How Claude picks sources (and why brands get under-cited)
Claude searches the web through Brave, cites the most selectively of any major engine, and defaults to its training corpus. Here is how it picks sources.
Claude is the AI engine your brand is least likely to show up in, and it is not because Claude is small. Anthropic now controls roughly 40% of enterprise LLM API spend, ahead of OpenAI, per Menlo Ventures' December 2025 enterprise survey. The under-citation is a design choice. Claude answers from its training corpus by default, searches the live web only when it decides the question demands it, and when it does search it cites the most conservatively of any major engine. Signals runs an aged Reddit account marketplace plus an editorial network for AI brand mentions across Reddit, Quora, Product Hunt, and Threads, and Claude is the surface where the brands we work with most often see a competitor named and themselves left out. This is how Claude actually selects and cites sources, where its retrieval differs from ChatGPT and Google, and why the gap is fixable.
How does Claude decide whether to search the web at all?
Claude does not search by default. Unlike a pure answer engine, its first move is to check whether its training corpus already covers the question. Anthropic's documentation is explicit: Claude searches when a request depends on information that is "current, changing, or outside its training data," and answers directly for "established facts, math, science fundamentals, or coding concepts."
That decision gate is the first place brands lose. If a user asks "what is the best project management tool," Claude may treat that as stable knowledge and answer from training, never firing a search, never opening a citation slot. No retrieval event means no live URL to cite, so your freshly published comparison page is invisible regardless of how good it is. The practical implication: for Claude, being baked into the training corpus as a known entity matters more than for ChatGPT or Perplexity, because Claude reaches for the live web less readily and trusts what it already knows.
What powers Claude's web search, and why does that matter?
When Claude does search, it is not searching Google. Claude's web search runs on Brave Search. TechCrunch documented this in March 2025 after engineer Antonio Zugaldia spotted Brave Search added to Anthropic's subprocessor list, and Simon Willison found Claude's search function carried a parameter called BraveSearchParams and returned citations identical to Brave's own results. Brave already powered Mistral's Le Chat, so the partnership fit an established pattern.
This matters because Brave's index is not Google's. Brave operates an independent crawler and a smaller, differently weighted index, so the pages that surface for a query in Brave are not always the pages that rank in Google for the same query. A brand can sit at position three in Google and be effectively absent from Brave's top results, which is the pool Claude draws from. Optimizing only for Google rankings and assuming Claude will follow is the most common mistake we see. The retrieval layer Claude actually reads is Brave, and few marketers ever check it.
How does Claude actually cite a source once it searches?
Claude returns citations as encrypted, block-level references, and it does so automatically. Per Anthropic's documentation, "citations are always enabled for web search." Each search result comes back with an encrypted_content field, and in multi-turn conversations that encrypted blob has to be passed back so Claude can keep referencing the source without re-exposing the full page. The retrieval is sealed, which is why operators cannot reverse-engineer Claude's source list the way they scrape Perplexity's.
The unit of citation is the text block, not the sentence. Each citation carries a url, a title, and up to 150 characters of cited_text lifted from the source. Claude cites a whole retrieved block rather than a phrase buried inside it, so a page that answers the question in one clean, self-contained block is far more citable than one where the answer is spread across three paragraphs. The newest tool version adds dynamic filtering, where Claude writes and runs code to discard irrelevant search results before they reach its context. A page that does not survive that filter never gets considered for a citation at all.
Which sources does Claude favor, and which does it avoid?
Claude is the authority engine. Across the majors it leans hardest toward established, high-Domain-Authority institutions and away from social and user-generated content. Profound's analysis of 3.25 billion AI citations put Claude's social citation rate at 3.99%, the lowest of any major engine, against 15.3% for Google AI Overviews and 11.3% for Perplexity. An arXiv study of more than 10,000 Claude health citations found 97.8% came from established institutions with a median Domain Authority of 92.
The bias extends to recency. Claude prefers stable, durable content over breaking coverage: only 36% of its journalism citations came from the past 12 months, versus 56% for ChatGPT, per the 5WPR AI Platform Citation Source Index 2026. One independent teardown of 2,170 Claude-cited URLs found 56% sat on /blog/ paths and only 3% were homepages, with a striking 28.1% on .ai domains. The pattern is consistent: deep, structured, named-author practitioner content from a credible domain, not homepages, press releases, or thin listicles.
| Source type | Claude's posture | What it means for you |
|---|---|---|
| Wikipedia and reference | Heavily favored | Entity presence matters before retrieval ever runs |
| Established institutions | Favored (97.8% of health cites) | High Domain Authority is close to a prerequisite |
| Practitioner and SaaS blogs | Favored (56% on /blog/ paths) | Deep, named-author analysis earns slots |
| Social and UGC | Avoided (3.99% social rate) | Reddit and forum seeding underperforms here vs Perplexity |
| Breaking news | Underweighted (36% from past year) | Durable content beats freshness plays |
| Homepages and PR wires | Rarely cited (3% homepages) | Your homepage is not the asset that gets cited |
Figures from Profound (3.25B citations), an arXiv health-citation study, the 5WPR Citation Source Index 2026, and a 2,170-URL Claude teardown. The through-line: authority and depth over freshness and reach.
Why do brands get under-cited in Claude specifically?
Because Claude applies two filters where most engines apply one. The first filter is the search-decision gate: if Claude answers from training, your live content never gets a hearing, so the only way to be present is to already exist as a recognized entity in the corpus, which means durable third-party coverage that predates the query. The second filter is the Brave-retrieval gate: even when Claude does search, it pulls from Brave's index and then down-weights social, UGC, and low-authority pages before citing.
A brand fails Claude when it is invisible at either layer. We see two failure modes constantly. The first is the new brand with no entity footprint: no Wikipedia presence, no named editorial coverage, nothing Claude's training run absorbed, so Claude has no reason to surface it even when asked directly. The second is the SEO-strong brand that ranks in Google but has no Brave presence and leans on social or affiliate content that Claude structurally discounts. Both are fixable, but they are different fixes, and treating Claude like ChatGPT misdiagnoses which one you need. This is the same causal split we trace in our forensic audit of why a competitor shows up in ChatGPT and you do not.
How do you earn a Claude citation?
You build presence at both layers Claude reads, and you do it with third-party authority rather than your own pages. Start with the entity layer: durable, independent coverage that a training run will absorb and that establishes your brand as a known thing in your category. That means named editorial mentions on credible domains, reference-grade presence, and consistent third-party description of what you do, not a denser homepage. This is the core of the GEO thesis we lay out in our pillar on getting mentioned by ChatGPT: unlinked brand mentions predict AI citation roughly 3x more strongly than backlinks, and the mechanism is exactly what Claude rewards.
Then address the retrieval layer. Check your Brave Search visibility for your top category queries directly, because that is the index Claude reads, and most brands have never looked. Publish deep, named-author analysis on a high-authority domain in the /blog/ structured format Claude favors, with one clean, self-contained answer block per question so it survives dynamic filtering and fits Claude's block-level citation. The editorial placements that move both layers at once are the ones on established third-party sites, which is precisely the asset our analysis of the 50 domains that drive most AI citations maps out.
How is Claude different from ChatGPT, Perplexity, and Google AI Mode?
Claude is the most conservative citer and the most training-dependent of the four. ChatGPT and Perplexity reach for live retrieval more readily and tolerate a wider source mix, including far more social and Reddit content. Perplexity in particular leans heavily on community sources, where Reddit makes up close to half its top-cited pool. Google AI Mode runs aggressive query fan-out across its own index and cites a broad set of domains, as we cover in our breakdown of how Google AI Mode picks sources.
Claude sits at the other end. It searches less often, draws from Brave rather than Google, discounts social content hardest, and prefers durable institutional and practitioner sources over fresh coverage. The operator consequence is that a Reddit-and-freshness strategy that earns Perplexity citations will underperform in Claude, and a strategy built on high-authority editorial presence and entity-level recognition will outperform there. You cannot run one undifferentiated GEO playbook across all four engines and expect even coverage. Claude rewards authority and depth; optimize for it specifically or accept being the brand it leaves out.
Frequently asked questions
Does Claude search the web for every question?
No. Claude answers from its training corpus by default and only triggers web search when the request depends on information that is current, changing, or outside its training data, per Anthropic's documentation. For questions it treats as stable knowledge, such as established facts or general "best tool" queries, it may not search at all, which means no live citation slot opens. Being recognized as an entity in the training corpus matters more for Claude than for engines that retrieve more eagerly.
What search engine does Claude use for web search?
Brave Search. TechCrunch documented the integration in March 2025 after Brave Search appeared on Anthropic's subprocessor list and engineers found a BraveSearchParams reference in Claude's search function plus citations identical to Brave's results. This matters because Brave runs an independent index that differs from Google's, so a page that ranks well in Google can be absent from the Brave results Claude actually reads. Checking Brave visibility for your category queries is a step most brands skip.
Why is my brand not cited in Claude when it shows up in ChatGPT?
Usually because Claude applies two filters ChatGPT does not apply as strictly. First, Claude answers from training more often, so if your brand is not an established entity in the corpus, it may never search for you. Second, when it does search, it pulls from Brave and heavily discounts social, UGC, and low-authority pages. A brand strong on Reddit seeding or recent SEO content but weak on durable third-party editorial presence tends to clear ChatGPT and fail Claude.
How does Claude format its citations?
Citations are always enabled for Claude's web search. Each result is returned as an encrypted content block, and citations include the source URL, title, and up to 150 characters of cited text. Claude cites the whole retrieved text block rather than a single sentence, so content that answers a question in one clean, self-contained block is more citable than content where the answer is spread across several paragraphs. The newest version also filters results with code before they reach Claude's context.
Does Claude cite Reddit and social media?
Rarely, compared with other engines. Profound's analysis of 3.25 billion AI citations put Claude's social citation rate at 3.99%, the lowest of the major engines, versus 15.3% for Google AI Overviews and 11.3% for Perplexity. Claude favors established institutions, reference pages, and named-author practitioner content over user-generated posts. A community-seeding strategy that works well for Perplexity will underdeliver in Claude, where authority and Domain Authority carry far more weight than social reach.
What kind of content does Claude cite most?
Deep, structured, named-author content on high-authority domains. A teardown of 2,170 Claude-cited URLs found 56% sat on /blog/ paths and only 3% were homepages, while a health-citation study found 97.8% of sources came from established institutions with a median Domain Authority of 92. Claude also skews toward durable content over breaking news, with only 36% of its journalism citations from the past 12 months. Practitioner analysis with clear methodology beats homepages, press releases, and thin listicles.
:::
