Claude citation statistics

Why the findings section is short

Worth setting expectations before you read further. This page carries more method than findings, and that ratio is deliberate rather than a gap we intend to paper over with borrowed numbers from other platforms.

Every other statistics page on this site cites a vendor study analyzing millions of citations. For Claude, that study doesn't exist in public yet.

This isn't an oversight in our research. We looked, and the absence itself is the finding. That's exactly the kind of gap this site is built to notice, and eventually fill with our own first-party data. We won't paper over it with a number borrowed from a different platform.

What is publicly known

Short, and worth reading carefully rather than skimming. Two of these are documented facts. The third is a qualitative impression that circulates widely and rests on much less than its confident phrasing suggests.

The Claude bot map

Three bots, three jobs, and only one of them determines whether you can be cited. This is the part of the page with genuinely actionable guidance, because the bot behaviour is documented even where the citation behaviour is not.

See the full registry for verification methods and robots.txt handling.

Configuring Claude bot access correctly

This is the one section of this page where the guidance is concrete rather than provisional, because the bot behaviour is documented even though the citation behaviour is not. It is also where the costliest mistake on this page gets made.

The mistake. A robots.txt rule intended to opt out of AI training that catches Claude-SearchBot alongside ClaudeBot. The site owner believes they have declined training. They have also removed themselves from Claude's search answers entirely, and nothing about that outcome is visible to them. There is no error message and no report. Just an absence.

What each decision actually does. Blocking ClaudeBot means your content is not collected for future model training. It has no effect on whether Claude can cite you today. Blocking Claude-SearchBot means Claude's web search tool cannot retrieve your page, so you cannot be cited. Blocking Claude-User means a page a person explicitly pastes into a conversation cannot be fetched, which mostly frustrates your own readers.

The configuration most sites actually want. Allow Claude-SearchBot and Claude-User. Decide on ClaudeBot separately, as a content-policy question rather than a visibility one. If your content is your product and you would rather it not train future models, block it, and note that this costs you nothing in citation terms.

How to verify. Check your server logs for each user agent individually rather than assuming the robots.txt is doing what you intended. Verify by IP rather than trusting the user-agent string, since strings are trivially spoofed. The exact hostname patterns for verification are in the AI Bot Registry.

One deprecated token worth keeping. Older block lists reference anthropic-ai and Claude-Web. These are deprecated rather than active, and leaving them in an existing robots.txt is harmless. Do not treat their presence as covering the current bots.

Claude's citation behavior, qualitatively

Available discussion, not a controlled study, describes Claude leaning toward sources that are easy to verify. It's reportedly skeptical of vague or promotional content when a better-cited alternative exists.

That fits Anthropic's broader focus on careful, well-grounded answers. But treat it as a general description, not a measured finding. No percentage breakdown like the ChatGPT or Perplexity figures exists for Claude yet.

Why this gap exists, structurally

"Nobody has studied it" invites the obvious question of why not. The answer is not that Claude is harder to study. It is a combination of incentives and timing, and understanding it explains a fair amount about how this whole field's evidence gets produced.

Almost all citation research is produced by visibility vendors. Not academics, not publishers, not the AI companies themselves. Vendors build citation-tracking infrastructure because they sell access to it. What gets measured follows what customers ask about.

Vendor coverage follows commercial demand, not research value. Customers ask about ChatGPT because it has the largest consumer footprint, and about Google because it is Google. A platform with a smaller but more technical user base generates fewer support tickets asking "am I visible there," so it gets built out later.

Later arrival compounds it. Claude's web search capability arrived after the tracking tooling for the other engines already existed. Retrofitting a new surface into an existing measurement pipeline is real engineering work with an unclear payback, so it sits behind other priorities.

Nothing about the platform makes it hard to measure. This is the part worth stating plainly. Claude cites sources in its answers. A researcher could run a query set against it and record what comes back, exactly as they would for any other engine. The gap is not technical.

That combination, a genuine gap with no technical barrier, is exactly the profile of a research opportunity. It is why this platform sits first in the Index roadmap rather than last, despite being the smallest of the surfaces tracked.

What transfers from other engines, and what does not

In the absence of Claude-specific data, the practical question becomes which findings from other engines it is reasonable to borrow. Some transfer better than others, and being explicit about which is more useful than a blanket "we don't know."

Likely transfers: content-level tactics with a mechanism. The Princeton study's findings on quotations, statistics and cited sources rest on a mechanism, that specific and attributable content is easier and safer to quote, which is not engine-specific. Any retrieval system composing an answer from sources faces the same problem. This is the most defensible borrowing available.

Likely transfers: access requirements. A bot that cannot fetch a page cannot cite it. This is not a finding so much as a physical constraint, and it applies identically everywhere.

Probably does not transfer: source-type skew. ChatGPT's encyclopedic lean and Perplexity's community lean are quite different from each other. That variance is itself evidence that skew is engine-specific. Assuming Claude resembles either would be a guess dressed as an inference.

Probably does not transfer: ranking-overlap figures. Overlap with Google's top 10 varies from roughly 12% to roughly 38% across the engines that have been measured. With that much spread, no single figure predicts an unmeasured engine.

Unknown: citation density. Whether Claude cites many sources per answer or few is reported only qualitatively. Since this varies substantially between the measured engines, borrowing a number here would be arbitrary.

The honest position: borrow the mechanism-backed content tactics, borrow the access requirements, and treat everything engine-specific as unknown. That is less satisfying than a full playbook and it is what the evidence actually supports.

Getting cited by Claude: current best guesses

Labelled guesses on purpose. Everything in this section would be stated more confidently on almost any other site covering this topic, and the confidence would not be earned.

With no Claude-specific data yet, the safest approach is to apply the tactics with the strongest general evidence: clear citations, direct quotations, disclosed statistics. Make sure your content is easily verifiable too. That one property keeps coming up in descriptions of Claude's source selection.

Treat this as a working guess, not a proven playbook.

How to measure Claude yourself

The absence of published research has one practical upside. Anyone can produce the first measurement of their own presence here, and it requires no tooling beyond a spreadsheet.

Build a fixed question panel. Fifteen to thirty questions your actual audience would ask. Weight it toward the specific and technical, since that is where Claude's user base skews and where a smaller site is most likely to appear at all.

Run each question with web search enabled. This matters. Claude answering from training data alone will not cite live sources. The measurement only means something when retrieval actually runs.

Record the full source list. Not just whether you appeared. Who else got cited is the more valuable data, because it tells you what kind of source wins on your topics when nobody has published an aggregate answer.

Repeat each question several times. Answers are non-deterministic. A source cited once and absent twice is noise. Consistency across runs is the signal.

Watch Claude-SearchBot in your logs. This is your leading indicator and the one most people skip. Crawl activity precedes citation. If Claude-SearchBot has never fetched your pages, your problem is access, not content, and no amount of writing will fix it.

Freeze the panel and repeat monthly. The same discipline that makes any citation measurement meaningful. A changing question list generates apparent movement that means nothing.

Doing this puts you ahead of the published research on your own topics, which is an unusual position and worth taking advantage of while it lasts.

A brief timeline of Claude's web access

Claude's live web search is a newer capability than ChatGPT Search or Perplexity's core product. It arrived as a tool the model can invoke, not as the product's original design.

That later arrival is the simplest explanation for why large-scale citation studies exist for other engines but not yet for Claude. The research and vendor tooling that produces those studies for other platforms just hasn't caught up here yet. It's not that Claude is inherently harder to study.

What a proper Claude citation study would look like

Since the study does not exist, the useful contribution is specifying it. This is the design the Index intends to run, published here so it can be checked against what eventually ships, and so anyone else who wants to run it first has a method to work from.

The query set. The same published 1,000-query set used across every engine the Index tracks. Reusing it is the entire point: a Claude figure measured on a different query set than the ChatGPT figure cannot be compared to it, which is how most cross-engine claims in this field become meaningless.

Collection conditions, stated in advance. Web search explicitly enabled. Consistent session state. A stated geography. Five runs per query, because a single run mixes signal with the engine's own non-determinism and there is no way to tell which is which afterward.

What gets recorded per citation. Query, run number, timestamp, cited URL, cited domain, position within the response, and a context snippet. Position matters because nobody has established whether being cited first differs from being cited fifth.

The primary outputs. Source-type distribution, so the Claude equivalent of the Wikipedia and Reddit skews finally exists. Citation density per answer, compared against the other engines on the identical query set. Overlap with Google's top 10. And concordance with each other tracked engine.

Pre-registered predictions. Published before collection, so they can be wrong in public. The working expectations: Claude's source distribution will be less concentrated than ChatGPT's, its ranking overlap will fall between ChatGPT's and Perplexity's, and its citation density will be lower than Perplexity's. Each of those could turn out wrong, which is what makes registering them worthwhile.

What gets published. The raw citation rows, not only the summary. Anyone should be able to recompute every reported figure from the released file, and disagree with the analysis if the data supports a different reading.

None of this is technically difficult. It is a query set, a collection script, and the discipline to publish the predictions before the results. That such a study does not exist says more about the field's incentives than about the difficulty of the work.

How to reason about a platform with no data

The broader skill this page is really about: what to do when you need to make decisions and the evidence does not exist. It comes up constantly in this field, not only with Claude.

Separate mechanism from measurement. Some claims rest on how a system works, others on what someone observed. Mechanism-backed claims travel further into unmeasured territory. "Specific, attributable content is easier to quote" is about how retrieval works. "47.9% of citations go to encyclopedic sources" is an observation about one engine and travels nowhere.

Prefer the low-regret action. When you cannot tell which of two approaches is better, weight the one that helps regardless. Writing more specifically and citing sources improves a page for human readers and classic search whatever Claude does with it. That asymmetry makes it the right call under uncertainty.

Fix the binary constraints first. Access is not a matter of degree. A blocked bot is a hard zero, and that is knowable without any citation research at all. Resolve the things that have definite answers before agonising over the things that do not.

Measure the thing yourself, at small scale. A thirty-question panel run monthly is imperfect and it is real evidence about your own situation. It beats waiting for someone else to publish an aggregate that may not describe your niche anyway.

Say "unknown" out loud. The failure mode this field rewards is confident guessing. Naming the uncertainty costs you nothing with a serious reader and protects you when the guess turns out wrong.

The research gap, explicitly

Nothing public answers these four questions yet. What share of Claude's citations go to which domain types? Does Claude's citation rate correlate with Google ranking at all? How citation-dense are Claude's answers compared to ChatGPT or Perplexity? How does Claude-SearchBot's crawl behavior compare to other retrieval bots? All four are registered as open questions in the Citation Index roadmap.

What would change this page

A page built around an absence should say what would fill it. These are the specific developments that would move this from an accounting of a gap to a statistics page like the others.

Any vendor publishing a Claude citation corpus. The fastest path, and entirely outside our control. If a visibility vendor extends its tracking to Claude and publishes even partial figures, the source-skew and citation-density questions get first answers immediately.

Anthropic publishing anything about retrieval behaviour. Unlikely, and it would be the highest-quality possible source. Even a general description of how sources get selected would let the field reason from mechanism rather than inference.

The Index's own first collection. The path this site controls. That produces first-party, openly published figures on the identical query set used for the other engines, which makes them comparable in a way most published cross-engine claims are not.

An independent researcher running it first. Genuinely welcome. The study design above is published partly so someone else can execute it. A second measurement by a different party would be worth more to the field than a first measurement by us.

A significant product change at Anthropic. This one would move the page in the other direction, resetting whatever had been established. Any figure published about a retrieval system carries the date of the product it measured.

When any of these land, this page gets rebuilt rather than amended, and the version history stays visible. A page whose central claim is "nobody has measured this" should not quietly become a page with numbers on it, as though the gap had never been there.

Who uses Claude, and why it matters here

One reason the absence of research on this surface matters more than its user numbers suggest: the audience composition is unusual, and it likely differs from the platforms that have been studied.

Claude has meaningful adoption among developers and technical practitioners, partly through coding assistants and IDE integrations rather than the consumer chat interface. It also sees professional and enterprise use in contexts where careful, well-sourced answers matter more than speed.

That composition has a specific implication for what gets cited. A user base weighted toward technical and professional questions asks different questions than a general consumer base. Documentation, technical references, and specialist writing plausibly matter more here than they would on a platform handling a broader consumer query mix.

Which means borrowing figures from a consumer-weighted platform is doubly unsafe. Not only is the retrieval system different, the query distribution feeding it likely differs too. Two independent reasons the ChatGPT source-skew number should not be assumed to describe Claude.

It also changes the commercial calculation for some sites. If you publish technical documentation, developer-facing content, or specialist professional material, this platform's audience overlaps your own more than raw usage numbers imply. That is an argument for measuring it yourself now rather than waiting for someone else's aggregate, which may be weighted toward query types you do not serve anyway.

Stated carefully: this is reasoning from observable adoption patterns, not from citation data. It is Hypothesis -grade, like most of this page. It is offered as a reason to investigate rather than a conclusion to act on.

Verification status

Everything on this page above the bot-map table is graded Partial at best, and most of it is Hypothesis . This page will be substantially rewritten — not incrementally updated — once the first Claude Citation Report ships.