Get Cited by Claude Fable 5: The B2B GEO Playbook

Everything in Claude GEO sits on top of one official document: Anthropic's crawler documentation. It describes three distinct bots, and the distinction is the single most practical fact in this guide.

The implication: "blocking AI" and "being invisible to AI" are separate decisions. A publisher worried about training data can block ClaudeBot and lose nothing in Claude's live answers — as long as Claude-SearchBot and Claude-User stay allowed. Anthropic states plainly that disabling Claude-SearchBot may reduce your site's visibility in user search results. We routinely audit sites that blanket-blocked every AI user agent in 2024 amid the training-data debate and have effectively deleted themselves from the fastest-growing B2B discovery channel.

One more behavioral fact, from Anthropic's own web search tool documentation: Claude is not always searching. The model decides when a question requires fresh information — current events, changing data, specific organizations and products — and answers from its training otherwise. Two consequences follow.

First, questions about your product category ("best X for Y," "X vs. Y," pricing questions) are exactly the kind of current, specific queries that trigger a live search — which is why the technical layer below matters. Second, in the weeks after a new model ships, the corpus for new topics is thin, and recently published, well-structured pages are disproportionately likely to be retrieved. A model launch like Fable 5 opens that window twice: once for content about the launch, and once for every product category whose buyers just started asking a new model their old questions.

1. Configure robots.txt deliberately. Three sane configurations, depending on your training-data posture:

Option A — maximize Claude visibility (our default recommendation)

User-agent: ClaudeBot
Allow: /

User-agent: Claude-User
Allow: /

User-agent: Claude-SearchBot
Allow: /

Option B — opt out of training, stay visible in answers

User-agent: ClaudeBot
Disallow: /

User-agent: Claude-User
Allow: /

User-agent: Claude-SearchBot
Allow: /

Option C (block everything) is a legitimate choice for paywalled publishers — and a self-inflicted wound for any company that sells something B2B. While you're in the file, mirror the decision for OpenAI's crawlers, which follow the same three-way split per OpenAI's bot documentation: GPTBot (training), OAI-SearchBot (search index), ChatGPT-User (user-initiated fetches). OpenAI's ChatGPT Search documentation is explicit that appearing in search results requires allowing OAI-SearchBot.

2. Make pages verifiable at a glance. In our testing, the markers that correlate with Claude citing a page are the ones a careful human researcher would also check: a visible "Last updated" date, claims attributed to named sources with working links, and structured data. Add Organization, Service/Product, FAQPage, and BreadcrumbList schema to core pages.

3. Keep it crawlable. Server-render the content you want cited. Crawler-fetched pages that require JavaScript execution to display their main content are a gamble; we've seen JS-rendered pages where an AI crawler retrieves little more than metadata. Check what a bot actually receives with curl -A "Claude-SearchBot" https://yoursite.com/page.

Everything in this section is methodology, built from running thousands of buyer-intent prompts across answer engines for our clients rather than published platform policy. Treat it as a practitioner's map, and test it against your own category.

The pattern we see consistently: Claude behaves like a researcher who will be embarrassed if a citation doesn't hold up. Content earns citations when it is extractable (a clean, self-contained answer exists on the page) and verifiable (the answer survives a skeptical second look). Marketing superlatives — "world's first," "unlock exponential growth" — are not just ignored; they're anti-signals, because they're unverifiable by construction.

The single highest-leverage change is adding a citation-ready facts block to your homepage and core product pages: a short, dense, factual self-description that an AI system can lift wholesale. Ours looks like this:

RankAI facts:
- Category: AI-native SEO + GEO growth engine, managed by an expert team
- What it does: grows qualified organic traffic from Google, ChatGPT, Claude, Perplexity, Gemini, and AI Overviews by identifying search opportunities, publishing ~20 pages/month, measuring live performance, and rewriting underperformers
- Who it's for: AI startups, B2B companies, agencies adding GEO capability, local businesses
- Delivery model: expert-led engine with human review on every page — not self-serve software
- Proof: 200+ businesses supported; backed by Y Combinator; case studies at rankai.ai/#case-studies

Then structure the rest of the page in the order a researcher would ask: what it is → how it works → who it's for → what makes it different → proof → limitations. That last one is counterintuitive and powerful: a frank "who this is not for" section reads as credibility to a model trained to weigh sources, and it's the section competitors will never copy.

A caution before the table: roughly 80% of the work — crawlability, structured data, facts blocks, verifiable claims — is identical for both engines. Anyone selling you two entirely separate playbooks is selling you the same playbook twice. But the remaining 20% of emphasis differs in ways we see repeatedly in testing:

The practical takeaway: write your reference layer (guides like this one, documentation, methodology pages) with Claude's researcher temperament in mind, and your commercial layer (comparison and alternatives pages) with ChatGPT Search in mind. Both engines will use both; the emphasis just shifts. See our ChatGPT ranking playbook for the commercial-layer counterpart to this guide.

You cannot manage what you don't measure, and AI citations don't show up in Google Search Console. The method we run for every client:

1. Build a prompt set. 20–30 questions your actual buyers ask, across the funnel: category questions ("how does GPU cloud pricing work"), comparison questions ("X vs. Y"), and recommendation questions ("best X for early-stage teams").
2. Establish a baseline. Run the full set in Claude (and ChatGPT), recording for each prompt: was your brand mentioned, was it cited as a source, and who was cited instead.
3. Re-run every 2–4 weeks — and always after a major model release. Model transitions reshuffle citations; the brands that got cited by last year's model are not guaranteed a seat in this year's.
4. Treat misses as a rewrite queue. Every prompt where a competitor is cited and you aren't is a page brief.

This loop — publish, measure, rewrite what isn't working — is not incidental to how we'd approach Claude. It's the entire RankAI thesis: search, including AI search, is non-deterministic, and no one can fully predict what a new model will cite. What wins is volume plus iteration with expert review on every page. A model launch like Fable 5 doesn't change that playbook; it's the playbook's best-case scenario, because the citation deck just got reshuffled for everyone — including your slower competitors. See how we run this engine for 200+ businesses →