# How to Get Cited by Claude: The Anthropic Retrieval Playbook

**Author:** John Morabito (Founder, /winston)
**Published:** June 14, 2026
**Reading time:** 12 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/how-to-get-cited-by-claude/

Most GEO advice is written for ChatGPT and Google, and it quietly assumes every AI engine works the same way. Claude does not. It retrieves through its own web search tool, through MCP servers, and through connectors inside Claude for Work, which changes what it takes to be the source it reaches for. Here is how Claude finds and cites pages, what is different from the other engines, and the checklist for becoming citable.

**Getting cited by Claude comes down to being easy to retrieve and easy to read.** Claude reaches your content through its web search tool, through MCP servers, and through connectors in Claude for Work, then it cites the sources it can parse cleanly and trust. There is no published ranking algorithm, so the honest play is to reduce retrieval friction (clean server-rendered HTML, markdown twins, connected schema) and lead every page with a direct, liftable answer. Do that and you are the source Claude names, whether the question comes from public web search or a private workspace.

## How Claude actually retrieves and cites

Claude does not have a search index of its own the way Google does. It reaches the live web and your systems through a set of tools, and understanding those tools is most of the strategy.

- **The web search tool.** When a question needs current information, Claude issues search queries, gets back a set of results, fetches the pages it wants to read, and attributes the sources it leaned on. This is the closest thing to classic AI-search citation, and it rewards the same discipline: be retrievable and be the clearest answer on the page.
- **MCP servers.** The Model Context Protocol lets a site or app expose structured, machine-readable content and actions that Claude can call directly. Instead of scraping your front end, Claude asks your MCP server for exactly the data it needs. If you run one, your content is retrievable without any parsing guesswork. Build guide: https://www.winstondigitalmarketing.com/playbooks/mcp-servers-for-marketing-teams/
- **Connectors and project knowledge.** Inside Claude and Claude for Work, users connect drives, tools, and documents, and they load reference material into project knowledge. Your content can be the cited source in those contexts without ever appearing in a public search result.

The through line: Claude cites what its tools can reach and read cleanly. Every tactic below is a way to make one of those paths smoother.

## What differs from ChatGPT and Perplexity

The fundamentals of generative engine optimization port across every engine. Answer-first structure, chunked sections, corroboration, and clean markup win citations everywhere. The retrieval plumbing is where Claude diverges, and that difference is worth building for.

ChatGPT search has historically drawn on Bing's index, which means a lot of ChatGPT optimization is really Bing SEO (https://www.winstondigitalmarketing.com/playbooks/bing-copilot-optimization/). Perplexity runs its own crawler and leans hard on freshness and dense citation (https://www.winstondigitalmarketing.com/playbooks/how-to-rank-in-perplexity/). Claude, by contrast, retrieves through its own web search tool plus a growing surface of MCP servers and connectors that hand it your content directly. So being reachable to Claude is partly a content job and partly an integration job. If you already do clean GEO for the other engines, you have done most of the content work. The extra lever is exposing structured versions of your content that Claude can pull without fighting your front end.

The one-line version: ChatGPT optimization is mostly Bing SEO. Perplexity optimization is mostly freshness and citation density. Claude optimization is clean retrieval plus integration: give its tools a frictionless path to your content and a plainly stated answer once they get there.

## Why clean server-rendered HTML matters to Claude

When Claude fetches a page through its web search tool, it works from what the fetch returns. If your content only appears after JavaScript executes, or it is buried under layers of layout, navigation, and interstitials, the model has to work harder to find the actual answer, and sometimes it does not find it at all. A page that renders its content in the initial HTML response, with the answer present in the markup rather than assembled client-side, is one the model can read on the first pass.

This is not a Claude quirk. It is true of every retrieval-based engine, and it is why we build this site as static, server-rendered HTML. The agentic web transformation (https://www.winstondigitalmarketing.com/services/ai-marketing/agentic-web/) we run for clients and on our own site is largely about making content reachable by machines, not just browsers. If a human needs your JavaScript to see the page, Claude probably needs it too, and it does not always run it.

## Why markdown twins matter

A markdown twin is a clean, text-only copy of a page served at a stable URL, stripped of navigation, scripts, and visual chrome. The point is to hand any retrieval system the content and nothing else. When Claude fetches a page, the less it has to strip away to reach the answer, the more reliably it reads and attributes it. Serving a markdown version to bots is a low-cost way to guarantee the model gets the clean version.

Two honest caveats. First, a markdown twin is not a secret ranking signal, and Anthropic has confirmed no such signal exists. It is a friction reducer, which is valuable precisely because retrieval is the whole game. Second, twins have an upkeep cost: if you edit the HTML page and forget the twin, the two drift apart and Claude may cite a stale version. Sync both on every edit, or automate the twin so it regenerates from the source. Our approach to serving markdown to AI bots: https://www.winstondigitalmarketing.com/playbooks/llms-txt-guide/

## How Claude weights source clarity and structure

Once Claude has your page in hand, the citation decision is the same judgment every strong AI engine makes: is this the clearest, most trustworthy, most liftable answer available? A few patterns move that decision in your favor.

- **Lead with the answer.** Open the page, and ideally each section, with a direct and complete response to the question, then expand. Claude lifts the tight answer and attributes it. A page that buries the answer under 600 words of preamble loses to one that states it up top.
- **Chunk by question.** One H2 per real question, answered completely in roughly 100 to 150 words, so a section can be extracted whole without losing meaning. Rubric: https://www.winstondigitalmarketing.com/playbooks/writing-faqs-ai-engines-cite/
- **State claims plainly and let them be corroborated.** Claude leans toward sources it can verify. Definite statements, named specifics, and claims that show up confirmed elsewhere read as more trustworthy than hedged generalities.
- **Confirm what you are with connected schema.** An Organization and Person graph joined by stable `@id` references, with Article and FAQPage on the post, removes ambiguity about who is making the claim. Pattern with examples: https://www.winstondigitalmarketing.com/playbooks/schema-markup-for-ai-engines-2026/

| Retrieval path | What makes you citable | Where it lives |
|---|---|---|
| Web search tool | Server-rendered HTML, answer-first content, markdown twin | Public web |
| MCP server | Structured, machine-readable content and data endpoints | Your infrastructure |
| Connectors | Documents and drives connected to a workspace | Claude for Work |
| Project knowledge | Reference material a user loads into a project | Inside a Claude project |

## Being citable inside Claude for Work and MCP contexts

Public web search is only one surface. Inside Claude for Work, a team can connect Claude to its own documents, drives, and internal tools, and it can load reference material into project knowledge. Your content can be surfaced and cited there without ever touching a public search result. This matters most if you publish documentation, a knowledge base, or a data source your customers actually use.

The move is to expose that content through an MCP server or a supported connector so Claude can retrieve it directly in those workspaces. When a customer's team asks Claude a question your material answers, you want your content to be the source it reaches for. That is a different kind of citation than a SERP placement, and for many businesses it is the more valuable one, because it lands at the moment of real work. The decision of what to build, a Skill, a tool, or an MCP server, is the subject of: https://www.winstondigitalmarketing.com/playbooks/anthropic-skills-vs-openai-tool-use/

## The honest part: there is no confirmed algorithm

Anthropic has not published a ranking or citation algorithm for Claude. Everything above is observed behavior across real retrievals, not a documented formula, and anyone selling a guaranteed method is guessing. That is not a reason to do nothing. It is a reason to invest only in work that would help any retrieval system: content a machine can reach, an answer it can lift, and signals it can verify. Those hold up as the tools evolve, which they will. Optimize for retrieval friction and clarity, not for a secret score, and you never have to unwind a tactic when the plumbing changes.

## The practical checklist

1. **Server-render the content.** The answer should be in the initial HTML, not assembled by JavaScript after load. If a bot with scripts off cannot read it, fix that first.
2. **Lead every page and section with the answer.** Direct response up top, expansion below. Chunk one question per H2 in 100 to 150 words.
3. **Ship markdown twins for priority pages** and keep them synced with the HTML, or automate them so they never drift.
4. **Connect the schema graph.** Organization, Person, Article, and FAQPage joined by stable `@id` references, validated and server-rendered.
5. **Expose an MCP server or connector** for the documentation or data your customers use inside Claude, so you are citable in private workspaces, not just public search.
6. **Spot-check Claude monthly** on your top 20 questions to see who it cites. When it is not you, that gap list is your content calendar.

## Frequently asked questions

**How does Claude decide what to cite?**
Anthropic has not published a ranking algorithm, so this is observed behavior rather than a documented formula. When Claude uses its web search tool, it issues queries, reads the pages the search returns, and attributes the sources it actually leaned on in the answer. In practice the pages that get cited are the ones that are easy to retrieve and easy to read: server-rendered HTML the fetch can parse without JavaScript, a direct answer near the top, sections chunked one question at a time, and schema that confirms what the page is. Claude leans toward sources it can verify and that state their claims plainly, so clarity and structure do more work than keyword density.

**Is optimizing for Claude different from optimizing for ChatGPT?**
The foundations overlap, the retrieval paths differ. Both reward answer-first, chunked, well-structured content that a model can lift and attribute. The difference is the plumbing: ChatGPT search has historically drawn on Bing's index, while Claude retrieves through its own web search tool and, increasingly, through MCP servers and connectors that hand it your content directly. That means being reachable to Claude is partly a content job and partly an integration job. If you already do clean GEO for ChatGPT and Perplexity, most of it ports to Claude. The extra lever is exposing structured, machine-readable versions of your content that Claude can pull without wrestling with your front end.

**Do markdown twins help Claude cite my site?**
They help by removing friction. A markdown twin is a clean, text-only version of a page served at a stable URL, with no navigation, scripts, or layout noise around the content. When Claude fetches a page, the less it has to strip away to reach the actual answer, the more reliably it reads and attributes it. Markdown twins are not a magic ranking signal and Anthropic has confirmed no such signal, but they make your content trivially parseable, which is the whole game for a retrieval system. The catch is upkeep: if you edit the HTML page and forget the twin, the two drift apart and the model may cite a stale version. Sync both or automate the twin.

**Can my content be cited inside Claude for Work or an MCP server?**
Yes, and that is a distinct surface worth building for. Inside Claude for Work, a team can connect Claude to its own documents, drives, and internal tools through connectors and MCP servers, so your content can be surfaced and cited in contexts that never touch public web search. If you publish documentation, a knowledge base, or a data source your customers use, exposing it through an MCP server or a supported connector makes it directly retrievable by Claude in those workspaces. The citation there is not a public SERP placement, it is your content being the source Claude reaches for when someone on that team asks a question your material answers.

**Is there a known Claude ranking algorithm?**
No. Anthropic has not published a ranking or citation algorithm for Claude, and anyone selling you a guaranteed formula is guessing. What we can describe is observed behavior: Claude cites pages its tools can retrieve and read cleanly, it favors sources that state claims plainly and can be corroborated, and it rewards structure that lets it lift a complete answer from one section. Treat every tactic here as a way to reduce retrieval friction and increase clarity, not as a lever on a secret score. That framing keeps you honest and, conveniently, it is also the work that ages well as the tools change.

Service: https://www.winstondigitalmarketing.com/services/generative-engine-optimization/
Audit: https://www.winstondigitalmarketing.com/contact/#audit