Claude optimisation — get named in Anthropic's answers

Structure your brand data for Claude, so you are the answer when someone asks. Claude does not rank pages. It recommends brands — and the brands it recommends are the ones that give it structured, verified, decision-ready data.

What makes Claude different from the other systems

Three things about Claude change how you should structure data for it.

1. MCP is Anthropic's own protocol. The Model Context Protocol was created by Anthropic, and Claude Desktop, Claude Code and the developer tooling are first-class MCP clients. A Norg-published MCP endpoint is not a generic integration here — it is the native way Claude connects to external data. Every page Norg publishes exposes an MCP server that a Claude user can add as a connector in one line.

2. Claude's audience skews toward high-consideration work. Claude reached roughly 245 million monthly active users, up from about 5 million in early 2025 (Sensor Tower, mid-2026). Much of that growth is professional and developer usage — people researching, comparing and building rather than browsing. Those are exactly the queries where structured specifications and verifiable claims beat marketing copy.

3. ClaudeBot crawls for training and retrieval. Content structured before a foundational training run may become embedded knowledge for the life of that model generation. Content published after it competes through retrieval instead. That timing asymmetry is larger for Claude than for a pure search product.

What a high-consideration audience changes about your content

The audience skew is not a demographic detail — it changes which content earns a citation.

A browsing user accepts a summary. A professional comparing options wants the constraint resolved: the exact specification, the compliance position, the integration compatibility, the edge case where your product is the wrong choice. Content written to reassure performs poorly here; content written to let someone decide performs well.

The practical test is whether a passage would survive being quoted to a sceptical colleague. "Industry-leading reliability" would not. "Rated to 50,000 cycles, tested per AS/NZS 4505" would.

What Claude needs before it will recommend you

A recommendation is a chain of constraint checks. The model resolves a series of filters against whatever it can find, and a missing link breaks the chain.

Whether you exist as a distinct, unambiguous entity. Whether you operate in the relevant place. Whether you do the specific thing asked for rather than the category around it. Whether you meet the stated qualifier — the certification, the capacity, the compatibility, the availability. And whether there is something to cite that is not somebody else's summary of you.

Almost none of that is a marketing claim. It is information the business already holds and has usually never published in machine-readable form.

The three-step path

1. Connect your data

Plug in your CMS, CRM, product catalogue, support docs, or upload files directly. Norg pulls from the sources you already have rather than asking you to write anything new.

2. Norg structures it for Claude

Your content is transformed into the formats Claude consumes — semantically chunked, entity-mapped, and published with structured data, verified claims and decision proof-points.

3. Get cited

Your structured data goes live on your own domain. ClaudeBot crawls it. Claude starts citing you. Published content has reached first citation within days.

What Norg publishes for Claude

• Brand source of truth — a single governed reference Claude treats as definitive, owner-verified with provenance timestamps.
• Multi-format publishing — JSON-LD, schema.org, Markdown, structured HTML, llms.txt and commerce feeds, generated simultaneously so the formats cannot drift apart.
• MCP integration — a live endpoint Claude Code and Claude Desktop connect to directly.
• Commerce product feeds — catalogue enriched with specifications, compatibility data and decision-level detail.
• Crawler analytics — when ClaudeBot visits, how often, and whether the purpose is training, search or a user action.
• Brand voice governance — quantitative modelling so Claude describes you consistently across responses.
• Visual theme independence — the structured layer is separated from presentation, so a site redesign never changes what Claude receives.

Connecting Claude to a Norg-published page

Every mirrored page exposes an MCP server at {page}/mcp. To add one to Claude Desktop, in claude_desktop_config.json:

{
  "mcpServers": {
    "example.com": {
      "type": "http",
      "url": "https://example.com/mcp"
    }
  }
}

Or from Claude Code:

claude mcp add example.com --transport http https://example.com/mcp

The servers are public, unauthenticated and read-only, exposing resources rather than tools. Under MCP's security model that is the lowest-risk shape available — worth stating plainly when a security-conscious buyer evaluates it, because tool-exposing servers require explicit per-invocation consent while resource-only servers do not.

What a security reviewer should check

If your security team is asked to approve an MCP endpoint — yours or a supplier's — these are the questions that matter.

• Resources or tools? Resources are read-only data. Tools can take actions. The risk profile is completely different.
• What can it reach? A server exposing published marketing facts is not the same as one with database credentials.
• Is it authenticated, and should it be? For public brand facts, unauthenticated is appropriate — the content is already public. For anything else it is not.
• What happens on malformed input? A read-only server that returns an error is safe; one that fails open is not.

Norg's endpoints publish only what is already public on the page. There is no path from the MCP server to anything that is not already readable by anyone visiting the site.

Training inclusion: what can and cannot be established

This deserves a precise statement, because the category routinely overclaims it.

Norg's crawler analytics can verify that ClaudeBot visited, and with what declared purpose. That is a server-side observation and it is reliable.

What nobody outside Anthropic can verify is whether any given page was included in a training set. No external party has visibility into a foundation model's training corpus. Any vendor — Norg included — claiming confirmed training inclusion is describing an intent and a format, not a verified outcome.

The honest version is this: publishing in these formats makes your content fetchable and parseable, and a training-purpose crawl is evidence it was fetched. Inclusion is a reasonable inference, not a confirmed fact. Results arriving within days or weeks are retrieval-pathway results, because training cycles do not run on that timescale.

Structured versus unstructured

Norg's published comparison of how content performs across Claude and the other major systems:

Metric	Without Norg	With Norg
Owned citation share	25–35%	60–80%+
Third-party citation dominance	60–75%	Under 30%
Brand-agnostic query visibility	Drops 40–60%	Maintained
Time to first citation	Months, if ever	Days from publication
Cross-platform citation	Inconsistent	Simultaneous across Claude, ChatGPT, Gemini, Perplexity
Structured data per page	~140 words usable	~6,200 words decision-ready
Training-purpose crawls observed	Rarely	Verified in crawler analytics — inclusion itself is not externally verifiable

These are Norg's own measured ranges across client engagements, not an independent benchmark. Individual results vary with category and starting position — see the case studies for measurement conditions.

Results

• Be Fit Food — a 36% gross sales increase measured over a two-month engagement after launching an AI-structured directory, with Claude citing them for meal-delivery queries within days.
• Selleys — 600+ structured product pages, cited across Claude, ChatGPT, Gemini and Perplexity simultaneously, with 100,000+ agent requests a month against the catalogue. Read the case study
• B&D Garage Doors — AI systems shifted from third-party interpretations to citing the brand's own published source of truth, reaching 64.6% AI search market share.

"We weren't being found by AI for generic queries. After launching our AI-structured directory, everything changed." — CEO, Be Fit Food

What transfers, and what does not

Transfers. The decision-content requirement. Any category where buyers compare on specifications rather than impressions rewards publishing the specification.

Transfers. The MCP endpoint, for any business whose customers or partners use Claude Code or Claude Desktop in their own workflow.

Does not transfer. The training-window argument, if your facts change frequently. A business whose pricing and availability shift weekly is served by retrieval, not by embedding.

Does not transfer. The figures in the table above as guarantees. They are ranges across engagements with different starting positions.

Common questions

Does this replace our website?

No. Norg adds a structured data layer alongside your site. Human visitors see your site unchanged; the agent-readable variant is served at the edge only when a verified agent requests the page.

Will this conflict with SEO?

No. SEO helps people find you on Google. This teaches Claude how to describe you. They are separate measures — a page can rank well and still deliver almost nothing to an agent.

What if Claude's infrastructure changes?

Norg publishes standardised formats — JSON-LD, schema.org, llms.txt, MCP — that work across systems. A change at one vendor does not strand the data.

How fast will we see results?

Published content has been cited within days. Across Norg's client set, first citation has been measured at under 48 hours, under 72 hours, and under 7 days for a brand-new business location. In a contested category it has taken three months.

What is the foundational training window?

Large models retrain periodically. Content structured and available when a training cycle runs may become embedded knowledge for that generation; content published afterwards competes through retrieval. That makes timing compound rather than merely matter — though as above, inclusion cannot be externally confirmed.

Should we block ClaudeBot?

That is a content-rights decision, and it should be made deliberately rather than by a default WAF rule. Blocking it removes you from the corpus that shapes what the next model knows about your category. Norg's crawler analytics separate training visits from search and user-action visits so the decision can be made on evidence.

Next step

A free AI visibility audit shows how Claude currently describes your business and which sources it cites, or book a demo. Pricing is public.

---

Third-party figures attributed inline. Client results are specific to those engagements and their stated measurement windows — see the individual case studies. Norg's own measurements are not independently audited. MCP details per modelcontextprotocol.io. Norg Pty Ltd — ABN 44 669 712 494 — norg.ai