AA on public sites by 2021, with penalties on the books), and "is this site accessible?" became a question institutions had to answer.

Agent-readiness is the same move for a new class of reader: the machine acting on a citizen's behalf. The precedent isn't hypothetical. It's already forming.

Published scorecards convert principle into action: the UN e-Government Index and the OECD's open-data index demonstrably push governments to chase a rank. What's missing is the standard itself, and the body to own it.

What "agent-ready" means
05

Three layers, from cheapest to most capable.

"Agent-ready" isn't one thing, and conflating the layers is how these conversations go wrong. Here they are, cheapest first, each against one government example: the GST/HST credit.

RAG

Be readable & citable

The AI reads your published page and grounds its answer in it, with a citation. You just publish clean, structured, machine-retrievable content. Read-only. This is 80% of the value and nearly all of what the report card measures.

"Am I eligible for the GST/HST credit?" → the AI cites the canada.ca eligibility page instead of guessing.
MCP

Be queryable

You expose typed tools an agent can call live, with authentication, so it gets an official answer instead of a paraphrase. This is the layer people mean by an "MCP (Model Context Protocol) server."

CRA exposes estimate_credit(income, family_size, province) → returns an authoritative dollar figure.
SKILLS

Be operable

You publish a packaged, procedurally-constrained workflow that teaches an agent the correct steps, forms, and edge cases, so it does the task right every time instead of improvising.

A CRA "file-benefit-application" skill encodes the exact sequence, documents, and field formats.

RAG means the model knows your content. MCP means it can query your systems. Skills mean it knows the correct procedure. A content site needs Layer 1. A service needs all three. None of it is exotic. Ontario is simply at zero on all three.

The Play
The play for CSA
06

Don't just audit the gap. Own the standard that closes it.

Anyone can run a scan. The durable, mission-aligned move for a standards body is bigger: define what "agent-ready public information" means in Canada (the WCAG for AI consumers), and let the report card be the evidence the standard is needed.

The slot is open. ISO, NIST, and W3C are busy on AI governance and risk; W3C is standardizing agent interaction. No standards body has defined the machine-retrievability of public information. CSA already runs Canada's AI-standards and accessibility-standards lane, and it adopted ISO/IEC 42001 nationally. This is the same instrument pointed at a new and urgent target.

Why it fits the Public Policy Centre, specifically

It lands on two of the Centre's four stated priorities at once: supporting public-sector innovation and enhancing the health and wellbeing of Canadians. And it operationalizes the exact thesis of A Complement, Not a Substitute: the risk now sits in moving too slowly. A standard is how you move decisively with guardrails.

The name follows: Citizen Answer Engine Optimization. It makes the machine's front door open onto the real, Canadian, authoritative source: by design, to a published standard, measured on a report card.

Reference
Plain-language glossary
§

Every term in this report, defined.

A quick reference for anyone who doesn't live in this vocabulary. Read it once and the rest of the report goes faster.

The core ideas

Answer engine
An AI system that answers a question directly instead of returning a list of links. ChatGPT, Claude, Perplexity, Gemini, and Google's AI Overviews are answer engines.
AI agent
Software that acts on a person's behalf. It reads, decides, and takes steps such as searching, retrieving a fact, or filling in a form, with little or no human clicking.
AEOAnswer Engine Optimization
Making your content easy for answer engines and agents to find, read, trust, and cite, the way SEO made content easy for search engines. Also called GEO, for Generative Engine Optimization.
Agent-ready
A website an AI agent can find, read cleanly, and act on. This report scores it on a 0-to-5 scale.
RAGRetrieval-Augmented Generation
When an AI pulls in your published page and bases its answer on that source, ideally with a citation, rather than on memory. RAG-ready content is structured, current, and machine-retrievable.
MCPModel Context Protocol
An open standard that lets a website expose tools and data an agent can call live and securely, returning an official answer such as a dollar figure or a status, instead of a paraphrase.
Agent Skill
A packaged, repeatable workflow (the steps, forms, and rules for a task) that a site publishes so an agent completes it correctly every time instead of improvising. Factual, not generative.
Readiness levels (0–5)
The capability scale used here: 0 Not Ready, 1 Basic Web Presence, 2 Bot-Aware, 3 Agent-Readable, 4 Agent-Integrated, 5 Agent-Native. Each level adds one capability.

The readiness signals (what the scan checks)

robots.txt
A plain-text file at a site's root that tells crawlers and agents which parts of the site they may access.
Sitemap
A machine-readable list of a site's pages, so an agent can find all of it, not only what happens to be linked.
Link headers
HTTP response headers (RFC 8288) that point an agent to related resources, such as an API catalog or documentation, without it having to parse the page.
DNS-AIDDNS for AI Discovery
An emerging standard for advertising a site's agent endpoints through DNS, so agents can discover them before they load a single page.
Markdown content negotiation
Serving a clean text version of a page when an agent asks for it, stripping the visual clutter so the AI reads it accurately and cheaply.
AI bot rules
Entries in robots.txt aimed at specific AI crawlers such as GPTBot, ClaudeBot, and Google-Extended, setting what each is allowed to do.
Content Signals
A robots.txt extension, led by Cloudflare, that states how AI may use your content: for search, as an answer input, or for training.
Web Bot Auth
A way for an agent to cryptographically prove its identity, so a site can wave through trusted agents and block impostors.
MCP Server Card
A published file that advertises a site's MCP server and the tools an agent can call there.
Agent Skills index
A published list of the skills, or workflows, a site offers to agents.
WebMCP
A way to expose agent-callable tools directly inside a web page.
API Catalog
A standard table of contents for a site's APIs (RFC 9727), so an agent can discover how to call them.
OAuth discovery & Protected Resource
Standard metadata (RFC 8414 and RFC 9728) that tells an agent how to authenticate and which authority guards a given resource.
Auth.md
A proposed convention that describes, in one file, how agents register and authenticate with a service.
A2AAgent Card
A published identity file used for agent-to-agent communication.
Agent commerce protocols
Emerging standards (x402, MPP, UCP, ACP, AP2) that let agents pay and check out on a user's behalf. Not relevant to information-only government sites, so they are excluded from the score.

Policy, standards & institutions

CSACanadian Standards Association
CSA Group, a century-old Canadian standards body. Its Public Policy Centre published Will Falk's paper, and it is the natural home for an agent-ready public-information standard.
WCAGWeb Content Accessibility Guidelines
The international standard, authored by the W3C, for making websites usable by people with disabilities. The model this report borrows.
AODAAccessibility for Ontarians with Disabilities Act
Ontario's 2005 law that made WCAG conformance a legal requirement for public-sector websites, with penalties on the books.
21st Century IDEA Act
A 2018 US law requiring federal websites to publish machine-readable content and rich, descriptive metadata.
SaMDSoftware as a Medical Device
Health Canada's category for software regulated as a medical device. Whether a tool falls into it turns on its represented intended use, not its technology.
W3CWorld Wide Web Consortium
The body that authored WCAG and is now standardizing agent interaction (WebMCP). The template for how a web standard becomes law worldwide.
ISO/IEC 42001
The international AI management-system standard, which CSA has adopted nationally, evidence that CSA already runs Canada's AI-standards lane.
Method & sources
·

Where the numbers come from.

Scan data. Cloudflare's isitagentready.com API, full scan across 21 signals, homepage-level, 2026-07-21. Levels 0–5, reconstructed from the tool's own gating logic. See the Cloudflare Agent Readiness announcement.

Canadian AI-health usage. CMA 2026 Health & Media Tracking Survey (Abacus Data, n=5,000, Feb 2026): 52% / 48% / 27%.

Global usage. OpenAI, "Introducing ChatGPT Health" (Jan 2026): 230M/week. Microsoft Research + Nature Health: 1.7M conversations, 109 countries.

Government AI errors. UK Open Data Institute chatbot study (Feb 2026); Moffatt v. Air Canada (BC Civil Resolution Tribunal, 2024); Globe & Mail CRA "Charlie" test (2026).

Accessibility precedent. Ontario AODA / WCAG 2.0 AA (2021); US 21st Century IDEA Act (2018); UK GDS "datasets ready for AI" (Jan 2026).

Standards landscape. ISO/IEC 42001, 23894, 8183; W3C WebMCP (Sept 2025); NIST AI RMF. No body yet defines machine-retrievability of public information.

Lead by example. This page implements the signals it measures: a Markdown rendering, an MCP server, and the .well-known capability files. Scan this report yourself ↗.

Economics. 2026 frontier-model API pricing; web-accessibility remediation benchmarks; SecurityScorecard / BitSight as the scorecard-to-services model.

Prepared by Space & Story as a companion field test to A Complement, Not a Substitute: Generative AI's Role in Canadian Healthcare in 2026 (CSA Public Policy Centre, Will Falk). Findings reflect a point-in-time scan of public homepages; scores change the moment a site is updated.