What the Leaked Claude Fable 5 Prompt Reveals About Ranking in AI Search
The leaked Fable 5 prompt spells out exactly how Claude searches, cites, and quotes. If you want AI search to mention your site, these are the rules it follows.
Horia Stan is a music producer and sound engineer at The One Records in Bucharest who has spent the last year obsessing over how AI search decides what to mention. When Claude Fable 5's system prompt leaked, it handed everyone a document most SEO people would pay for: the actual rules Claude follows when it searches the web and cites sources. I read it with that one question in mind. Here is what it tells you about getting your content into AI answers.
Why this matters now
AI search stopped being a side channel. By the first quarter of 2026, AI assistants like ChatGPT, Claude, Perplexity, and Google's AI answers were handling an estimated 12 to 18 percent of English informational queries. A year earlier it was under 2 percent. The curve is steep, and it does not care how well you rank on the old blue links.
That last stat is the point. A full quarter of the leaked Fable 5 prompt is dedicated to how it searches and how it cites. This is not a footnote in the model. It is a core function, and now the rulebook is public.
When Claude decides to search
Most people imagine the model just knows things. The prompt says otherwise. Claude is told to run a web search whenever a query touches anything that could have changed since its January 2026 knowledge cutoff.
Specifically, it searches for breaking news, for "binary events" like deaths, elections, and incidents, and for present-tense questions about who currently holds a role or position. If someone asks who runs a company right now, Claude is instructed not to trust its training data and to go look.
The practical read for you: if your topic is time-sensitive, freshness is not a nice-to-have. It is the trigger that gets you pulled into the answer at all. Stale content does not get searched for. Current content does.
The rule that changes everything: Claude rewords, never quotes
Here is the single most important line in the prompt for anyone making content. When Claude uses something it found in search, it is told to reword it. Not paraphrase loosely. Reword, so that even short phrases differ from the source. The prompt is explicit that citation tags exist "for attribution, not permission to reproduce original text."
Claude does not quote you. It rewrites you in its own words and links your name. You win by being the clearest thing it can summarize.
This flips the old SEO instinct on its head. You are not writing a quotable sentence and hoping it gets lifted. You are writing the cleanest, most extractable explanation of an idea so that when Claude rewords it, your version is the one it reaches for, and your name is the one it attaches. The copyright limits reinforce this: 15 or more words from a single source is flagged as a severe violation, and one quote per source is the maximum. The model is built to synthesize, not to copy. I traced this rule and the rest in the full Fable 5 prompt breakdown.
How to write for a model that rewords you
If the model summarizes instead of quoting, your job is to be the most summarizable source on your topic. That is a real, specific way to write.
Lead with the direct answer, then explain. A model scanning for the cleanest take grabs the one that states it fastest, not the one that buries it under an intro.
Use the actual phrasing people ask. A clear question followed by a tight answer mirrors how the model retrieves and reassembles information. An FAQ block earns its place.
Write sentences that still make sense pulled out of context, because that is exactly what happens to them. One idea, one source, stated plainly.
Claude synthesizes what it can read. Content stuck behind client-side JavaScript may never reach it. Make sure the words are in the HTML.
The model favors sources it can connect to a clear, real entity. Be a recognizable name on your topic across the web, not an anonymous page chasing a phrase.
Notice that none of this fights the model. It works with how the prompt says it behaves. This is the whole reason I structure the writing on this site the way I do, including the tools I build over on the digital side.
Measure it, because you can
The old SEO loop was rank tracking. The AI search loop is citation tracking, and it is more manual but more honest.
Pick the 30 queries that matter most for your work. Run them through Claude, ChatGPT, Perplexity, and Gemini. Write down where you get mentioned and where a competitor does instead. That is your baseline. Re-run it weekly. You are no longer chasing a position number, you are chasing whether the machine says your name.
It takes an afternoon to set up and ten minutes a week to maintain. It will tell you more about your real AI visibility than any third-party score, because it is the actual output users see.
The uncomfortable part
There is a cost buried in all this. If Claude rewords your work and attaches a link, some people get the answer and never click. Synthesis means the model captures the value of your explanation while sending you a fraction of the traffic a blue link used to. That is the trade. The leaked prompt makes the mechanism undeniable, which is better than guessing.
The honest strategy is to be cited often enough, by name, that the brand recognition compounds even when the click does not happen. Being the source the machine trusts is its own kind of ranking. For how this whole episode is reshaping who even has access to these models, see Anthropic pulling Fable 5 offline.
Frequently asked questions
Does Claude quote my content directly?
Rarely, by design. The leaked prompt instructs Claude to reword search results rather than quote them, and flags 15 or more words from one source as a severe violation. It synthesizes your information and attributes it with a link, instead of lifting your sentences.
How do I get cited by Claude?
Be the clearest, most current, most self-contained source on your topic, in server-rendered HTML, tied to a recognizable entity. Answer questions directly in the first sentences. The model reaches for the source it can summarize most cleanly, and attaches that name.
Is GEO different from SEO?
They overlap but diverge on intent. SEO optimizes for a ranking position you click through. GEO, or generative engine optimization, optimizes to be the source an AI synthesizes and names in its answer. Strong fundamentals help both, but writing for synthesis is a distinct skill.
How do I track AI citations?
Manually, and it is worth it. Run your most important queries through the major AI assistants on a fixed schedule, record who gets cited, and watch the trend. There is no reliable automated rank tracker for AI answers yet, so direct observation is the ground truth.
Continue reading
Anthropic Pulled Claude Fable 5 Offline. The Two-Tier AI Problem Is Now Real.
A jailbreak, a 'secret sabotage' scandal, a US government order, and a one-line resurrection. The wildest week in AI ended with the best public model going dark.
Claude Fable 5's Entire System Prompt Leaked. I Read All 1,585 Lines.
Within 24 hours of launch, Claude Fable 5's full system prompt hit GitHub - 120,000 characters, 1,585 lines. Here is what is actually inside it, and what it tells you.