Lead Research Pipeline

How it works

The pipeline does the research a sales team would do by hand: find companies, find their websites, read them, decide who is worth contacting and why. This page explains each step and how the LLM is used.

From the whole region to a short list

The numbers from the run behind this demo.
  1. 120,104
    1. Active companies
    NE, SR, DH, TS and DL postcodes
  2. 29,244
    2. Small, trading companies
    after accounts and industry filters
  3. 200
    3. Sampled for this demo
    checked against the live register
  4. 69
    4. Possible website found
    via web search, then checked
  5. 55
    5. Website verified
    the LLM agrees it is theirs
  6. 35
    6. New leads
    20 were already in the CRM

Only 28% of sampled companies have a website that is verifiably theirs. Property companies almost never do (1 of 34), which is why the demo hides them by default.

The six steps

  1. Step 1
    Find companies

    Starts from the official register of UK companies, not a bought list.

    Companies House data is filtered by postcode area, company size (the accounts they file) and industry codes, dropping dormant and holding companies. Each sampled company is re-checked against the live register.

  2. Step 2
    Find their websites

    Companies House doesn't record websites, so the pipeline searches for each one.

    A web search for the company name and town. Directory and social media sites are removed, including any domain that turns up for several unrelated companies. Each remaining site is checked for the company's registered number and postcode.

  3. Step 3
    Read the website

    Fetches the homepage and about page, the way a researcher would.

    Respects each site's robots.txt, waits between requests and identifies itself. Email addresses and phone numbers are stripped before anything is stored.

  4. Step 4
    Profile each company with an LLM

    Claude (claude-haiku-5-5) reads the page text and fills in a fixed form.

    It decides whether the website really belongs to the company, then records what they sell, who they sell to, whether that is B2B or B2C, and any size or growth facts stated on the site. It is told to use only what the pages say and to leave fields empty rather than guess.

  5. Step 5
    Score fit and draft an opener

    Each profile is scored from 0 to 100 against your ideal customer.

    The score comes with a one-sentence reason citing the facts behind it. Leads scoring 50 or more get a two-sentence opening line for a cold email that mentions something specific about the company.

  6. Step 6
    Clean for your CRM

    Matches new leads against what you already have.

    Your CRM export is standardised (names, websites, postcodes, towns, dates) and de-duplicated, then each lead is matched by company number, website or name and postcode. New leads are exported as a HubSpot-ready file.

How the LLM is used

The model never browses or searches on its own. It only reads text the pipeline has already collected, and it must answer in a fixed format that is checked before anything is saved.
Call 1: company profile

One call per company, reading its homepage and about page.

Website belongs to them
yes, no or unsure, with a reason
What they sell
one plain sentence, plus up to five products or services
Who they sell to
one sentence on their customers
Market
B2B, B2C, both or unclear
Sector
a short label, e.g. roofing contractor
Size and growth signals
countable facts only, e.g. 3 branches, established 1891, hiring
Call 2: fit score

One call per company and customer type, reading the short profile rather than the whole website, which keeps it cheap.

80 to 100Strongclearly matches the ideal customer, with evidence
50 to 79Possibleplausibly a customer, some evidence missing
20 to 49Weaka partial match
0 to 19Not a fitnot a sensible prospect, or too little to judge

Plus a one-sentence reason, and an opener for scores of 50 or more.

Rules the model works to
  • Use only what the website says; leave a field empty rather than guess.
  • A different address on the site is not on its own evidence against a match, because many small firms register at their accountant's office.
  • Never name individual people; refer to them by role. Output is also checked automatically for names after every run.
  • Don't lower a fit score just because a website doesn't state headcount, when the kind of work clearly implies staff.
  • Openers must only mention facts from the profile: no praise, awards or guesses about how successful a company is.
  • Every answer is validated against a fixed schema before it is stored, and every call is logged with its token count and cost.
$0.54
Per 1,000 companies profiled
$0.28
Per 1,000 fit scores
$0.29
LLM cost of this whole project
including discarded prompt versions

What I can build for you

The same pipeline, pointed at your market: your region, industries and company sizes, scored against your ideal customer in your own words.

  • A list of qualified leads with the reasoning for each score
  • Matching against your existing CRM, so you only get companies you don't already have
  • A HubSpot-ready import file, or a CSV for any other CRM
  • An accuracy check against hand-labelled examples, so you know how far to trust it

It uses company-level public data only, with no personal names, emails or phone numbers, which keeps it simple under UK GDPR.

Get in touchView the source code