Claude is more conservative than ChatGPT about which sources it cites. When evidence feels thin, Claude will say so rather than constructing a confident-sounding answer from weak sources. When a page relies on secondary summaries rather than original analysis, Claude tends to skip it. When claims are not sourced, Claude treats them as unverifiable rather than giving the page the benefit of the doubt.
This makes Claude harder to manipulate and more consistent to optimise for. The same things that make content genuinely good — original data, named sources, structured evidence, authoritative authorship — are the things Claude's retrieval system systematically favours. Generic content has always been a weak AEO bet. With Claude specifically, it is almost worthless.
Claude has two distinct citation modes, and optimising for the wrong one is one of the most common mistakes brands make. Understanding which mode is active for a given query type determines what work actually produces citations.
What Are Claude's Two Citation Modes?
Claude in its default configuration on claude.ai answers from training data without browsing. Citations in this mode come from whatever Claude learned during training — typically content published and widely referenced before the training cutoff. Optimising for this mode means building a strong open-web entity footprint over time: Wikipedia presence, named coverage in established publications, Wikidata entry, consistent entity signals across authoritative third-party sources. This is the same entity-building work covered in the entity-based AEO guide.
Claude with web search enabled — in claude.ai with search turned on, in the API with tool use configured, or in partner products that wrap Claude with retrieval — cites live web sources the same way Perplexity does. In this mode, Claude uses Brave Search for web retrieval, per TechCrunch reporting confirmed in March 2025 and aligned with Anthropic's published subprocessor details. Profound's 2025 analysis found 86.7% overlap between Claude and Brave Search citations. Optimising for this mode means being indexed and trusted by Brave Search's crawler, which follows standard web crawl principles and responds to the same technical signals as Bingbot and Googlebot.
The practical implication: more of Claude's behaviour depends on training data than ChatGPT's does, because ChatGPT triggers web browsing more reflexively for factual and commercial queries. Claude tends to answer from internal knowledge first and treats web search as a deliberate tool call rather than a default. This makes long-term entity building — the kind of work that gets your brand into Claude's training knowledge — proportionally more valuable for Claude than for any other major engine.
What Does Claude's Constitutional AI Framework Mean for AEO?
Anthropic trained Claude using a process called Constitutional AI, which embeds specific values around honesty, accuracy, and harm avoidance at the training level. For AEO practitioners, this surfaces in one consistent pattern: Claude applies a higher evidence bar than other engines before citing a source confidently.
Content that contains promotional language, unsupported superlatives, or claims that feel unverifiable gets cited with hedging or not cited at all. Content that shows its work — named sources, dated statistics, specific outcomes rather than vague claims — gets cited with higher confidence and cleaner attribution.
Stridec's April 2026 analysis of Claude's sourcing patterns documented three specific content patterns that Claude systematically down-weights: generic affiliate aggregator content, low-trust SEO listicles with no original data, and unsourced claims. These are exactly the content types that still earn some citation traction on ChatGPT in browse mode. For Claude, they do not.
The positive flip side: brands that publish genuine first-hand analysis, case studies with specific numbers, and content with verifiable claims have a disproportionate advantage in Claude citations relative to their traditional SEO authority. Claude favours content with genuine epistemic credibility, not just domain authority.
What Technical Access Does Claude Require?
Claude uses two separate crawlers — ClaudeBot for training data collection, and Claude-SearchBot for live retrieval when web search is enabled. These are independently configurable in robots.txt, the same way OpenAI's GPTBot and OAI-SearchBot are separate directives.
Blocking ClaudeBot keeps your content out of future training datasets but has no effect on live retrieval citations. Blocking Claude-SearchBot removes your pages from Claude's web search results. Most brands that want citation without training exclusion should allow Claude-SearchBot and optionally block ClaudeBot based on their own policy:
# Allow live retrieval citations
User-agent: Claude-SearchBot
Allow: /
# Optionally block training collection
User-agent: ClaudeBot
Disallow: /Check your server logs for both user-agents. If Claude-SearchBot has never appeared in your logs, either your WAF is blocking it before it reaches robots.txt, or Claude's web search is not surfacing your domain as a candidate for your target queries. The AI crawlers guide covers the WAF and CDN check in full.
Brave Search crawls the web independently. Brave's crawler user-agent is BraveBot. It follows standard robots.txt rules and standard technical SEO signals — crawlability, page speed, content in initial HTML response, structured data. A page that Brave cannot crawl or trust will not appear in Claude's web search results regardless of how well it is written.
What Content Types Does Claude Favour in Live Retrieval?
Stridec's analysis identified four content patterns that earn Claude citations consistently in web search mode:
Entity-first writing. Every passage should name the subject entity in the sentence itself, not rely on pronouns or prior context. "Claude uses Brave Search for web retrieval" is entity-first. "It uses this search engine for retrieval" is not — it cannot be extracted as a standalone passage because "it" is ambiguous. Claude's Constitutional AI training specifically favours content that can be cited with accurate attribution. Ambiguous entity references undermine that.
Inline sourced claims. Every significant claim should carry its source in the same sentence, not in a footnote or a references section. "Perplexity processes 780 million queries per month (Bloomberg, 2026 CEO statement)" is directly citable with attribution. "Perplexity processes hundreds of millions of queries monthly" is a vague claim Claude cannot verify. Claude's grounding behaviour means it cross-references claims against what it knows. Named, dated, specific claims survive that check. Vague claims do not.
Original data or first-hand analysis. Content that synthesises what others have published earns weaker Claude citations than content representing a genuine original contribution. Case studies with specific client outcomes, proprietary research with named methodology, or analysis that draws a conclusion not available elsewhere are the content types Claude's training process specifically favours, per Hashmeta's February 2026 analysis. The E-E-A-T guide covers the experience signals that signal first-hand knowledge.
Named authorship with verifiable credentials. Claude applies stronger E-E-A-T weighting than most other engines. Content bylined by "the editorial team" or attributed to no author is treated as lower-authority than content with a named author, a linked professional bio, and verifiable credentials in the subject area. Person schema connecting the author to their professional profiles is the machine-readable implementation of this signal.
How Do You Build Training Data Presence for Claude?
Training data presence is the long-game strategy for Claude — getting your brand's accurate information into the model's parametric memory so it describes you correctly even in default (no web search) mode.
The same entity-building work that affects Google Knowledge Graph and Wikidata affects Claude's training data. Wikipedia articles, Wikidata entries, LinkedIn company profiles, editorial coverage in publications Claude treats as authoritative, and review platform profiles all feed into the training corpus. Content published before a training cutoff date has a chance of being incorporated into model weights — creating a more durable citation pathway than live retrieval alone.
One practical action: publish evergreen content well before anticipated model training refreshes, not reactively. Anthropic's model release announcements typically signal upcoming training updates. Content that has been indexed, cited by multiple sources, and stable for several months before a training update is more likely to make it into model memory than content published and indexed the week before.
Run branded prompts through Claude monthly — "What is [your brand]?" "What does [your brand] do?" "What are [your brand]'s main features?" — and document exactly what Claude says. Discrepancies between Claude's descriptions and your current product are training data gaps. The brand hallucination guide covers the detection and correction process for exactly this pattern.
Claude citation tracking requires separate treatment from ChatGPT and Perplexity because of the two-mode dynamic. The NotionCue Prompt Tracker runs your tracked prompts through Claude with web search enabled weekly, so you can see live-retrieval citation rate. The Citation Tracker monitors what Claude says about your brand in default mode — catching training-data hallucinations before they shape buyer perception at scale. Both matter, but they require different fixes when they diverge.
Frequently Asked Questions
Does blocking ClaudeBot affect whether Claude cites my content?
Blocking ClaudeBot affects training data collection only. Claude-SearchBot handles live web retrieval for queries where Claude's web search tool is activated. You can block ClaudeBot (to exclude content from future training runs) while allowing Claude-SearchBot (to remain citable in web search mode) independently. Both directives can coexist in the same robots.txt file.
Why does Claude sometimes cite my competitor even though my content is more recent?
Claude's source selection in web search mode prioritises evidence quality over recency more aggressively than Perplexity does. A competitor page with more original data, more specific named sources, or more clearly structured evidence may be preferred over a more recent page that is less evidentially grounded. Check what the competitor page is doing differently from yours — named sources, specific statistics, original data, or clearer BLUF structure — and fix the underlying gap rather than just updating the date.
How long does it take to appear in Claude's training data after publishing new content?
Training cycles vary and Anthropic does not publish exact schedules. Content published several months before a model update has the best chance of inclusion. The more a piece is cited by other sources, the stronger its training data signal. Prioritise getting content cited by authoritative third-party sources — editorial coverage, Reddit upvotes, Wikipedia references — before each anticipated training update rather than trying to time publication to the week of an update.
Is Claude the right first engine to optimise for in a new AEO programme?
No. Claude should come after Perplexity and Google AI Overviews. Perplexity responds fastest to technical and content changes, provides explicit citation data, and signals quickly whether changes are working. Google AI Overviews reaches the largest audience. Claude's value comes from its conservative, high-evidence sourcing pattern — brands that build genuine authority see disproportionate returns, but that authority takes time to build. Start with the faster-feedback engines and build the evidence base that Claude rewards naturally.