How Claude Picks Its Sources When Buyers Ask About Software
· 6 min read · By Perciva Team
When a buyer asks Claude to compare vendors or shortlist software, Claude draws on two layers: what it learned during training, and — when the question calls for current facts — live web search, which retrieves pages and cites them in the answer. Which layer dominates determines everything about how your product is represented: training knowledge is broad but frozen at a knowledge cutoff, while searched answers are current but depend on which pages Claude's retrieval surfaces.
Understanding how each layer forms its picture of your product tells you exactly where to intervene. This post walks through both, plus the crawler-level plumbing that decides whether your site can be part of the picture at all.
Layer 1: What Claude Learned in Training
Claude's parametric knowledge of your product is a compression of everything written about you across the public web up to its training cutoff: your site, documentation, review platforms, comparison articles, news, and community discussion. Three properties matter for B2B teams:
- Consensus wins. A claim that appears consistently across many independent sources becomes something Claude states confidently. A claim that exists only on your site tends to be recalled weakly or hedged.
- It ages in one direction. Whatever the web said about you before the cutoff is what Claude remembers — old pricing, discontinued tiers, your positioning from two rebrands ago. Fixes you shipped since then do not exist in this layer.
- Anthropic's training crawler is the intake. ClaudeBot is the crawler Anthropic uses to gather public web data. Sites that block it are simply absent from the firsthand record — Claude then knows you only through what third parties wrote.
Layer 2: What Claude Finds When It Searches
Since Anthropic added web search to Claude, buyer-style questions about pricing, current features, or "best tools in 2026" can trigger live retrieval. Claude issues searches, reads the results, and composes an answer with citations to the pages it used. In this mode, Claude behaves much like other search-grounded engines: the winners are pages that are retrievable, current, and quotable.
The plumbing has its own user agents worth knowing:
| Crawler / agent | Purpose | Practical implication |
|---|
| ClaudeBot | Collecting public web data for training | Blocking it removes your site from future training snapshots |
| Claude-SearchBot | Indexing to improve search result quality | Blocking it hurts your retrievability in searched answers |
| Claude-User | Fetching a page because a user asked | Blocking it breaks direct "look at this page" requests |
As with other engines, audit robots.txt, CDN bot rules, and server logs — an AI crawler you have silently blocked at the WAF level costs you visibility no amount of content work can recover.
What Makes Claude Different in Practice
Teams that monitor multiple engines notice consistent behavioral differences in how Claude handles buyer questions:
- It hedges honestly. Claude is comparatively willing to say pricing may have changed or that it is unsure — which means thin or stale public information about you produces visibly cautious answers, not confident advocacy.
- It resists superlatives. Claude tends to frame recommendations as trade-offs ("X suits teams that need..., Y suits teams that...") rather than crowning a single winner. Your goal is to be the product whose fit-description is accurate and compelling, not to chase a "best overall" slot that Claude rarely awards.
- Balanced sources fit its style. Content that acknowledges trade-offs gives Claude exactly the material its answer style prefers — another reason honest comparison pages outperform one-sided ones.
We compared this behavior directly in Gemini vs Claude for product evaluation prompts.
The Playbook for Claude Visibility
- Unblock the three agents in robots.txt and your CDN, and verify visits in logs.
- Fix the consensus record. Audit the top third-party pages about your product (reviews, comparisons, directories) and correct stale facts — this is what both training and search synthesize from.
- Publish trade-off-honest comparison content. State clearly who you are for and who you are not for; this is the framing Claude reuses.
- Keep evaluation-critical facts current and dated — pricing, security posture, integrations — so searched answers quote fresh, specific pages.
- Test with real buyer phrasings. Run your actual evaluation prompts against Claude and note when it searches versus answers from memory; the same question phrased with "in 2026" or "current pricing" will pull the live layer. Build the prompt set from questions buyers genuinely ask, not marketing keywords.
Where Claude Shows Up in the B2B Buying Process
Claude has a distinctive footprint among buyers, and it changes what you optimize:
- Technical evaluators lean on it. Developers and technical buyers use Claude heavily for tooling questions, which makes accurate representation of your docs, APIs, and limits disproportionately valuable if you sell to engineers.
- Documents get pasted in. Buyers hand Claude long documents — RFP responses, security questionnaires, your own whitepapers — and ask for summaries and comparisons. In those moments Claude is reading your material directly; clear, well-structured documents survive summarization with your framing intact, while dense marketing prose gets flattened into generic claims.
- Enterprise deployments mean invisible conversations. Claude runs inside many companies' internal tools, and the conversations you can never observe still draw on the same two layers you can influence: the training record and the searchable web.
Common Mistakes
- The forgotten blanket block. Many sites blocked every AI crawler during the first backlash wave and never revisited the decision; check whether your robots.txt still blocks the agents above.
- Treating hedges as noise. When Claude says pricing "may have changed", that is a signal your public record is thin or stale on that exact point — a fixable gap, not model quirkiness.
- Superlative-stuffed copy. Claude's style discounts unsupported "leading" and "best-in-class" claims; specific, verifiable statements travel much further.
- Neglecting documentation. For technical categories, docs are often the most-retrieved and most-pasted material you own — treat them as a marketing surface, not just a support surface.
None of these require new budget — they are audit items. An afternoon spent on robots.txt, on the facts that trigger hedges, and on your top ten documentation pages moves more than a quarter of net-new content.
Watching for Drift
Claude's picture of you shifts on two clocks: model releases move the training layer in jumps, while the searchable web moves continuously. An answer that was accurate at your last check can silently change after either. Because engines disagree — often substantially — it pays to track Claude alongside ChatGPT, Gemini, and Perplexity on the same prompt set and diff the answers over time; why engines disagree is a topic of its own. Perciva runs this comparison continuously so you see the moment Claude's story about you changes.
Give special attention to prompts where Claude's hedging changes. A shift from confident, accurate description to hedged uncertainty often precedes a wrong answer by one model or index refresh — the record thinned before it broke. Treat new hedges on commercially important prompts as early warnings, and refill the record before a rival's version of the story fills the vacuum.
The Bottom Line
Claude picks its sources the way a careful analyst would: prefer consensus, hedge when the record is thin, check the live web when facts might have moved. Your job is to make the record about your product deep, consistent, and current — on your site and on the third-party pages that outnumber it. Do that, and both of Claude's layers start telling the story you want buyers to hear.