How do you research an account in five minutes with AI?

Why does the AI get the pages instead of the question?

Ask a model "what does this company do and what changed this year" and you get fluent text with no promise any of it is true. OpenAI's September 5, 2025 research post defines hallucinations as "plausible but false statements generated by language models" and says training and evaluation "reward guessing over acknowledging uncertainty."

The vendors say the same about search. Per OpenAI's ChatGPT search help page (read 2026-10-06), "Search results and citations can be incomplete, outdated, or incorrect," and it tells you to open a cited source and check that it supports the answer.

Build a closed box instead. If the model only has three pages you opened, every sentence either came from them or was invented, and you can tell which in a minute.

Swipe to see more →

What are the five minutes, minute by minute?

Use a timer. The cap is the point.

Swipe to see more →

Page not found in 40 seconds? Write "not found" and move on. A thin brief in five minutes beats a fat one in twenty. Whether an account earns five minutes at all is covered in how much time to research a prospect.

Why these three? The about page says what they sell, open roles show where they are about to spend, and the newest dated item shows what changed (vocabulary in sales trigger events). For a public company, SEC's EDGAR finds filings by name or ticker, and full-text search covers filings since 2001 (SEC pages, read 2026-10-06); the newest 8-K counts as page 3.

Both prompts are ours, built on the role-input-output-rules recipe from AI prompts for sales prospecting.

Prompt 1, the brief builder:

Below are three pages I copied from [COMPANY]'s own website. Using ONLY these pages, write a brief of no more than 80 words, not counting the tags with these lines: what they sell and who buys it; the newest dated change, with the date; what their open roles say they are about to spend on; one hypothesis about a problem this creates for [THE JOB TITLE WE SELL TO], labeled "inference". Put [P1], [P2] or [P3] after every factual line to show which page it came from. If a line would need a fact that is not in the pages, write "not in pages" instead of guessing. PAGE 1 (about): [PASTE] PAGE 2 (jobs): [PASTE] PAGE 3 (news, dated): [PASTE]

Prompt 2, the opener:

Here is a verified brief about [COMPANY]: [PASTE BRIEF]. Use only facts in the brief; add no dates or numbers. Write three opening lines for a first cold email to the [JOB TITLE], each 25 words or fewer. Each line must name one specific fact from the brief and connect it to a problem that person owns. No greetings, no compliments, no "I noticed", no questions about their day, no exclamation points. Under each line, write the tag of the brief line it came from.

How do you catch a hallucination in sixty seconds?

Run our three-check, in this order:

1. Tag check. Every factual line carries [P1], [P2] or [P3]; only the line labeled "inference" goes untagged. No tag, or a tag whose page lacks the fact: delete it. Do not fix it.
2. Date check. The news line must carry the date you copied. If the AI wrote "recently" or moved the year, restore the page 3 date. Our rule: news older than 90 days is background, not a reason to write today.
3. Swap test. Swap in a competitor's name. If the brief still reads true, it is generic; cut until it could only be about this account.

Asking the model to audit itself is fine, but the grader is still the model. Your eyes are the check.

Citations are where verification starts, not where it ends. Per Perplexity's help page (read 2026-10-06), each answer "includes numbered citations linking to the original sources," and Claude's web-search page (same date) says it provides direct citations. A link can still point to a page that does not say what the brief says. If you search instead of pasting, add to Prompt 1: "For every fact, give the URL and quote the sentence." Then open every URL.

Our guidance: keep a tally. If more than 1 ask-mode brief in 10 needed a deletion, move that list to paste mode and keep it there.

Worked example: what does a five-minute brief look like for one account?

Invented example: the company, the roles and the numbers below are made up to show the routine. Brightmoor Payroll sells payroll and benefits software to companies with 50 to 500 employees. We sell onboarding software for new sales hires.

The pages, every detail invented. Page 1 (about): "Payroll and benefits for growing teams of 50 to 500," headquarters Columbus. Page 2 (careers): 11 open roles, five SDRs and one Head of Sales Development, plus five engineers; the SDR posting says "you will join a new mid-market team launching in Q1." Page 3 (press): "Brightmoor launches payroll for companies with 500 to 2,000 employees," dated September 18, 2026, first customers live in Q1. Prompt 1 returns a 92-word brief, and one line reads "Brightmoor raised a Series B earlier this year" with no tag.

The three-check. Tag check: no tag, and none of the three pages mentions funding. Deleted. That was the hallucination, caught in about 20 seconds, and the brief drops from 92 words to 84 (the line was 8 words). Date check: the AI wrote "recently launched"; we put back "September 18." Swap test: "a fast-growing company investing in its sales team" fits any competitor, so it shrinks to what only fits Brightmoor: a new mid-market team with six open sales seats. With our two hand-added lines, the brief lands at 73 words, tags not counted.

Verified brief (invented, 73 words):

The inference is labeled, so it passes the tag check untagged. The contact and question lines we add by hand after the check; the AI never writes them.

The first opener candidate ran 28 words. We cut "by Q1" and "at once" and got 24, under the 25 cap:

The math: 40 + 40 + 40 + 30 + 60 + 50 + 40 = 300 seconds, or 5:00. Six sales seats out of 11 open roles is 6 / 11 = 55%, just over half their hiring. Saved: brief, opener, three URLs, run date.

Where does the five-minute brief go, and what does it feed?

Save the brief, opener and three URLs in one field on the account record. A brief that lives in a chat window is lost by Friday.

The opener becomes the first line of the first email (structure in cold emails that work). Who you write to is a separate step: how to find the decision maker.

Now the volume math (our math). At 5 minutes each, 20 accounts a week is 100 minutes, about 1 hour 40 minutes. Over 4 weeks that is 80 accounts and 400 minutes, about 6.7 hours. Against our house plan of 504 prospects per rep per month, 80 accounts at two contacts each is 160 prospects (160 / 504 = 31.7%). The rest get segment messages. For the send side, the calculator estimates sends, replies and meetings.

When should you skip the AI entirely?

• When the company name is common. A search for "Summit" or "Meridian" can cite the wrong one, even with citations. Paste or skip.
• When you already have a reply thread. Their last email is a better page 3.
• When there is no news page and no filings. Write "no dated change found" and lead with the open role.

Start from a list you trust: how to build an ideal customer profile.