Where your Claude product exposes web search, use it to draft a cited brief on a narrow, public account trigger, then check the source pages and let a person decide whether to make contact. The Messages API and Managed Agents offer domain controls; a prompt alone may not restrict sources. Score only verified fit signals. Claude cannot prove buying intent, create permission to contact someone or guarantee leads.

Before sending a first email, confirm that the account fits, the public trigger is current and the message offers a relevant next step.

This workflow uses Claude’s Messages API or Claude Managed Agents as concrete examples. Confirm that the Claude product your team chose supports web search and citations; the API and Managed Agents domain settings described here do not automatically apply to other Claude products. If your product has no domain filter, open and verify every cited page yourself.

The human decision still matters. A cited summary is a starting point, not proof that a business wants to buy.

Choose a narrow public signal that connects to a service you can actually deliver.

Examples for a Cape Town or Johannesburg B2B team:

  • a regulated industrial business publishes a new facility, expansion or safety notice;
  • a professional-services firm has a credible service offering but no clear conversion route for a named buyer group;
  • a travel business launches a new destination or product range but its website lacks proof, structure or an enquiry flow for the segment.

Do not search for “companies that need a website.” That judgement belongs after research. Start with a condition that can be observed and cited.

Create a constrained, cited brief

For Claude’s Messages API, set allowed_domains or blocked_domains on the web_search tool; Claude Managed Agents have related domain settings. A prompt asking for official pages does not itself enforce a domain restriction. Check the current web search documentation for the controls available in your product. Then request a short brief grounded in official company pages and a small number of credible independent sources. If you are using another Claude product, verify its search and citation behavior before applying this workflow.

If you use the Messages API, web search has a usage charge in addition to model token costs. Check Anthropic’s current API pricing before estimating a research budget; pricing and product billing can change.

Use a prompt such as:

Research this Johannesburg engineering firm using its official website and public regulatory sources only. Create a short account brief with its named services, location, recent publicly announced changes, the page URLs supporting each statement, and unanswered buyer questions. Mark anything not supported by a source as unverified. Do not infer revenue, private contacts or buying intent.

These constraints make it easier to distinguish sourced facts from unsupported inference. They do not replace checking the cited pages.

Open the citations before you score the account

Review every cited page. Then have a person score the account against these checks:

Test Question
Fit Does the organisation match our sector, geography and service scope?
Trigger Is there a current, public change or gap worth discussing?
Evidence Can we point to the company’s own page or a credible source?
Offer Can we propose a helpful next step without pretending we know their numbers?
Contact route Is there a legitimate public way to start a business conversation?

Score only what you can verify. “Likely needs a redesign” is not evidence. “The service page contains no decision path for procurement teams, despite a published expansion announcement” is a hypothesis you can test in a respectful conversation.

Write the message after the research, not before it

Generic personalisation repeats the company name and congratulates the recipient on a milestone. Useful personalisation names a relevant observation and offers a small, credible next step.

Treat this as sample wording and use it only after checking that the observation is accurate and the proposed deliverable is real: “Your Cape Town training programme explains the technical curriculum clearly. I could not find a route for an operations manager to compare the delivery options before enquiring. I could sketch a shorter decision path for specialist training providers. Would that be useful?”

The message should remain true if the recipient forwards it to a colleague. Do not say Claude found a problem, claim access to private data or imply that an account has a known conversion issue.

Keep research separate from contact permission

Public web information can help you understand a business, but its availability does not by itself settle whether you may collect or use a person’s personal information for outreach. The Information Regulator’s POPIA direct-marketing guidance discusses different requirements for different communication channels.

Keep only the research needed for a defined sales purpose, retain source links, protect personal information and make the applicable objection or unsubscribe route clear. Check the rules for your contact method and the information you collect before sending outreach. Seek legal advice if you are unsure how POPIA applies to your contact channel or planned use of personal information.

Record a reason for each account you disqualify. This gives the team a basis for deciding whether to keep or remove an account instead of using AI to make the list larger.

Measure whether the research changes sales quality

Use a small scorecard for each segment:

  • accounts researched and verified;
  • messages approved and sent;
  • positive replies;
  • first conversations;
  • qualified opportunities;
  • proposals and work won.

After a month, review the disqualification reasons for patterns that may call for changes to the ideal customer profile (ICP), trigger or offer. Those patterns are more useful than the number of briefs an AI tool produced.

To check whether Claude makes the work worthwhile, compare equal-sized groups of accounts from the same segment and trigger type, researched manually and with Claude. Use a 30-day window from first outreach for both groups. Record researcher minutes for both methods. For Claude-assisted accounts, record any web-search charges for the product you use and API token costs if you use the Messages API. Also track the share of citations that need correction and qualified conversations during that window. Treat this as an internal process comparison, not a result you can generalise to other teams.

Make your business easier to verify too

Evidence-based B2B outreach can start a conversation with a relevant account. Inbound visibility gives buyers another way to assess the evidence behind your business.

IDJoy can audit how answer engines currently describe your business, improve the proof behind your service pages and build a link strategy around assets worth citing. Review the AI search visibility scopes.