TaskbyTask.Practical AI guides for work
Claude or ChatGPT · Practical walkthrough

Research an account before a call without inventing the news

You have eleven minutes before a call with an account you haven't touched since March, so you ask an assistant what's going on with them. What comes back is fluent and current-sounding, and somewhere in the middle is a sentence that will make you look ridiculous if you repeat it.

Task by Task15 min read + practice
A built cargo harbour stands beside an outlined extension that has not been constructed.

What you'll make

a short claims table about one account, with a source link and a publication date on every row, plus a list of what you couldn’t verify, which becomes your questions for the call.

What you'll need

ChatGPT or Claude with web search available in your work account, the company’s official name and website, the reason for the call, and about ten minutes. Gemini or Copilot can do the same job if they show sources you can open.

Example result.

Verified:   New operations director, Dana Reed, from 1 September (company news page, 14 Aug 2026)
Verified:   Plans a second packing site for early 2027, subject to planning approval (company news page, 20 May 2026)
Ask:        "Where have you got to on packaging suppliers?"
Ask:        "Is the growth plan still UK-focused?"

1. Decide what you need, in three questions

Before you type anything, write the three things that would change how you run the call. For most account conversations they’re some version of:

  • Has anything changed in their business that affects what we sell them?
  • Has anything changed in who makes the decision?
  • Is there a reason this is a bad week to ask for something?

Note the exact company name and website; similar names produce plausible but irrelevant news. Leave pipeline notes, pricing and your internal view of the account out of a web search: none of it improves the results.

Why this works: a prompt that asks for “everything about” a company gets a brochure, and the brochure is where invention lives.

What you should see: three questions, a company name and a domain.

2. Turn search on, and say so in the prompt

In ChatGPT, open the + menu in the message box and choose Web search. In Claude, click + at the lower left of the chat and select Web search (a checkmark shows it’s on); in the newer Claude experience there’s no toggle and Claude decides when to search. On a Claude Team or Enterprise plan, an Owner or Primary Owner first has to enable web search in Organization settings › Capabilities.

Then say “Search the web” in the prompt as well. Belt and braces costs four words.

Why this works: a model with no search and a request for recent news will produce recent-sounding news, because that’s what was asked for.

What you should see: search switched on. If it isn’t available, stop: an answer from memory is not a current brief.

3. Ask for a table, not a briefing

This is the whole technique. Change the shape of the answer and the honesty follows.

Search the web. Research [company name], [website domain], for
[your three questions], over the last [time period].

Return a table only, one row per claim, with these columns:
Claim | Source link | Date published | Event date | Source type

- Every claim must come from a page you found in this search. If a
  claim has no source, write "no source found" and keep the row.
- Date published is the date on the page, or "undated".
- Keep announced plans separate from things that have happened.
- Source type is one of: company's own site / named publication /
  official register / inferred.
- No summary, narrative or recommendations.

[company name] and [website domain] pin down the right entity; [your three questions] come from step 1; [time period] is usually 6–12 months.

Why this works: in a table, an unsourced row looks as bare as it is. “No source found” gives the model a legitimate way to produce nothing.

What you should see: five to ten rows, some with “no source found” or “undated”. That’s the table doing its job.

Pick the three rows you’d most like to mention and open their links. For each one:

  • Does the page say the thing? The usual problem is a real source supporting a weaker claim. “Opened a new site” turns out to be “plans to open a site”.
  • Whose words are they? The company’s own page is current and flattering; trade coverage is interpreted and sometimes wrong.
  • Is it the right company? Calderstone Group plc is not Calderstone Foods Ltd. A dull error, and an extremely common one.

A link that doesn’t open is not a source. Rows that survive go into your note; rows that don’t go to step 5.

This isn’t the method for everything. For revenue, ownership, credit or insolvency, go to the statutory record, such as Companies House. And don’t use it to build a profile of a named person: confirm their role on the company’s own site and stop there.

Why this works: three links take four minutes and cover what you’d say out loud.

What you should see: each row marked supported, weaker than claimed, or unverified.

5. Turn what you couldn’t verify into questions

Don’t delete the unverified rows. Rewrite them as questions with no premise in them. “I saw you’re consolidating suppliers, is that right?” is a bad question, because you didn’t see it; a model mentioned it. “How are you thinking about suppliers this year?” is the same curiosity with the fiction removed.

Why this works: an unverified claim is something you suspect matters and don’t know, which is exactly what a good question is for.

What you should see: two or three open questions, ready for the call.

Check the result

  • Every claim you plan to say has a link you’ve opened.
  • Every row has a publication date or says “undated”.
  • Plans read as plans, with their conditions (“subject to planning approval”).
  • Your questions don’t assume anything you couldn’t verify.

The usual failure: a plan becomes an event. An undated article or a loose summary turns “plans to open in 2027” into “opened”, and it reaches the table looking just as sourced as everything else. Spot it by checking every past-tense row against a dated page.

Correct this row: [claim]. The page [link] says [what it says]. It
supports [the narrower fact], not [the overclaim]. Keep the publication
and event dates separate and move any remaining uncertainty to the
questions list. Do not replace it with a guess.

[claim] and [link] identify the row; [what it says] and [the narrower fact] come from the page; [the overclaim] is what the row wrongly added.

Worked example: Calderstone Foods

Fictional example. You have a renewal call tomorrow with Calderstone Foods, a mid-sized food manufacturer. The table came back like this (links described, not written out):

Claim                                    | Source              | Published  | Type
New operations director appointed        | company news page   | 14 Aug 26  | company's own site
Opened a second packing site             | trade magazine      | undated    | named publication
Moving to a single packaging supplier    | no source found     | —          | inferred
Expanding into Ireland next year         | no source found     | —          | inferred
Certified to a new food safety standard  | certifier's register| 02 Jun 26  | official register
  • Row two is the most dangerous row on the table. It has a source and no date, and it uses the past tense. Searching the company’s own site found the 20 May announcement: a site planned for early 2027, subject to planning approval. The repair prompt changed the row to “Plans a second packing site for early 2027, subject to planning approval (company news page, 20 May 2026)”.
  • Rows three and four are the model being helpful. Plausible, the sort of thing a food manufacturer does, no source. Neither goes near the call as a statement; both became the questions in the example result.
  • Rows one and five hold up. Row one is checkable in thirty seconds on their own site, and a renewal looks different to someone in their first month. Row five is on the certifier’s own register.

Ten minutes produced two verified facts, one corrected row and two good questions: a better position than a confident paragraph about a site that hasn’t been built.

Try it

Inspect a research brief that makes a plan sound completed

Example responses, written to show what each prompt tends to produce.

Weak: Sales hook

The prompt

Find an exciting sales angle from this company news.

Fictional practice source:
Fictional announcement dated 20 May 2026: Calderstone Foods plans a packing site for early 2027, subject to planning approval. An undated article says capacity options are being assessed.

Example response

Select a phrase in the response to inspect its support or weakness.

Why this prompt changes the task

An enthusiastic audience brief must not outrun source status.

Include a citation

The prompt

Summarise the announcement with its date and a source reference.

Fictional practice source:
Fictional announcement dated 20 May 2026: Calderstone Foods plans a packing site for early 2027, subject to planning approval. An undated article says capacity options are being assessed.

Example response

Select a phrase in the response to inspect its support or weakness.

Why this prompt changes the task

A source reference helps inspection but does not prevent a claim losing its qualifier.

Evidence and outreach question

The prompt

Separate announced plans from completed events. Use the dated announcement, retain early 2027 and planning-approval condition. Keep the undated capacity article separate. Suggest one question without claiming a need or decision.

Fictional practice source:
Fictional announcement dated 20 May 2026: Calderstone Foods plans a packing site for early 2027, subject to planning approval. An undated article says capacity options are being assessed.

Example response

Select a phrase in the response to inspect its support or weakness.

Why this prompt changes the task

Use research to prepare a question, not fabricate an account event. Check the current primary source before real outreach.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Search one named source before widening

When a broad search returns the wrong company or old news, narrow it. A restricted search can miss coverage, so record what you checked rather than calling it complete.

Search [official domain] for dated updates between [start] and [end]
about [topic]. Return direct links and dates. If none is found, say
which pages you checked and leave the question open.

Then widen to a named trade or regulatory source if it serves the question.

When there genuinely is no news

Most companies don’t generate news. A stable, privately held business may go years without appearing anywhere, and that’s when a research prompt is most tempted to fill the gap. A table with three “no source found” rows is a good result. It tells you to spend the call asking, which was probably the right plan anyway. It doesn’t prove nothing has happened: you may have searched the wrong entity or hit a site you couldn’t read.

Combine public research with account history

Keep public findings, customer statements and internal notes in separate sections, each with its date. An internal note saying the customer plans to expand may predate a public cancellation. Build a one-page account brief covers the internal side; for comparing several companies, see competitor research you can check.

Sources

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.