How to research an account with AI before reaching out

A repeatable AI research routine that finds one specific, checkable problem in a company's public writing, with the prompts and verification checks that keep it honest.

Get started

Short answer

AI account research means pasting a company's own words into a model, meaning site pages, recent posts and an interview, then asking it for one gap between what the company says and what it does. The output you want is a single observation with a quote behind it, not a drafted message. You write the message, and you check the quote first.

What a useful observation looks like

Most account research starts by hunting for something nice to say. That's the wrong target. Complimenting a founder's growth or a product launch is easy to automate and easy to ignore, because it tells the prospect nothing about what they're dealing with this quarter.

What you want is a gap. A gap is a place where what the company says it wants and what its public footprint shows don't line up. A careers page hiring four SDRs while the demo form still asks for six fields. A CEO posting about moving upmarket while the pricing page lists one self-serve plan. Both facts need to be recent, or you're describing the company as it was two years ago.

One gap per message. Stack three observations into a first line and the prospect reads a report and answers none of them. Pick the one closest to the problem your product solves and keep the others for a follow-up.

Generic praise

  • I saw you are growing fast and have a great team.
  • Your product looks like a leader in the AI space.
  • I've been following your company and love the mission.

Specific observations

  • Your CEO posted three times about shifting toward self-serve, but your pricing page still hides the numbers behind a call.
  • You just shipped a Salesforce integration, which puts you in front of CRM admins instead of the marketing buyers you were writing for last year.
  • The careers page has four open SDR roles and the demo form still asks for six fields.

Collect the raw text before you open the AI

Hand a model a URL and ask what the company is struggling with, and you'll often get an answer built from what companies like that usually struggle with. Browsers miss pages, cut long posts short and summarize text that isn't on the page. You then build a message on a sentence nobody at the company ever wrote.

Paste the source text yourself. Three minutes of copying removes most of the ways this goes wrong, and it gives you the exact wording to quote later. Copy current pages and write down the date you pulled each one, because half of these pages change within a quarter.

These are the sources worth pulling, and what each one is good for.

Skip press releases and generic blog posts. They're written to sound finished and rarely contain the friction you're looking for. A job ad or a founder's comment under someone else's post will tell you more than a launch announcement.

Careers page

Source
Where the company is spending on people, and which teams are under pressure.
What it tells you
Job titles, team names, tools and requirements listed in the ads.

Founder or exec posts

Source
What leadership is telling the market and themselves right now.
What it tells you
The last five posts, plus any comments where they answer a question.

Company page posts

Source
What marketing thinks is worth announcing, and the language they use for it.
What it tells you
Post text and the date of each launch.

Pricing page

Source
Who the product is built for and how buying works today.
What it tells you
Plan names, what's gated behind a call, the steps to checkout.

Podcast or interview

Source
Longer answers about what's hard, usually said with less polish than on a marketing page.
What it tells you
Two or three paragraphs where the founder talks about process or bottlenecks.

The prompt that returns an observation you can use

Ask a model to write the outreach message and you get the shape every model produces: a compliment, a line of context and a claim about your product. The research is the part worth handing over. The sentence is yours.

Give the prompt a narrow job. Read the pasted text, return one gap and quote the sentence on each side of it. Tell it to answer only from the pasted material. Without that instruction, models fall back on general knowledge about the industry and the answer drifts away from the account.

This prompt works with any chat model. Replace the bracketed part with the problem you sell against.

If the model hands back something like "they're investing in growth", delete it and rerun with a named instruction: find the contradiction between the careers page and the CEO's last two posts. A vague answer usually means the pasted text was too thin.

Paste this job

Act as a B2B researcher. Below is text copied from [company]'s own pages, posts and interviews.

Find one gap: a place where two parts of this text point in different directions, or where a stated priority is not reflected in how the company actually operates. The gap should relate to [the problem my product solves].

Rules:
- Use only the pasted text. If it does not support a gap, reply "no gap found" and stop.
- Quote the exact sentence for each side of the gap.
- No adjectives about the company. No praise.
- Output: one sentence describing the gap, then the two quotes.
- Keep the whole answer under 60 words.

Check the quote before you type a single word

Models fill gaps. When the pasted text doesn't answer your question, a model will often answer anyway, borrowing from what it knows about companies of that size and stage. Send that and you've written to someone about a rebrand that never happened, or a funding round that was actually a customer win.

The fix is a quote you can find yourself. Ask for the exact sentence from the source, then search the pasted document for it. If it isn't there word for word, drop the observation and move to the next account. That's two minutes per account, and it means you never send a claim you can't defend.

Run these checks in order.

  1. 1

    Ask for the quote

    Reply to the model with: quote the exact sentence from the pasted text that supports each side. An answer with no quote is a guess.

  2. 2

    Search the source

    Paste the quote into the page, post or transcript you copied. If it doesn't appear word for word, the observation is out and you start the next account.

  3. 3

    Check the date

    Find when the post, job ad or page was published. If the change you're citing is more than a quarter old, phrase it as background or find a fresher signal.

Example

  1. You

    Saw two open AE roles on your careers page aimed at mid-market, and your CEO's post last week said the demo still goes to a founder. Are you planning to move that off the founders this quarter?

  2. Prospect

    That's the plan, it's just not built yet. What do you do?

Decide which accounts get the full routine

Not every account earns the same five minutes. Spend the full routine on accounts where something changed recently. A trigger gives your observation a date, and a dated observation is much harder to ignore.

For everything else, run the light version: one sentence from the LinkedIn page and a note to check back later. The prompt still runs, you just won't find a gap, and you'll know that before you write anything.

Keep a list as you go. Account, the two quotes, the date you pulled them, the sentence you used. Three weeks later when someone replies, you'll want to remember what you saw and whether it's still true.

  • A new leader in the function you sell to, hired in the last quarter
  • A product launch or pricing change that implies a different kind of buyer
  • A job post for a role that only exists when a company is fixing something, like a RevOps hire or a second support lead
  • A funding round with a stated plan for the money

Turn the observation into a short opening line

The observation is the hard part, so keep the sentence around it small. Name what you noticed, name the two facts, stop. Anything past that is usually you explaining why it matters, which the prospect can work out alone.

End with a question you'd actually want answered. On a connection request, one fact and one short question is enough; the pitch can wait until they accept, and often it should wait until they ask. Our LinkedIn buying signals checklist has more on which triggers are worth writing about.

Same account, two openers.

Weak opener

  • I love what your team is building and would love to connect.
  • Looks like you're scaling fast. Congratulations on the growth.
  • I help companies like yours do better outbound.

Strong opener

  • Four SDR roles open and a demo form with six fields. Is the form doing the qualifying for them?
  • Your CEO keeps writing about mid-market, but the pricing page still only shows the self-serve plan.
  • You shipped the Salesforce integration two weeks ago. Does the demo still go to a founder?

Doing this with Kinlead

The copying and quoting is the part Kinlead skips. The Steal Customers agent watches competitor company pages and, if you add them, founder or employee profiles. It reads the comments on their recent posts and turns commenters who look like buyers into leads. Each lead keeps the post URL, what the post was about and what the person commented, so the reason for writing is recorded before you write.

Leads are scored 0 to 100 against the customer profile you described, and anything below the fit line isn't kept. Connection notes are written under 200 characters and open with why that person became a lead. If you want to check the wording first, queued messages can be read and edited on the Actions page before they send, and any campaign can be paused.

Frequently asked questions

Related

Get started with Kinlead

Prefer markdown? How to research an account with AI before reaching out.md