How to Get Cited by Claude: The Memory-First Playbook

Claude answered purely from training data until March 2025, when Anthropic switched on web search for paid U.S. users -- years after every competitor. Search reached all plans globally in May 2025. That late start tells you where Anthropic's priorities sit: the model's own knowledge does the work, and search is the exception path.

The exception has rules. Anthropic's web search documentation says Claude searches when a request depends on information that is current, changing, or outside its training data -- recent events, live prices and data, or details about specific organizations and products that may have changed. Ask Claude a stable category question ("what is the best CRM for a small agency?") and it will usually answer from memory. Ask what changed this quarter and it searches.

What powers the search? Anthropic has never officially said. When search launched, TechCrunch reported that Brave Search had just appeared on Anthropic's subprocessor list and that Claude's citations matched Brave's results almost exactly. Since then Anthropic has also started operating its own search crawler, Claude-SearchBot, which its crawler documentation says "navigates the web to improve search result quality for users." The honest summary for operators: Claude's search launched on Brave's index per 2025 reporting, Anthropic now crawls for search quality itself, and the backend has never been formally disclosed -- so optimize for being crawlable and indexable in general, not for one index's quirks.

One more architectural fact matters: staleness is built in. The current top models -- Claude Fable 5, Opus 4.8, and Sonnet 5 -- carry a reliable knowledge cutoff of January 2026. Whatever you published after that does not exist in Claude's memory; it can only reach users through the search path. Both games are always in play.

Start with the number that reframes everything: Claude runs the software the rest of the industry is built with. Menlo Ventures puts Anthropic at roughly 54% of the enterprise coding market against OpenAI's 21%, and every major AI coding tool -- Copilot, Cursor, Sourcegraph, Replit, Cognition, GitLab, Vercel -- runs on Claude. Claude Code alone crossed a $2.5B run-rate (Anthropic). This is not a challenger poking at a lead; in the segment that builds AI products, Claude is the lead.

It is also the fastest-climbing engine in the field. Sensor Tower has Claude's U.S. audience up 452% year over year and now the third assistant globally; Apptopia tracked its U.S. mobile share climbing from under 2% to roughly 17% across the first half of 2026 while ChatGPT's slipped. And Claude converts better than anyone: the highest paid-conversion rate of any assistant, on a user base that skews affluent -- roughly 80% in six-figure households.

The honest caveat: by total headcount Claude is still the smallest of the major assistants. Pew Research (published June 2026) has 6% of U.S. adults using Claude, behind ChatGPT's 44% and every other named assistant. But for GEO, raw reach is the wrong lens. The people on the other side of a Claude conversation are disproportionately developers, analysts, and teams evaluating vendors -- the high-intent buyers you most want to be recommended to -- and Claude's answers are wired into the tools they already work in. A Claude mention reaches fewer people and more of the right ones.

That reach compounds one more way. Because citations are always enabled for web search in the Claude API -- Anthropic requires developers to display them -- being a source Claude trusts pays off across every Claude-powered product, not just claude.ai. On a company now at a $47B run-rate (Series H, up from about $10B at the end of 2025), that is a lot of surfaces.

In our 102-brand, five-engine study, Claude posted the highest unbranded recognition of any engine: on the first tracked run, it mentioned the brand in 51.5% of unbranded category prompts -- questions like "best payment platform for Indian startups" that never name the brand. ChatGPT, Perplexity, and Gemini sat between 18.7% and 23.9%; Grok at 12.0%.

Before you re-budget everything toward Claude: part of that lead is brand mix, not the engine. Claude and Grok are Agency-tier engines in our dataset, so they run on fewer brands, and that smaller cohort skews toward higher-stature brands -- which the brand-stature ladder shows are exactly the brands every engine recognizes more. Read Claude's number as "memory-first engines reward brands they already know" rather than "Claude is easier."

Three things sharpen that picture. Claude's lead is widest on the hard, specific prompts -- on problem/solution questions it mentions the brand 36.9% of the time against a 10.8% five-engine average -- exactly the long-tail questions that never map onto a convenient ranked list. Its visibility is also stickier: across brands with weeks of tracking, Claude's per-run slope was statistically flat while ChatGPT and Perplexity drifted measurably down, because memory does not churn the way retrieval does. And in our CRM deep dive Claude spread its mentions across brands more evenly than any other engine -- less winner-take-all at the top of an answer, so challenger brands have more surface to win.

Here is the trap in the phrase "get cited by Claude." In the same CRM deep dive, we measured how often each engine's brand mentions arrived with a source link: Perplexity 95%, Gemini 35%, Grok 20%, ChatGPT 15% -- and Claude 10%. When Claude answers from memory, there is nothing to link. The mention itself is the prize.

This has two practical consequences. If you measure Claude the way you measure Perplexity -- counting linked citations -- you will conclude Claude ignores you while it is actively recommending you in answer text. Track mentions, not just links, as the cross-platform hub argues for every engine. And referral traffic will understate Claude's influence more than any other platform's: a linkless recommendation shows up in your pipeline as a branded search or a direct visit three days later, not as a Claude referral.

Claude's memory is built from web-scale training data collected by ClaudeBot and frozen at each model's cutoff. You cannot submit anything to it. What you can do is make the open web say more about you -- durably, consistently, and in places crawlers treat as canonical.

Build the third-party record, not just your site

Across our full citation corpus, only 2.9% of AI citations point to a brand's own domain -- 75.2% go to corporate and competitor pages, category roundups, and the rest of the third-party web. Training corpora have the same shape: a model learns what you are from what everyone else says about you. Comparison pages, category roundups, the "best X" lists AI engines lean on, documentation communities, and press coverage all compound into parametric recall.

Make your entity unambiguous

Memory-first engines need to resolve which thing you are before they can recommend you. The entity-stack playbook -- consistent naming, Wikidata, Crunchbase, LinkedIn, and the schema that ties them together -- is more valuable for Claude than for any retrieval-first engine, because a confused entity cannot be recalled cleanly.

Play the model-generation clock

Training cutoffs mean memory work pays with a lag: the third-party coverage you earn this quarter surfaces in the next model generation, not tomorrow. That is a reason to start now, not a reason to skip it -- the brands Claude recommended most in our study are the ones whose web record has been accumulating for years, exactly the stature effect the 73/44/11 ladder quantifies.

When a prompt trips Claude's search triggers -- current, changing, or post-cutoff information -- the memory game pauses and a retrieval game begins. Three moves matter most.

Be the freshest credible answer for your category

Claude searches precisely when freshness is required, so dated, recently-updated, factual pages are what the search path selects for: pricing pages with visible effective dates, changelogs, annual reports with the year in the title, data pages that state when they were last refreshed. If your category content was last touched in 2024, you are invisible on the exact queries where Claude goes looking.

Stay crawlable by the two bots that matter at answer time

Claude-User fetches pages when a user's question triggers retrieval; Claude-SearchBot crawls to improve search quality. Anthropic's own documentation warns that blocking either "may reduce your site's visibility" in results. Neither bot renders JavaScript-only content reliably -- like every AI crawler we have measured, what matters is what is in the raw HTML.

Remember the API layer multiplies the stakes

Claude-powered products can restrict search with allowed_domains and blocked_domains and localize it with user_location, per the web search tool docs. Third-party builders are actively curating which domains their Claude apps may cite. Being a domain developers trust enough to allowlist -- documentation, data, reference pages -- is a Claude distribution channel that has no ChatGPT equivalent.

Anthropic splits its crawling into three user agents with three different jobs, and that granularity is the whole point: you can opt out of model training while staying fully visible in Claude's answers.

The configuration most brands actually want -- keep answers and search on, keep training opt-out on the table -- looks like this in robots.txt:

# Visible in Claude answers and search
User-agent: Claude-User
Allow: /

User-agent: Claude-SearchBot
Allow: /

# Optional: opt out of model training only
User-agent: ClaudeBot
Disallow: /

Blocking training is a legitimate licensing position -- but know what it costs in a memory-first engine: ClaudeBot is how you get into the next model's parametric memory, which Game 1 says is where most Claude mentions come from. Blanket "block-all-AI-bots" rules are the worst of both worlds: they take you out of training and out of live answers in one line. Test what each bot can actually reach with curl -A "Claude-User" https://yoursite.com -- or run our free crawler inspector, which fetches your page as all three Claude agents alongside ten other AI bots.

One honesty note on llms.txt: as of mid-2026 there is no evidence Claude's crawlers request it. Anthropic publishes an llms.txt for its own docs but has never said its bots consume the standard, and log-file analyses report that no major LLM provider fetches the file. Ship one -- it costs minutes and may matter later -- but treat robots.txt and raw-HTML quality as the real levers for Claude today.

Claude will not reward a content sprint the way a retrieval engine can. Its answers lean on a memory that updates in model generations, drawn from a web record that takes quarters to shift, in front of the smallest but highest-value audience in the field -- and the fastest-growing one. That feedback is slower than Perplexity's, and stickier: once you are in the memory, you tend to stay. So run both games in parallel -- build the third-party record and entity clarity the next model will memorize, keep your search-facing pages fresh and crawlable for the queries that force retrieval today, and measure mentions rather than links so you can see it working.