Claude Gets Your Company Wrong: What to Do

Claude produces two kinds of answers, and they fail for different reasons.

• Search off (parametric): Claude answers from what it learned in training. The knowledge is frozen at a cutoff date that Anthropic publishes per model.
• Search on (retrieval): Claude runs a web search, reads the results, and cites the text it used.

As of October 2026, web search in Claude.ai is a user toggle, and Anthropic publishes no default for it. On the API, developers turn the web search tool on, and the integrator's system prompt steers when Claude uses it. So the same question can get a grounded answer for one user and an ungrounded one for another.

The practical test is citations. If the wrong answer shows source links or quoted snippets, it was search-on, and you can trace it to a page. If it shows no sources, assume it came from training data.

Then classify the error itself. A stale fact (last year's pricing) needs a different fix than an invented one (a feature you never had). Our guide to telling a hallucination from an outdated answer walks through the difference.

A search-off answer reflects whatever Claude saw before its knowledge cutoff. If you changed your pricing, renamed a product or dropped a feature after that date, the model cannot know. If the old fact was widely repeated online, the model may repeat it confidently.

Nothing a publisher does this quarter reaches a search-off answer. New content only enters parametric knowledge when a future model is trained, and that latency is a full model generation. Nobody, including the vendors, has published a measurement of how much a given page affects what a later model recalls.

What that means for you:

• Publish the correct fact clearly now, so it exists for future training runs and for search-on answers today.
• Don't expect a fast change in answers that show no sources.
• Treat the next model release as the earliest point a search-off answer might shift.

For the full approach to facts that changed after a cutoff, see correcting outdated facts past a knowledge cutoff.

Anthropic documents three agents in its help center, and each has its own robots.txt rule:

• ClaudeBot: collects content that may be used for training.
• Claude-SearchBot: crawls to improve Claude's search results.
• Claude-User: fetches pages when a user's request triggers it.

Blocking ClaudeBot signals that your content should not be used in training. It does not remove you from Claude's search results, and allowing it does not make Claude search cite you. The two decisions are independent. A common mistake is a blanket AI block that catches the search agent along with the training one.

If you want to opt out of training but stay reachable for search, the rules look like this:

User-agent: ClaudeBot
Disallow: /

User-agent: Claude-SearchBot
Allow: /

Robots.txt is necessary but not sufficient. Anthropic also warns that IP blocking impedes its ability to read your robots.txt file at all.

When Claude searches, it can only quote what its crawler actually received. Three access checks come before anything else.

Is the fact in the raw HTML?

Vercel's crawler study found that Claude's crawlers fetched JavaScript files (23.84% of requests) but did not execute them, so they can't read client-side rendered content. Independent testing re-confirmed this for ClaudeBot and Claude-SearchBot as of June 2026. If your pricing or product list only appears after JavaScript runs, Claude search may never see it. Render that content on the server.

Does your firewall let Claude in?

A CDN or bot-management rule can return a 403 to a crawler that robots.txt allows. Nothing in your dashboards will tell you. Since 2026-09-15, Cloudflare blocks Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers, while Search stays allowed by default. Check your zone's settings rather than assuming. The free AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

Is the fact stated plainly?

Anthropic's API cites text spans of up to 150 characters. A specific sentence ("Our Pro plan costs $X a month") on a fetchable page gives Claude something it can quote. That is the only on-page content pattern with a positive experimental signal, and it is still not a guarantee.

If a search-on answer is wrong, open the cited sources. The error usually lives in one of two places.

On your own site. Check whether the same fact appears differently on your pricing page, about page, footer, docs or old blog posts. A stale page that is still live and crawlable is a valid source for Claude to quote. Update it, redirect it, or remove the old figure.

On third-party sites. Directories, review sites and old press coverage often carry outdated facts. When engines cite other sites instead of yours, being named correctly on those sites is associated with appearing correctly in the answer. The evidence shows association, not a promised effect. Our playbook on asking third-party sites to fix wrong facts covers how to request changes.

Watch for a rarer risk too. Researchers have shown that manipulated web content can steer LLM search agents once it lands in the results (arXiv:2606.16821). If a page about you contains hidden instructions, that is a reputational exposure to report, not a tactic to copy.

Skip the folklore. Schema markup, llms.txt files and "AI-friendly" rewrites have no credible evidence behind them as fixes for wrong answers.

Work in this order:

1. Collect the wrong answers, and note whether each one showed sources.
2. For sourced answers, list the cited URLs and fix or request fixes on each.
3. Confirm robots.txt allows Claude-SearchBot, and that your CDN returns a 200 to it.
4. Confirm key facts are in server-rendered HTML.
5. Remove contradictions across your own pages.
6. Re-ask the same questions with search on, logged out where possible, and compare.

Anthropic does not publish how Claude ranks search results, and no tool can see the queries Claude actually runs. Measure Claude on its own rather than blending it with other engines. The free checker on this site gives a snapshot of how AI answers describe you for three buyer questions. For which facts tend to break first, see the brand facts AI gets wrong most often.