Why does Claude answer differently in different apps?

"Claude" is not one fixed product. It's a family of models that many apps call, and each app shapes the answer. Three settings do most of the work.

• The model. Apps choose which Claude model to run. As of October 2026, the Claude.ai system prompt names Opus 5 as the currently selected model, with other models selectable. An app built on the API can pin a different one. Models have different training data and different knowledge cutoffs, which we cover in whether switching AI models changes what it says about your brand.
• The system prompt. Every app that builds on Claude writes its own instructions. A system prompt can tell Claude to stay brief, avoid naming vendors, prefer certain sources, or search only when asked.
• Search state. Claude decides question by question whether to search the web. That decision is steered by the app's instructions, and in some apps it's switched off entirely.

Change any one of these and the answer about your brand can change. Change all three and you're effectively asking a different assistant.

Search is the biggest lever, because it decides whether Claude can see your current site at all.

On Claude.ai

Web search is a user toggle. As of October 2026, Anthropic publishes no default for it. Two buyers asking the same question on Claude.ai can get different answers simply because one of them turned search on.

Through the API

Developers add search with Anthropic's web search tool. The API's org-level default is on, but that setting does not describe Claude.ai consumer behavior. The developer still decides whether to offer the tool and how the system prompt frames it. Each search costs $10 per 1,000 searches plus tokens, according to Anthropic's documentation, so some integrators limit or skip it.

Inside third-party products

When a support bot, writing tool or research app runs on Claude, its builder writes the system prompt. Claude then decides within those instructions. You can't see that prompt from outside, which is why comparing API answers with consumer app answers rarely lines up exactly.

With search off, Claude answers from what it learned in training. That knowledge stops at a cutoff date. Anything you launched, repriced or renamed after that date isn't in the answer.

This is the stage nobody controls this quarter. You can't edit training data, and the effect of allowing or blocking training crawlers is unmeasured, including by the vendors. Anthropic separates its crawlers: ClaudeBot collects training data, Claude-SearchBot supports search quality, and Claude-User fetches pages when a person asks.

Say you run a hypothetical project management tool that cut its prices in spring 2026. A buyer asks Claude.ai with search off and gets your old pricing. A colleague asks with search on and gets the current page. Neither answer is a bug. They're two different retrieval states.

The mechanism is documented. The size of the gap is not.

• Documented: Claude's search is discretionary, steered by the integrator, and a toggle on Claude.ai. Anthropic's search results expose a URL, title and page_age, and Claude cites text spans of up to 150 characters, which differs from how Perplexity cites whole sources.
• Documented: most AI crawlers, including ClaudeBot and Claude-SearchBot, do not execute JavaScript. Vercel found Claude's crawler fetched JavaScript files on 23.84% of requests but did not run them.
• Not measured: no published study reports how often Claude's brand answers differ between apps, or by how much.

The wider research says the same thing about all engines. A 2026 critical survey of 45 GEO studies found no technique with a stable, cross-platform causal effect on discoverability. Any single score for "Claude visibility" averages answers given with and without search, so treat it as a sample, not a fixed property of your brand.

• User-triggered fetches. When someone pastes your URL into Claude, Claude-User may fetch it directly. That's a different path from search and from training.
• Blocked at the edge. Your robots.txt can allow Claude-SearchBot while your firewall or CDN returns a 403. Anthropic warns that IP blocking can stop it reading your robots.txt at all.
• JavaScript-only content. If your pricing or product names load after JavaScript runs, a search-enabled Claude may still miss them.
• Memory and personal settings. Logged-in sessions can carry context that shifts answers. Test logged out where you can.
• Location. Answers can also vary by market, covered in whether AI describes your brand differently by country.

Start with the surfaces your buyers actually use, not every app that runs Claude.

1. List the surfaces. Ask sales and support where prospects research: Claude.ai, a Claude-powered tool in their workflow, or an API integration you know of.
2. Test each one with real buyer questions. Run the same three to five questions on each surface. Record the model, whether search was on, and whether sources were cited.
3. Compare against your facts. Note wrong prices, outdated product names and missing features. A wrong fact with no citation usually points to training data. A wrong fact with a citation points to the page cited.
4. Fix what you control. Check that robots.txt allows Claude-SearchBot, that your firewall lets it through, and that key facts are in the HTML before JavaScript runs. The Claude SEO guide covers each step.
5. Repeat on a schedule. Models and defaults change. Tracking Claude's answers about your brand explains how to keep the comparison consistent.

For a quick first read, the free checker on this site runs up to three buyer questions and scores the answers. It's a snapshot, so pair it with your own surface-by-surface tests.

Does Claude.ai search the web by default? Anthropic does not publish a default for Claude.ai as of October 2026. Web search is a user toggle, so whether a given answer used live search depends on the person asking. The Claude API defaults web search on at the organization level, but Anthropic notes that setting does not describe Claude.ai consumer behavior. If an answer shows no citations, assume it may have come from training data.

Can I see the system prompt a Claude-powered app uses? Usually not. Integrators write their own system prompts and rarely publish them. That prompt can steer whether Claude searches, how long it answers, and whether it names vendors. Because you can't inspect it, the practical approach is to test the app directly with your buyers' questions and record what comes back, rather than assuming it matches Claude.ai.

Does blocking ClaudeBot remove me from Claude's search answers? Anthropic documents separate agents. ClaudeBot collects training data, Claude-SearchBot supports search quality, and Claude-User fetches pages a person asks for. Blocking one is not the same as blocking the others. Check your robots.txt per agent, and confirm your firewall or CDN isn't returning a 403 to the agent you meant to allow.

Why does the Claude API give a different answer from Claude.ai? The API call may use a different model, a different system prompt, and a different search setting. Claude.ai adds its own instructions and lets the user toggle search. Your API test sends whatever you configured. Match the model and search state as closely as you can, and still expect differences, because the consumer app's full setup isn't something you can reproduce exactly.

Will adding schema or an llms.txt make Claude describe me correctly? There's no evidence for that. Anthropic doesn't document reading llms.txt, and the best matched test of schema found no meaningful lift in AI citations. What has evidence behind it is access: let Claude-SearchBot in, keep it reachable through your firewall, and put specific facts like prices in the HTML so a search-enabled Claude can fetch and quote them.