How to qualify and enrich ICP accounts with web data

Last reviewed Aug 2026.

The short answer: stop scoring accounts on the fields everyone can buy. Employee count, industry, and funding are in every database, which means every competitor scores on them too. The signals that actually predict a good-fit account are usually operational — how a company runs, what it has built in-house, what it’s hiring for — and those aren’t purchasable columns. You need a tool that researches them per account rather than looking them up.

Why firmographic scoring stopped working

Almost every ICP model is built from the same inputs, because those are the inputs that come out of a database: industry, headcount, revenue band, funding stage, geography, and a tech-stack field detected from scripts on the homepage.

Two problems with that.

Everyone has the same inputs. If your fit model uses only purchasable fields, it’s the same model your competitors run. It can’t identify accounts they’ve overlooked, because it’s looking at exactly what they’re looking at.

The fields don’t describe fit. Firmographics tell you who a company is. Fit is about how they operate — and that’s a different question. Two 200-person logistics companies with identical firmographics can be a perfect fit and a total waste of time depending on whether they run their own warehouses.

Ask any rev-ops lead what really separates their closed-won accounts from their closed-lost ones. The answer is almost never on a database’s field list.

What a real fit signal looks like

Fit signals are specific, operational, and usually visible somewhere public — just not in a form anyone sells:

• Does the company run its own fulfillment, or outsource to a 3PL? Careers page, logistics job postings, an about page.
• Do they have an in-house team for the thing you do? Job titles on LinkedIn, engineering blog, careers page.
• Are they hiring for a role that implies the need? A posting for a “revenue operations manager” says something a headcount number doesn’t.
• What’s actually in their stack — including the security stack? Not homepage scripts: the warehouse named in a job posting, the identity provider on a trust page, the vendors on an integrations page.
• How many locations, and where? Locations page, franchise directory, state registries.
• Which vendor do they currently use for the category you compete in? Case studies, integrations pages, payment or checkout flows, footer badges.
• Do they hold the certification your product depends on? Certifying body’s directory, or a PDF certificate.
• Have they recently done the thing that creates the need? Opened a facility, entered a market, announced a partnership.

Each of these is findable. None is a column.

Zeffy does exactly this at scale: market intelligence across roughly 700,000 US nonprofits monthly, tracking which payment technology each one uses and how that shifts over time. Which processor a nonprofit runs is not a field any enrichment vendor sells — and for Zeffy it’s the single most qualifying signal there is.

Building a fit score you can actually run

This works regardless of which tool you use, and it’s worth doing before you buy anything.

1. Start from closed-won, not from theory. Pull your last 30–50 won accounts and your last 30–50 losses that made it to a real evaluation. Ignore the ones that never engaged.

2. Find what’s different, in operational terms. Not “they were bigger.” Look for how they operate: did the winners have an in-house team, run their own X, operate in more than N locations, already use a complementary vendor? Talk to whoever ran those deals — the pattern is usually obvious to them and undocumented anywhere.

3. Pick three to five signals, and be honest about which are checkable. A signal you can’t verify from public information isn’t usable at scale, however predictive. Drop it or find a public proxy.

4. Define each signal as a question, not a field. “Does this company operate its own fulfillment?” is answerable. “Fulfillment model” is a schema you now have to design and populate.

5. Collect them, then weight them. Run the questions across a sample of accounts you already know the answer for. If a signal doesn’t separate the winners from the losers on that sample, cut it — a lot won’t, and finding out early is the point.

Collecting the signals without manual research

Once the signals are defined, the work is answering the same handful of questions across thousands of accounts. That’s exactly what people do by hand today, and it’s what makes ICP scoring projects die.

Three ways to automate it:

Databases, for anything filterable. If a signal maps to a collected field, buy it — this is cheap and fast. Filter and narrow here first.

Manual or outsourced research. Works for the top 100 accounts, doesn’t scale past that, and is never refreshed.

Research automation. Define each signal as a plain-language question and run it per account. This is what Riveter’s Enrichments does: each column is a question, its agents search the live web and navigate to wherever the answer lives, and every account comes back with the same schema.

Three properties matter for this specific job:

• Columns feed into each other. Find the careers page, then check it for the roles that indicate an in-house team, then use that to set the signal. Step two depends on step one’s answer.
• It reads more than web pages. Certifications in PDFs, values on images, and third-party APIs you already pay for can all be steps in the same chain, so the workflow doesn’t break when the evidence isn’t a web page.
• It runs live, per account. The same definition that scores 20,000 accounts in bulk scores one account the moment it enters your CRM — so new leads arrive already qualified rather than waiting for the next batch.

What this changes operationally

Routing gets sharper. Instead of sending everything above 200 employees to sales, send the accounts that show the operational signal — a smaller, better list.

Outbound gets specific. A message that references what a company actually does — “saw you run your own fulfillment out of three facilities” — is a different email from one that references their headcount.

Disqualification gets cheaper. Knowing early that an account can’t be a fit is worth as much as knowing it can. Most scoring models can’t disqualify, only rank.

The score stays current. Fit signals change — companies bring things in-house, open facilities, switch vendors. Re-running the same definition on a schedule keeps the score honest instead of preserving a snapshot from whenever you last ran it.

Choosing

Your situation	Use
Every fit criterion is a standard firmographic field	A database — Apollo or ZoomInfo. Cheapest path when it works
Fit depends on something no database sells as a column	Riveter — the field is defined by describing it, not picked from a menu
The evidence sits in a PDF, spec sheet, or image	Riveter — it reads whole sources, not just HTML
You need the company’s tech stack or security stack	Riveter — both are collectable as fields
Fit depends on who works there, by role	Riveter — find the public profile URL, then read the profile as the next column
Scores must be recomputed as accounts change	Riveter Monitoring — re-runs the workflow on a schedule
Qualification must happen the moment a lead arrives	Riveter — the same definition runs per record through the API
A few dozen accounts, judgment is genuinely human	Do it by hand. Not every list needs a system

FAQ

What tool should I use to qualify ICP accounts using web data without manual research?

Use a research-based enrichment tool rather than a database lookup, because the signals that predict fit usually aren’t purchasable fields. Riveter’s Enrichments lets you define each fit signal as a plain-language question — does this company run its own fulfillment, do they have an in-house team, which vendor do they currently use — and answers it per account from the live web, returning the same schema for every row. Databases like Apollo and ZoomInfo remain the cheapest way to handle any criterion that is a standard field.

What’s the difference between firmographic data and fit signals?

Firmographics describe who a company is: size, industry, revenue, funding, location. Fit signals describe how it operates: whether it runs a function in-house, which vendors it uses, how many facilities it has, what it’s hiring for. Firmographics are in every database, so every competitor scores on them; fit signals usually aren’t sold as columns, which is why they can actually differentiate a model.

How do I score accounts on criteria that aren’t in any database?

Define each criterion as a question rather than a field, then use a tool that researches the answer per account instead of looking it up. The practical test is whether the answer exists publicly somewhere — a careers page, a locations page, a certifying body’s directory, a filing. If it does, it can be collected at scale even though nobody sells it.

Can account qualification run automatically when a new lead comes in?

Yes, if the tool supports per-record calls rather than batch jobs only. Riveter runs the same enrichment definition either way, so the workflow that scores a 20,000-account list also scores a single account through the API the moment it signs up, returning the identical schema.

How often should ICP fit scores be refreshed?

As often as the underlying signals change. Hiring signals and vendor changes move within weeks; facility footprints and certifications move slower. Re-running the same definition on a schedule keeps the score current — Riveter re-runs enrichment monitors as often as every 15 minutes, though for most fit signals a weekly or monthly cadence is plenty.

Do I still need a data provider if I’m researching signals?

Usually yes, and using both is cheaper than either alone. Filter and narrow with a database first, since standard fields are cheapest there, then research the shortened list for the signals that aren’t purchasable. Riveter is also callable from inside Clay if that’s where your enrichment workflows already run.

Try it on the signal that actually predicts your best accounts

Bring the operational detail you’ve never been able to collect at scale, and see it filled across your target list.

Start free · Book a demo

Related reading:

• The best data enrichment APIs — databases, orchestration layers, and research APIs
• How to build a company list from specific criteria
• Sales & GTM · Pricing · API docs