How Does Claude Work? When It Searches and What It Cites (2026)

Short answer: Claude is a family of large language models built by Anthropic, sold as the claude.ai app, a developer API and the Claude Code coding agent. It answers from what it learned in training, which for the current models runs to June 2026. When it judges that a question needs current information, it searches the web and cites the pages it used. The API docs from Anthropic put it plainly: "Claude determines when to search based on the prompt." On 96 B2B SaaS agency-buying questions, Claude through the API with web search on (DataForSEO, 2026-11-10, 96 questions, one run each) searched on 89 (92.7%). It wrote a mean of 2.18 search queries per searched answer. It listed a median of 10 results and cited a median of 7.

How that compares with the other engines on the same 96 questions:

Claude figures: Claude through the API with web search on (DataForSEO, 2026-11-10, 96 questions, one run each). ChatGPT, Perplexity and AI Overviews figures: every Peec answer for the same 96 questions on 2026-11-09. Method at the end.

What is Claude?

Claude is Anthropic's model family. The models overview opens with "Claude is a family of" large language models "developed by Anthropic." Anthropic describes itself as a public benefit corporation working to secure the benefits of AI and mitigate its risks.

You meet it in three places:

• The app. claude.ai on the web, plus desktop and mobile apps.
• The API. Anthropic calls it the Claude Platform. Developers build Claude into their own products here, and it's where our measurements were taken.
• Claude Code. In Anthropic's words, "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools."

The app and the API don't behave identically. Anthropic says the web interface and mobile apps "use a system prompt to provide up-to-date information, such as the current date, to Claude at the start of every conversation," and that "these system prompt updates do not apply to the Claude API."

As of November 2026, the models overview lists four current models:

All four have a 1M-token context window. Opus 5.5 launched on September 22, 2026 as "the first model in our new Claude 5.5 family."

The consumer plans on claude.com/pricing:

Every plan includes web search, the free one too, so a buyer on the free plan can get a researched, cited answer about your category.

How big is Claude? Usage numbers are thin. Menlo Ventures, an Anthropic investor, estimated in November 2026 that Anthropic's share of enterprise API spend grew from 40% to 55% in a year. Similarweb says claude.ai's traffic "broke sharply upward starting in February 2026," while "ChatGPT still receives more web visits worldwide than Gemini, Claude, and Perplexity combined." I read that as Claude being the smaller consumer engine and a heavy enterprise one, which is exactly where B2B buyers sit.

How does Claude produce an answer?

In two stages, months apart. Training builds what Claude knows. The answer is written later, with or without a search.

Training. Anthropic says Opus 5.5 "was pretrained on large, diverse datasets to acquire language capabilities," then "underwent substantial post-training." The data is "a proprietary mix of publicly available information from online sources, public and private datasets, user data, and synthetic data generated by other models," cleaned with "deduplication and classification." The current four models list a training data cutoff of June 2026.

The constitution. Anthropic trains Claude against a written set of principles. The method started as Constitutional AI, described in a December 2022 paper: "The only human oversight is provided through a list of rules or principles." It has "both a supervised learning and a reinforcement learning phase," and the second uses "'RL from AI Feedback' (RLAIF)" instead of human labels for harmful outputs. In January 2026 Anthropic published a new constitution, saying "its content directly shapes Claude's behavior." It asks Claude to be "Broadly safe," "Broadly ethical," "Compliant with Anthropic's guidelines" and "Genuinely helpful," and to prioritise them in that order when they conflict.

Answer time. Five steps, from Anthropic's docs and our own API runs:

In the app, Claude shows "an indicator that Claude is searching the web" and then a cited answer.

Where does Claude get its information?

From three places: what it learned in training, what it finds on the web when it answers, and what the user hands it. You can move the second one fastest.

Training data comes first. Anthropic's privacy policy lists the sources: "Publicly available information via the Internet," "Datasets that we obtain through commercial agreements with third party businesses," data users and crowd workers provide "(unless users opt out)," explicit feedback, materials flagged for safety review, and "Data that we generate internally." On the public web side, Anthropic's training crawler is ClaudeBot. An earlier Transparency Hub entry says that when Anthropic's general-purpose crawler reads public pages, "we follow industry practices with respect to robots.txt instructions."

Web search at answer time is the second source. It's the only one that knows about your launch last week or the list you got onto yesterday.

The third is whatever the user hands it. Paste a URL and Claude can read it. Anthropic calls this web fetch, which "allows Claude to access and analyze the full content of articles, blog posts, and other web pages you want to discuss." Anthropic's API docs also say web fetch "currently does not support websites dynamically rendered with JavaScript." If your pricing renders client-side, assume, at least through the API, that Claude's fetch sees an empty shell and answers from search results or training instead.

When does Claude search the web?

When it judges the question needs current information. The rules differ a little between the app and the API.

In the app. Web search reached every Claude plan, Free included, on May 27, 2025. The current help article describes two setups. On Team and Enterprise, "An Owner or Primary Owner must first enable web search for the entire workspace," after which members switch it on from the "+" menu. In "the new Claude experience," though, "there's no web search toggle. Claude searches the web when it helps." Users can force it ("include "Search the web" or "Use web search" in your prompt") or tell Claude not to search.

In the API. The web search tool docs list what tends to trigger a search: "Recent events, news, or announcements," "Current prices, rates, scores, or statistics," "Information about specific organizations, people, or products that might have changed" and "Explicit requests to search or look something up." The third one covers questions about your company. Claude answers directly for "Established facts, math, science fundamentals, or coding concepts," for creative writing, for analysis of content it already has, and for greetings. Anthropic says single-fact questions typically take one to three searches and comparative research "can use 10 or more."

What we saw. Through the API, with web search on, Claude searched on 89 of our 96 agency-buying questions. Of the 7 it skipped, six were keyword-style phrases, such as "Boutique SEO agency for B2B SaaS brands," where Claude asked what the user wanted instead of searching. The seventh, "How do I tell a real AEO agency from an SEO agency that just added GEO to its website?", it answered from general knowledge. None of the 7 named LoudFace.

The queries Claude wrote were short and dated:

• 194 queries across the 89 searched answers. Two queries on 71 answers, three on 17, one on 1.
• 86 of the 194 (44.3%) contained "2026", and 80 of the 89 searched answers had at least one query with the year.
• 100 of the 194 contained "best" or "top".
• None used a site: operator.

Third-party data puts Claude's overall search rate far lower, and intent explains the gap. Search Engine Land, summarising figures Jonathan Clark shared from a Profound session, reported that "Claude used web search in 36.6% of prompts, compared with about 90% for ChatGPT," but that recency prompts such as "best XYZ" triggered search 81% of the time. Most of our 96 questions are of that kind. It also reported that "Prompts such as "how does," "what is," and "steps to" were less likely to trigger Claude to search the web. When Claude doesn't search, it can't cite web pages." The sample size wasn't given, so treat it as direction.

On "best agency for X" questions, Claude searched almost every time in our API runs, so live pages win. On "what is X" questions, the third-party data says it often doesn't, so whatever it learned before June 2026 is the answer.

Which search index does Claude use?

Brave Search, by the best evidence available, though Anthropic's product pages don't say so. Neither the help article nor the API docs name the text-search provider. The help article names one provider, for images: "Image search is powered by Bing."

The clearest record is Anthropic's own subprocessor list, and tests back it up:

• Simon Willison quoted a Trust Center notice from Anthropic, which reads "As of March 19, 2025, we have made the following changes," with Brave Search under "Subprocessors added". He then asked Claude to search for pelican facts. Claude "ran a search for "Interesting pelican facts" and the ten results it showed as citations were an exact match for that search on Brave." The function schema also carries the title "BraveSearchParams".
• TechCrunch reported the same change on March 21, 2025, calling the subprocessor list "the list of Anthropic partners who process Claude data." It asked Anthropic for comment; no reply appears on the page.
• A later capture of the list. Anthropic's live subprocessor page renders with JavaScript, so we couldn't read it. A third-party capture by StackPosture, dated June 10, 2026, lists Brave Search for "Web search (all products)" and TurboPuffer for "Web search (all products except Claude for Government)".
• Profound tested 3 commercial queries in March 2025 and found "a statistically significant overlap of 86.7% (13 out of 15 total results)" between Claude's cited results and Brave's top non-sponsored results. In the same test ChatGPT matched Bing's top results at "only 26.7%". Three queries is a small test.
• Search Engine Land's report of another Profound session said Claude "appears to use Brave's top 10 results directly," and also found "much higher overlap with Google rankings, at 64%."

So Brave appears on Anthropic's subprocessor list for web search in both dated readings, Anthropic's notice of March 19, 2025 and a third-party capture of June 10, 2026, and result-matching tests agree. We couldn't read today's live list, and Anthropic's help and API pages name no text-search provider.

Test it cheaply. Run your five most important buyer questions in Brave Search and see whether your pages show up. If they don't, you have a rough early read on why Claude doesn't cite you.

How does Claude choose and show sources?

Anthropic documents the display in detail and the selection barely at all.

The app promises that "Every response includes citations, so you can easily verify sources yourself." A search answer includes "Direct citations to sources," "Source links for further reading" and "Relevant quotes when appropriate."

The API is more specific. Each search result carries a url, a title and page_age ("When the site was last updated"). "Citations are always enabled for web search," and each citation includes cited_text, "Up to 150 characters of the cited content."

How it picks is less clear. The docs describe controls, such as restricting results to set domains or localising them, but no ranking rules. The best clue is from Anthropic's engineering write-up on its Research feature. Testers noticed "our early agents consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources like academic PDFs or personal blogs," and "Adding source quality heuristics to our prompts helped resolve this issue." The team graded output on "source quality (did it use primary sources over lower-quality secondary sources?)".

That was one feature's agent, in 2025, so read it as a direction and not a ranking rule. Anthropic's own team didn't want content farms winning.

On our questions, through the API with web search on, Claude cited most of what it read. Per searched answer it listed a median of 10 results and cited a median of 7. Across all answers, 650 of 853 listed results were cited (76.2%), and every searched answer cited at least one source.

The sources were agency websites and third-party lists. Claude cited 251 distinct domains across its 96 answers. The most-cited domain was loudface.co, in 43 answers. LoudFace publishes many agency lists on exactly these questions. The next nine were all agency or consultant sites, each cited in between 11 and 25 answers. The most-cited single URLs were third-party agency lists: a GEO agencies list on thedigitalelevator.com (11 answers), clearbrand.com's best SaaS SEO agencies (9), and lists on nathanojaokomo.com, breakingb2b.com, seoprofy.com and discoveredlabs.com (8 each).

Wikipedia was cited once. Medium, G2, Clutch and Quora were never cited.

How is Claude different from ChatGPT, Perplexity and AI Overviews?

On our 96 questions, with Claude measured through the API with web search on, each engine trusted a different corpus, and Claude's leaned hardest on agency sites and third-party lists.

It cited most of what it read. Claude cited 76.2% of the results it listed. ChatGPT cited 17.6% and Perplexity 30.4% on the same questions. Google AI Overviews was higher still at 96.5%, but it listed fewer pages per answer. On ChatGPT, being retrieved and being cited were far apart. On Claude, they were close. Our ChatGPT explainer goes into how ChatGPT chooses which sources to cite.

It skipped the social web. No YouTube, Reddit or LinkedIn in 96 answers. AI Overviews cited YouTube in 29 answers, Reddit in 26 and LinkedIn in 18 on the same questions. If your AI search plan is mostly Reddit threads and YouTube videos, it's tuned for Google. On these questions, Claude read elsewhere. Our Gemini explainer has the AI Overviews source mix.

It barely shared sources. When Claude and another engine both cited something for the same question, only about a tenth of Claude's cited domains were also cited by the other engine: 10.7% for AI Overviews, 9.2% for ChatGPT and 9.6% for Perplexity. About half the pairs shared no domain at all (43 of 85 with AI Overviews, 38 of 80 with ChatGPT, 49 of 89 with Perplexity). Search Engine Land's report found "Claude's citations overlapped with ChatGPT's in only 8% of cases." That's close to our 9.2%, though ours counts shared cited domains per question, a different measure.

It showed its queries. Claude exposed its search queries on every searched answer. ChatGPT showed queries on 60 of 96 answers, Perplexity on all 96 (one each), AI Overviews on none. How Perplexity picks what to cite is in our Perplexity explainer.

Winning in ChatGPT tells you little about Claude, so measure each engine on its own.

Which Anthropic crawlers visit your site, and what can you block?

Three robots, and blocking each one does something different. Anthropic's crawler help article spells it out:

Blocking the training crawler is a separate decision from blocking the two search agents. A B2B SaaS that wants to be named in Claude should, at minimum, leave Claude-User and Claude-SearchBot open. Blocking ClaudeBot keeps your future pages out of training, which hurts you on the "what is" and "how does" questions Claude often answers without searching.

The mechanics:

• Anthropic honours "industry standard directives in robots.txt" and supports "the non-standard Crawl-delay extension."
• Rules apply per subdomain, so block or allow "for every subdomain that you wish to opt out from." Your docs and blog subdomains need their own robots.txt.
• IP blocking "may not work correctly or persistently guarantee an opt-out."
• Anthropic says "we will not attempt to bypass CAPTCHAs for the sites we crawl." If a bot wall sits in front of your marketing site, Claude can't read it.
• "As of October 30, 2026," Anthropic recognises a machine-readable training preference signal based on an IETF draft. A robots.txt rule that names ClaudeBot still wins.

Before you change anything, our server-log playbook shows how to read the AI-bot traffic already in your logs.

What does this mean for a B2B SaaS brand that wants to be named in Claude?

Claude names you from training or from what it retrieves. On buyer questions, retrieval dominates. Five moves follow from the evidence above.

1. Keep the doors open. Allow Claude-User and Claude-SearchBot on every subdomain. Make sure key pages render without JavaScript and sit behind no CAPTCHA.

2. Get onto the lists Claude reads. Claude's most-cited URLs on our questions were third-party agency lists. A buyer asking Claude for the best tool in your category gets an answer assembled from comparison pages and roundups on other people's sites. Being on those pages, with an accurate one-line description, is the off-page half of the job. Our guide to getting into the best-of lists AI answers pull from covers how.

3. Match the queries Claude writes. In our API runs with web search on, Claude wrote two queries per answer, 44.3% of them carried the year, and over half contained "best" or "top". A dated title and a visible update date fit the queries it writes. Our piece on fan-out queries explains why the prompt you track isn't the query the engine runs.

4. Write citable sentences. A citation carries up to 150 characters of cited text. A sentence that names your buyer and your proof in one line is the unit Claude can lift. The team behind Anthropic's research feature pushed its agents toward primary sources, which is one more reason original data beats a rewrite of what already ranks.

5. Be the page in the results. On our questions, naming mostly followed having a page in the results. Through the API, Claude named LoudFace in 42 of 96 answers (43.8%), and in 38 of those 42 a loudface.co page was among the results Claude listed or cited; 4 named LoudFace with no loudface.co page in the results. Over 2026-11-03 to 2026-11-09, Peec measured LoudFace named on the same 96 questions in 23.1% of ChatGPT answers, 19.5% of Perplexity and 14.9% of Google AI Overviews. The designs differ, one Claude run per question on one day against seven days of daily Peec runs, so these figures don't rank the engines. We saw the same pattern across engines in our study of whether AI engines name brands they haven't read.

Being cited and being named are still different events. Of 12 agencies we picked in advance and counted by name, the ones Claude named most often across the 96 answers were Omniscient (57 answers), iPullRank (46), First Page Sage (41) and Directive (41). Names travel through other people's lists as well as your own pages.

If you want to use Claude itself for SEO work, or want the detail on getting your own site cited by Claude, read our Claude SEO guide. LoudFace runs this work for B2B SaaS teams through its generative engine optimization service. The free AI visibility audit checks where your brand stands across ChatGPT, Claude, Gemini and Perplexity against your top competitors.

How did we measure this?

We measured Claude's search-trigger rate on B2B SaaS agency buyer questions, meaning how often it actually searches the web versus answering from training data. We took the 96 non-brand questions LoudFace tracks in Peec, after removing the 4 that name LoudFace, and ran them through DataForSEO's Claude LLM Responses endpoint. On 2026-11-10, between 22:22 and 22:24 UTC, each one went once to Claude Sonnet 5.5, with web search on, search left to the model (not forced), a US search location and a 2,048-token output cap. The run cost $5.06 in total.

This is Claude through the API with Anthropic's web search tool. The claude.ai app adds its own system prompt and its own search behaviour, so treat these numbers as the closest measurable proxy for the app, not a reading of it.

For the comparison, we used every Peec answer on 2026-11-09 for the same 96 questions: 96 each for ChatGPT, Perplexity and Google AI Overviews. Naming rates for those engines cover 2026-11-03 to 2026-11-09.

Definitions:

• Listed: a result in the answer's search results or sources list.
• Cited: a result the answer attached a citation to.
• Searched: the answer ran at least one web search.
• Overlap: for the same question on adjacent days, where both engines cited at least one domain, the share of Claude's cited domains the other engine also cited. That gave 85 pairs with AI Overviews, 80 with ChatGPT and 89 with Perplexity.
• Named: the brand's name appears in the answer text.

One run per question is a snapshot. Claude's answers vary, and a second run would move individual numbers. The questions are agency-buying questions, so the source mix describes that slice of B2B SaaS buying, not Claude in general.