What you'll be able to do after this guide: teach an agent to label every finding by how sure it is, so it never presents a guess as a fact in a CRM record or a message to a prospect.
Prospecting agents are confident writers. Without a rule, "this company probably uses Competitor B" becomes "I see you use Competitor B" in a message, and a prospect who doesn't use it stops reading. Grading every finding protects your credibility and makes research checkable.
| Tier | Meaning | Examples | How it can be used |
|---|---|---|---|
| Tier 1: Confirmed | Stated by the company itself or a primary source | Their own job advert naming the product; a branded login page linked from their site; a vendor case study naming them; a named accreditation on their site | Safe to cite in a message |
| Tier 2: Strong inference | Good evidence, not first-hand | An enrichment tool's technology tag; technical fingerprints in their website's code (tell-tale file names that reveal which product built a page); a matching reviewer on a review site | Corroborate before stating as fact; otherwise ask as a question |
| Tier 3: Circumstantial | Plausible, not evidenced | "Companies in this sector usually..."; a generic job title | Never cited, and never enough on its own to add a record |
Ask the agent to tag every signal in its notes with its tier, for example "Competitor B login page at a branded subdomain (an address like learn.brightside.example) linked from their footer (Tier 1)".
A Fernhill agent researching Brightside Ltd finds:
An enrichment tool lists Competitor B in their technology. Tier 2.
Their careers page has an advert for a "Competitor B administrator". Tier 1.
They're in a sector where Competitor B is popular. Tier 3.
Finding 2 can go in a message. Finding 1 supports it. Finding 3 is noise and stays out of the record.
These came from real mistakes, and each is now a standing rule.
Look at the real page before concluding "nothing there". An agent once dismissed several strong prospects because search results and the companies' own marketing suggested a polished setup. Opening the actual pages showed sites assembled from several add-ons and an old website system repurposed as a login portal: exactly the problem Fernhill solves. Negative conclusions need more evidence than positive ones.
Branding can be removed; fingerprints can't. White-labelled products (rebranded as the customer's own) hide logos but keep their URL patterns, file paths and stock wording. List the tell-tale fingerprints for the tools you care about in the spec, so the agent doesn't rely on visible branding.
Test against a known negative. If an agent checks whether a subdomain exists, it must also check a nonsense subdomain. Some hosts answer every request, so "the address works" proves nothing. If the real address gives the same response as the nonsense one, the company isn't a customer of that product. Only a page showing the product's real login counts as evidence.
Enrichment data is a lead, not a fact. Technology tags go stale when companies rebrand, and large groups' tools get attached to their small subsidiaries. Never use an enrichment figure as "evidenced revenue"; use the company register or published accounts, and say when a figure is an estimate.
Know where your data lies by sector. For recruitment and staffing firms, profile-based headcounts include placed temporary workers and can be several times the real number, and their job adverts are often for clients. Encode exceptions like this by name.
Search results have limits. A site search returns a limited number of results with no paging. An empty result is inconclusive, not a "no". Where it matters, have the agent open every link in the site's navigation and footer as well.
Confirm identity before enriching. Find the registered company name from the site's own legal pages (terms, privacy notice) before searching the company register. Guessing the legal name from the domain once returned a real but unrelated company.
Check the company is alive. A live website and recent social posts don't mean a company is trading. Check its status on the official register as part of the same visit.
Machine-read numbers must add up. If an agent reads figures from scanned accounts, make it check them against other figures on the same page before writing them. A misread digit once turned a five-figure turnover into a six-figure one.
Below Tier 2, flag it; don't guess it.
Never guess an email address, a domain or a person's identity. If it can't be confirmed, leave it blank and say so.
Never create a record from Tier 3 evidence alone.
In messages, turn unconfirmed signals into questions. "Are you using Competitor B at the moment?" is honest. "It looks like you're on Competitor B" is still a claim.
Next: move on to Build the Sourcing Agent, starting with Anatomy of a Sourcing Run.
