How Claude Chooses Sources

Claude decides which sources to cite the same way a careful analyst does: it will only commit to naming a specific business or fact when it either remembers it clearly from training or can see it, right now, in a document it has been handed. And of the major assistants, Claude is the one most willing to say "I'm not sure" instead of guessing. If you want Claude to recommend you, you have to make the safe answer and your name the same thing.

I've now written about how ChatGPT chooses sources, how Perplexity chooses sources, and how Google's AI Mode chooses sources. Claude is the fourth, and it's the one people ask me about the least and misunderstand the most — usually because they assume it behaves like ChatGPT. It doesn't. Claude has a distinct temperament, and that temperament changes the AEO playbook in ways worth spelling out.

First, understand the two ways Claude can cite anything

Every answer Claude gives comes from one of two places, and the distinction is the whole ballgame. I called these the two memories of AI in an earlier piece, and Claude makes the split unusually visible.

The first memory is training — the compressed, lossy recollection of the web the model absorbed before its cutoff. When you ask Claude a question with no tools attached, it answers from this memory. It has no live view of the internet in that mode. It cannot look anything up. It is reasoning from what it internalized months or years ago, and it knows that memory is imperfect.

The second memory is retrieval — documents placed into Claude's context at the moment you ask. That happens when Claude has web search enabled, when it's wired to a set of tools or connectors, or when someone pastes a page in. In that mode Claude is grounding its answer in text it can actually read, and it will cite the passages it used.

The reason this matters more for Claude than for its competitors: Claude is comparatively cautious about the first memory and comparatively rigorous about the second. It is trained to distrust its own recall of specifics — exact names, numbers, quotes, current facts — and to lean on retrieved evidence when the stakes of being wrong are high. So the question "how do I get cited by Claude?" is really two questions: how do I get into its trained memory as a name it's confident about, and how do I get into the documents it retrieves at answer time?

Why Claude is the most conservative namer of the bunch

If you run the same "recommend a business for X" prompt across the major assistants, you'll notice a pattern. ChatGPT will often produce a confident, well-formatted list. Perplexity will hand you a ranked set with citations. Claude is the one most likely to give you a framework, name a couple of well-established players it's genuinely sure of, and then tell you to verify — especially for anything local, recent, or high-consequence.

This isn't a bug and it isn't Claude being unhelpful. It's a design choice about calibration. Claude is trained to make its stated confidence match its actual evidence. When it doesn't have strong, corroborated grounds to name a specific small business, it would rather describe how to choose one than fabricate a recommendation. From an AEO standpoint, that has a sharp implication: Claude sets a higher bar for naming you than the other engines do. You don't clear that bar with volume or clever phrasing. You clear it with evidence Claude finds hard to argue with.

It corroborates before it commits

The single most useful thing to understand about Claude is that a claim it can verify from multiple independent angles gets treated very differently from a claim it saw once. This is the same consensus-threshold dynamic I've described for AI search generally, but Claude leans on it hard. A business that shows up consistently — same name, same category, same location, same core facts — across many independent sources reads to Claude as a safe thing to say. A business that appears in one glowing page and nowhere else reads as a risk. Claude resolves risk by hedging or omitting, not by taking a flyer on you.

It hedges instead of guessing

Where a more confident model fills a gap with a plausible-sounding specific, Claude tends to flag the gap. "I'd want to double-check the current details" is a very Claude sentence. That means the failure mode you're fighting isn't Claude naming your competitor over you — it's Claude naming nobody and handing the user a checklist instead. Your job is to be so clearly, verifiably the answer that hedging would look like Claude withholding obvious information.

What actually gets Claude to name you

Be quotable at the passage level

When Claude is grounding an answer in retrieved documents, a citation is earned sentence by sentence, not page by page. Claude wants to write a sentence and point at the exact passage that supports it. Pages that make a specific, standalone, checkable claim — "X costs between A and B," "the process takes N steps," "the difference between P and Q is R" — give Claude something to lift and attribute. Pages that are vague, hedged, or buried in throat-clearing give Claude nothing to defend, so they don't get cited even when they're retrieved. Write like you're feeding a careful editor pull-quotes.

Get into the sources Claude's tools can actually reach

In retrieval mode, Claude can only cite what its tools bring back. Increasingly that's not just an open web crawl — it's search results, and, in a growing number of setups, specific connectors: a company's own documentation, a knowledge base, a set of trusted feeds. The businesses that win here have made their key facts machine-readable and available where those tools look: clean pages, clear structured data, an accessible llms.txt, and content that doesn't hide behind scripts or logins. If a retrieval tool can't parse you, Claude can't quote you, full stop.

Build a corroborated entity, not just content

For Claude's trained memory — the part that lets it name you without looking anything up — what matters is entity strength. Claude needs to know who you are, what category you belong to, and that these facts are agreed upon across the web. This is entity authority doing its familiar work, and it's why I keep pointing people at the boring infrastructure: consistent naming everywhere, a presence in the reference layer, third-party corroboration you didn't write yourself. A thin entity is exactly the kind of thing Claude decides not to risk asserting.

A test you can run in five minutes

Here's a diagnostic I give clients so they can feel the difference themselves. Take a real buyer-intent question in your category and ask it twice. First ask a version of Claude with no web tools, so it answers from trained memory. Then ask a version with search enabled, so it answers from retrieval. Save both.

The gap between those two answers is a map of your problem. If the no-tools answer doesn't name you but the search answer does, your public entity is thin — Claude doesn't remember you confidently, and you're relying on live retrieval to rescue you. Fix that with corroboration and reference-layer presence. If the no-tools answer names you but the search answer hedges or drops you, your live footprint isn't quotable or retrievable enough — the tools either can't reach your pages or can't find a passage to lift. Fix that with cleaner, more parseable, more specific content. If both answers hedge, you have work to do on both fronts, and at least now you know it. I run this same split on competitors, too; watching which of them survives the no-tools version tells you who Claude actually trusts versus who's merely retrievable.

Where people get this wrong

The most common mistake is treating Claude like ChatGPT and being frustrated when the same content produces a warmer response from one and a shrug from the other. They are different instruments. ChatGPT will reward a strong, well-organized page more readily; Claude will reward corroborated, verifiable, specific evidence and punish thinness by staying quiet. If you're getting named confidently by ChatGPT and hedged around by Claude, that's not a content-quality problem — it's a corroboration problem. The fix is more independent evidence, not more words.

The second mistake is assuming Claude "doesn't do recommendations." It absolutely does, when it's sure. I've watched Claude name specific, well-established businesses without hesitation the moment the evidence supports it. The reticence isn't a policy against naming; it's a threshold. Clear the threshold and the hedging disappears.

The third mistake is ignoring the tool layer. A lot of the highest-value Claude usage now happens inside products, workflows, and connectors where Claude is answering from a specific, curated set of documents rather than the open web. If you're not present in the sources those integrations pull from, you don't exist in that context no matter how strong your public entity is. AEO for Claude increasingly means asking, for each place your customers use it, "what documents does Claude see here, and am I one of them?"

My read on where this goes

My read — and this is a judgment call, not a guarantee — is that Claude's caution is a preview of where all the serious assistants are heading. As these systems get pushed deeper into work that has real consequences, the ones that guess confidently and turn out wrong become liabilities, and the pressure is toward exactly the calibrated, evidence-first behavior Claude already exhibits. That's good news for anyone building genuine authority and bad news for anyone hoping to game their way into an answer. The businesses that win the cautious models are the ones that are actually, verifiably the right answer — and made that fact easy to check.

If you only optimize for the confident engines, you'll do fine until the confident engines get more careful. Optimize for Claude — the hardest audience to convince — and you tend to earn the others for free. I'd rather build for the model that demands proof, because proof is the one thing that travels: it works on Claude today, and it works on whatever the rest of them become tomorrow.

About the author

Scott Tischler is the Founder & Chairman of AIrecommend.ai and a practitioner-authority on AI search and Answer Engine Optimization. With 20+ years in marketing technology — including American Express, MetLife, and UBS — and executive and professional study at Wharton, Harvard, and Oxford, he helps businesses become the ones AI recommends.