How to Research a Company With Claude Before a Sales Call: 5 Free Prompts (2026)

Jump to: set up · copy the prompt pack

Example: a call brief for gong.io from one prompt

Say you have a first call with Gong tomorrow and five minutes to prepare. This is the first prompt in the pack and Claude’s reply. One line in it is wrong (the free trial), one is half right (Gong has a pricing page, though it shows no prices), and one Claude hedged on missed the careers and case-study pages. Our check is under it.

Prompt

Brief me on gong.io before a first sales call. Use AgentData. Keep it to one screen: what they do, who they sell to, the tools they run that matter for a sales conversation (say how each was detected and when it was last seen), and anything you could not find. Don’t guess.

We gave Claude this prompt and AgentData’s lookup result directly, in a separate Claude session, not through the Claude app. With the connector added, Claude makes the lookup itself, so its wording will differ.

What makes this useful is not the length. It is that Claude says how each fact was found, separates a DNS record (a setting Gong made itself) from a page scan, and lists what it could not find instead of filling the gaps. That tells you which lines to repeat on the call and which to check first. The data behind it is on Gong’s public company page, which anyone can open.

Disclosure: AgentData is our product. We use it here because Claude can use it with no account and no key. The replies are Claude’s; the checks under them are ours.

Set up: connect company data to Claude

1. In Claude, open Customize, then Connectors, choose Add, then Add custom connector, and name it AgentData.
2. Paste https://mcp.agentdata.run/mcp as the server address and leave sign-in off. On Team and Enterprise plans, an Owner adds it under Organization settings, then Connectors, and members then connect it under Customize, then Connectors. Free plans can add one custom connector.
3. Paste a prompt from the pack below with a real company’s website address.

Claude Code users can add it with one command instead:

claude mcp add --transport http agentdata https://mcp.agentdata.run/mcp

If we have not profiled the company yet, the lookup crawls it on request, usually within a few minutes (about an hour when busy), and the reply says when to ask again. The full setup for Claude, ChatGPT, Cursor and VS Code is in the MCP setup guide.

The five prompts, with Claude’s real replies

Each prompt below was run on 22 October 2026 for a seller of a website analytics and session-replay tool. Swap in what you sell. Prompt 1, the call brief, is the example above. As with the example, we gave Claude the lookup results directly rather than through the Claude app.

What AgentData is thin on: news, and the person you are meeting. For a company’s recent launches, funding or hiring, turn on web search in the same chat and ask for the last 90 days with a link and a date for each item, then check the links. For the person, paste their role and anything they have published, and ask Claude what they are likely to care about, marked as guesses. With an API key from a trial or plan, AgentData adds named people where we have found them.

Prompt 2: which of their tools matter to what you sell

Use it when you need an angle: is there a rival to replace, something you plug into, or an empty space? For Gong, Claude found a likely rival in place and called it a switch conversation, not a first purchase.

Prompt

I sell a website analytics and session-replay tool. Using AgentData, which tools at gong.io matter to me? Split them into: a competitor already in place, tools mine would sit alongside, and gaps. Give the detection method and last-seen date for each, and say how sure you are.

Prompt 3: which account to call first

Paste your own list. Claude looks each company up and ranks them on evidence, which is quicker than five browser tabs and says why.

Prompt

I sell a website analytics and session-replay tool. Here are five accounts: gong.io, pipedrive.com, airtable.com, typeform.com, planable.io. Look each one up in AgentData and rank them for a first call, with one line of evidence each. Size is a website estimate, so rank on the tools they run, not on company size.

Prompt 4: discovery questions you can defend

Questions built from what a company actually runs sound prepared rather than generic. Asking for a source on each one keeps Claude honest.

Prompt

From the gong.io profile in AgentData, write five discovery questions for my first call. Each one should come from something AgentData found (name it) or from something it could not find.

Prompt 5: a first email with one true fact

The rule that matters is the last line of the prompt: one fact the reader would recognise, and nothing Claude cannot back up. Note that it picked the fact seen twice over the sharper one seen once.

Prompt

Write a first email of no more than three sentences to someone in marketing at gong.io about what I sell. Use one fact from AgentData that a person at Gong would recognise as true, say in a line under the email which fact you used and how AgentData found it, and use nothing you can’t back up.

With an API key, the same lookups go further: they return the email addresses we found at a company, each with its verification status, and named people where we have found them (none yet for Gong), so prompt 5 can go to a person. In Claude Code the key goes in as a header (the command is in the setup guide). Anthropic’s connector help page (read 22 October 2026) says a custom connector can now send a fixed request header too; we have not tested a key that way yet. Start a 7-day free trial, no card needed.

Copy the whole pack

All five, with the parts to change in square brackets.

1. Call brief
Brief me on [company.com] before a first sales call. Use AgentData. Keep it to one screen: what they do, who they sell to, the tools they run that matter for a sales conversation (say how each was detected and when it was last seen), and anything you could not find. Don't guess.

2. Stack fit
I sell [what you sell]. Using AgentData, which tools at [company.com] matter to me? Split them into: a competitor already in place, tools mine would sit alongside, and gaps. Give the detection method and last-seen date for each, and say how sure you are.

3. Who to call first
I sell [what you sell]. Here are my accounts: [a.com, b.com, c.com, d.com, e.com]. Look each one up in AgentData and rank them for a first call, with one line of evidence each. Size is a website estimate, so rank on the tools they run, not on company size.

4. Discovery questions
From the [company.com] profile in AgentData, write five discovery questions for my first call. Each one should come from something AgentData found (name it) or from something it could not find.

5. First email
Write a first email of no more than three sentences to someone in [team] at [company.com] about what I sell. Use one fact from AgentData that a person there would recognise as true, say in a line under the email which fact you used and how AgentData found it, and use nothing you can't back up.

Other ways to give Claude company data, compared

Claude in the app (web, desktop and mobile) knows what it learned in training and, with web search on, what pages say. It does not read a site’s scripts or DNS records, which is where most of the tool evidence above came from. These are the main ways to add company data, from each vendor’s own pages, read on 22 October 2026.

Option	To start	What it adds
AgentData (ours)	The connector address. No account or key needed	Company profile, tools with how and when each was found, email format; customer lists for 10 technologies on Pro ($99 a month)
Clay connector	A Clay account (made for you when you connect); 500 free Clay credits	Contact details, enrichment providers, research agents, and signals including tech stack, job changes and funding
ZoomInfo MCP	A ZoomInfo account. Search is free; enrichment and account research use credits	Contacts, company records and intent from ZoomInfo’s database
Claude with web search	Nothing extra; included in your plan	What public pages say; no script or DNS checks

Sources, read on 22 October 2026: Clay’s Clay in Claude announcement (26 January 2026); ZoomInfo’s Claude MCP guide (30 July 2026), which says enrichment uses bulk data credits, up to 25 per call, and account research uses AI action credits.

Save it so it runs every time

• In a Claude project, paste the pack into the project’s instructions, then start each chat with just the company’s address.
• In Cowork, Anthropic’s Sales plugin has an account-research skill that reads the tools you connect for it, such as your CRM, email and the web, and a call-summary skill for the follow-up (Anthropic’s own guide, read 22 October 2026). Add AgentData to that list for the accounts you have never spoken to.
• In Claude Code, the same prompts work at the command line, and you can loop over a list. Our Claude Code prospecting guide shows how.

What the data can’t tell you

The catch: AgentData reads a company’s own website, DNS records and HTTP headers. It cannot see tools that leave no trace there, such as a CRM used only inside the sales team, and a page scan can miss scripts, as it missed Google Tag Manager on gong.io. Yes/no flags like “has a pricing page” come from checks of the site’s pages and can be wrong, as four were for Gong. A tool’s first-seen date is when we first saw it, not when the company bought it. Size is an estimate from the website. Treat medium and low confidence as “check first”.

From AgentData’s MCP guide and published limits, read on 22 October 2026. When an allowance runs out, the reply says when it resets. Coverage and freshness are published hourly on the live data page: on 22 October 2026, 30 million companies profiled and 4,254 technologies detected.

How we checked this

On 22 October 2026 we looked each company up with AgentData’s lookup tool, with no key. We gave Claude each prompt in a separate Claude session with those results and nothing else, set up as if it had made the lookups through the connector. We did not run them through the connector in the Claude app; there, Claude makes the lookups itself and its wording will differ. Each reply was run once and is shown unedited. We then checked the claims about Gong against gong.io, its DNS records and its tag manager in a browser the same day; the other four companies were checked against their profiles only.

Related reading

• How to find companies using a specific technology, with the evidence for each
• Give your own AI agent company data, with code
• Sales prospecting in Claude Code with company data
• The MCP servers for sales and B2B data, scored