How Fokals intent scores are built, and how to read one

A score is only as useful as your understanding of what feeds it. This post follows the Fokals method for intent data from signal to row: what counts as a signal, what gives a signal its weight, how 90 days of signals become a number from 0 to 100, when that number is flagged as an intent surge and what evidence travels with it. With it you can take any row of Intent Scores, say what produced it and put it to work in a ranking, a model or a call.

A signal is a dated public action

The unit behind every score is a signal: something a company did in public, on a date, that points to a coming purchase or build. Each signal counts towards one or more of 69 topics, which sit in nine groups: advertising, marketing, commerce, bookings and events, sales and service, data and technology, website technology, business systems, and growth. Signals come from four kinds of evidence.

• Website changes. A company adopts or removes an advertising platform or a tool, migrates to another platform, verifies its domain with a software vendor, or adds a market, a language, a currency or an app. A new sales motion counts as well, such as a pricing page, a demo or a free trial, and so does a new programme, such as affiliate, subscription or wholesale.
• Job postings. A posting names a tool, with more weight when the tool is in the title. The labels on a posting add others: a paid-media role, a role that selects tools, an international role, an AI role, a new initiative, or a leadership hire that builds a function.
• Funding. A private capital raise reported in a regulatory filing, weighted by the amount raised.
• Announcements. A company announcement or regulatory disclosure that has an event type becomes a signal for the topics that event implies: funding counts towards new funding, expansion towards international expansion, and an acquisition or a leadership change towards building a new function.

Classification adds a fifth. The text of a company's website and of its postings is classified against the taxonomy, and a topic is assigned where the text shows the company adopting, buying, building or expanding in it. These topic signals are produced under a named, frozen version, like every label in the data.

Each of these is an act of the company itself. The score uses no browsing data, no publisher co-operative and no advertising bid stream, which is the difference set out in company signals vs third-party intent data.

What gives a signal its weight

Each signal in Company Signals carries its observation date, its kind, its source, the topics it counts towards, its weight and a line of detail. The weight states how strong the evidence is, and three principles describe it.

• A commitment outweighs an adjustment. A platform migration or a new advertising platform weighs more than a new language on the site.
• Specific outweighs general. A tool named in the title of a posting weighs more than one named in the body, and a paid tool adopted on a site weighs more than a free one.
• Size counts. A capital raise is weighted by the amount raised.

One safeguard keeps a single day from dominating. Signals of one kind from one observation of a site are capped, so a company that opens dozens of markets at once keeps a score in proportion.

Company Signals is also wider than the score. It delivers dated signals on the sales organisation, such as a founding sales hire, sales expansion into a new country and a move upmarket, as events to use in their own right.

From signals to a weekly score

A score is computed for each company, topic and closed week, and a week runs from Monday to Sunday in UTC.

1. Take the dated signals of the 90 days to the week's end.
2. Weight each one by recency, so a signal from the scored week counts for more than one from two months before.
3. Sum the weights for each topic.
4. Map the sum to a scale of 0 to 100. The scale climbs fast with the first strong signals and levels off as more evidence arrives.
5. Write the row once, after the week closes. It is never revised.

One adjustment reaches across topics: new funding, itself a topic of the growth group, raises the company's other topics as well.

Recency is easiest to follow with a case. Take Acme Robotics, an illustrative company, and three signals of equal weight that count towards the CRM topic.

The three are summed, and the sum is what the scale maps. The shape of the scale tells you two things about reading the result. A second strong signal moves a low score more than a tenth signal moves a high one, and a few points of difference near the top of the scale say little. Take the score from the row, which is the record of what was known when the week closed.

Recency also explains why a score falls. With no new signals, each signal counts for less week by week and leaves the window after 90 days, so a company moves down a ranking as its public activity on the topic goes quiet. A falling score with old evidence dates records a quiet period, and one fresh signal lifts it again.

The surge flag

The surge flag is set when a strong score is at least double the company's own average for that topic over the previous twelve weeks, or when it is a first strong week on the topic. The flag marks a change from the company's own normal, which a high score alone does not. The three cases below are illustrative.

The second row is the one to remember. A company that scores high on a topic every week is steadily active, and the flag is reserved for the newly active. Sort on the score when you want the busiest accounts and on the flag when you want the newly busy ones.

The evidence behind the number

Each row of Intent Scores carries the number of signals behind the score and the five strongest as evidence, each with its date, kind, source, weight and detail. Every signal, and not the strongest five alone, is delivered daily in Company Signals. The query below lists the signals that count towards the CRM topic for one company in the 90 days to a week's end, strongest first.

-- Postgres, with topics loaded as jsonb
select observed_at, source, kind, weight, detail
from company_signals
where company_id = :company_id
  and topics ? 'crm'
  and observed_at <= :week_end
  and observed_at >  :week_end - interval '90 days'
order by weight desc, observed_at desc;

Use the query to read a score. A signal is dated by the act itself: a posting carries its own date and a filing its filing date. The weekly row is the point-in-time record. It holds what was known when the week closed and it is never revised, which is what a backtest or a model needs.

How to read one

• One topic at a time. Each row is one company, one topic and one week. Choose the topics that describe what you sell or study, and rank within each.
• A missing row is a quiet week. A company in the index with no row for a topic was quiet on that topic in that week.
• Read the evidence before the number. Three signals of different kinds, from the website, the careers page and an announcement, are a different case from one large funding signal.
• Keep the version. Each score carries the version it was produced under. A change to topics, weights or the scoring formula is released under a new version, with at least 90 days' notice for a breaking change, in the same way as a label version.

The guide to prioritising accounts with intent scores joins the score to fit and turns it into a weekly ranking.

What a score gives you

A Fokals intent score is evidence-backed: every score carries the dated signals behind it, so a rep, an analyst or a model reviewer can open the source and check it. Three uses follow from that.

• A reason to call. The score says where a company's recent public activity on a topic is concentrated, and the evidence gives the rep a dated fact to open with: the role posted, the tool adopted, the raise reported.
• A feature for a model. Scores are produced under a named, frozen version, and weekly rows are written once after the week closes. A feature built on them keeps its meaning from one quarter to the next.
• A point-in-time series. Each row is what was known when its week closed, so a backtest reads the score exactly as it stood.

Calibrate the score on your own outcomes. For each band of scores, count the accounts that went on to open an opportunity, and set your cut-off where the bands separate. Keep the cut-off fixed for a quarter so that the bands stay comparable.

The intent dataset delivers Intent Scores, Company Signals and Company Funding, and the methodology is the reference for the method.

Frequently asked questions

The queries and code on this page are examples to adapt. Test them in your own environment before you rely on them.