Go-to-market tools sell signals. A company visited your pricing page. A company is "researching" your category. A company is hiring. A contact changed jobs. Each signal arrives with an implied promise: this account is ready to buy, so act now.
Some of those signals do predict buying. Many predict nothing, and a few predict the opposite. A student writing an assignment visits the pricing page. A competitor attends the webinar. A company hiring fast may be too busy to change payroll software this quarter. CloudNala's view is that a signal is a hypothesis about buyer behaviour, not evidence of intent. It earns a place in your score and your plays only after your own outcome data shows that it predicts something, and it stays there only as long as it keeps doing so.
That turns signal management from a purchasing decision into a small, continuous research programme. It is cheaper than it sounds, and it is the only way to know which of your signals are working.
Lowveld Ledger is fictional, and its systems and figures are illustrative. Lowveld Ledger is the Mbombela accounting and payroll software company from earlier parts. Its revenue data now has written lifecycle definitions and a single system of record per field (Part 3). Its sales team has a long list of things it believes indicate a good prospect, and no way to tell which are true.
The list of things everyone believes
When the GTM engineer asked the commercial team what made a prospect worth calling, they produced a list in fifteen minutes. Webinar attendance. Pricing-page visits. A referral from an accountant. Opening three emails in a row. A trial sign-up with a company email address. Company size between 20 and 80 employees. A job ad for a payroll administrator. The SARS employer reconciliation season.
Every item was plausible. Every item had a story behind it about a deal that closed. None had been checked against the outcomes of the hundreds of deals that did not close.
This is the normal state of most commercial teams, and it is not a criticism. Experienced sellers develop good instincts. But instincts are formed from memorable wins, and they are hard to update when the market changes. A signal catalogue makes them testable.
Four kinds of signal
It helps to separate signals by what they claim to tell you, because each kind fails differently.
Fit signals say whether a company could be a good customer at all: size, industry, province, payroll frequency, current accounting tool. They change slowly. They fail when the target profile they encode is out of date.
Intent signals say a company may be looking for something like your product: pricing-page visits, comparison-page visits, third-party "research" data. They fail through noise. Many visitors are not buyers, and third-party intent data is an inference built on someone else's model that you cannot inspect.
Engagement signals say a person is paying attention to you: webinar attendance, email replies, trial activity. They fail when attention is mistaken for buying authority. The bookkeeper who attends every webinar may not choose the software.
Timing signals say now may be a better moment than later: a trial reaching a certain day, a job ad for a finance role, a compliance deadline. For a payroll product in South Africa, the SARS employer reconciliation periods are an obvious candidate. In 2026, SARS opened the interim EMP501 period on 21 September and closed it on 31 October, after the annual period that ran from 1 April to 31 May (checked on 31 October 2026 against SARS announcements). Timing signals fail when everyone in your market uses the same calendar and the buyer is drowning in identical messages.
Every signal gets a catalogue entry
The GTM engineer replaced the list with a signal catalogue. Each signal became an entry with the same fields.
Name and type. Plain language, and one of the four types above.
Source. The system of record from Part 3. A signal from a system that is not the source of truth for that field is not admitted.
Hypothesis. One sentence that could be false. "Trials that run a test payroll by day 10 convert to paid more often than trials that do not." Not "payroll activity is important".
Test. How the hypothesis will be checked and over what period. Usually a comparison of outcomes between accounts with and without the signal.
Decay. How long the signal stays meaningful. A pricing-page visit is stale in days. A company-size fit lasts until the company grows.
Owner. The person who reviews the signal's results. At Lowveld Ledger, the GTM engineer owns the catalogue and reviews it with the head of sales.
Status. Testing, proven or retired.
The catalogue started with fourteen entries, all marked testing.
How to test a signal with your own data
Testing a signal does not require a data-science team. It requires the outcome data from Part 3 and some honesty about sample sizes.
The simplest method is a backward look. Take every account that entered the trial or reached MQL in a past period. Split them into those that showed the signal and those that did not. Compare what happened next: SAL, opportunity, customer, and for customers, whether they were still paying after six months. If the difference is large and the groups are not tiny, the signal is promising.
The stronger method is a forward test. Use the signal for some accounts and deliberately not for a randomly chosen comparison group, then compare outcomes after a set period. This is what Part 6 calls a holdout, and it protects you from the most common mistake in signal analysis: confusing a signal that predicts buying with a signal that sellers simply act on more.
Lowveld Ledger's backward look produced three useful surprises.
First, a test payroll run by day 10 of the trial was by far the strongest signal in the catalogue, and nobody in sales had been able to see it before the product data was synced.
Second, opening three emails in a row predicted almost nothing. Some email clients and security tools open messages automatically, so the signal was partly measuring software, not people.
Third, webinar attendance predicted conversion only when two or more people from the same account attended. One attendee, on its own, was close to noise. That finding only became visible after Part 3 separated accounts from contacts.
These are illustrative results for a fictional company. The point is the method: your own data, not a vendor's claim, decides which signals you use.
Proven is not permanent
A signal that worked last year may not work this year. Buyers change, products change, and when competitors start using the same signal, its value as a differentiator falls.
Lowveld Ledger re-tests every proven signal each quarter. If a signal no longer separates good outcomes from bad, it goes back to testing or is retired. Retired signals stay in the catalogue with a one-line reason, so that nobody proposes the email-opens signal again in a year without knowing it was tried.
The catalogue also has a rule for adding signals: any new signal, including any new data feed a vendor offers, enters as testing with a hypothesis and a test. It does not affect scores or routing until it is proven. That rule alone saves money, because it turns every "you should buy our intent data" conversation into "give us a trial period and we will measure it".
Signals and the people behind them
Some signals involve personal information about named people: who visited a page, who changed jobs, who attended a webinar. How you collect and use that information is governed by POPIA and by your own privacy controls, and some data you can buy should not be used at all. Part 10 of this series covers the governance, and the CloudNala guide on which data AI can touch is a useful plain-language starting point. For now, one design rule: every signal's catalogue entry records whether it uses personal information and on what basis, and a signal that cannot answer that question is not admitted.
What changed at Lowveld Ledger
Of the original fourteen signals, four were proven within the first quarter, six stayed in testing because the samples were too small, and four were retired. The four proven signals became the inputs to the lead score in Part 5. The trial payroll signal, and its mirror image (a test payroll started but not finished), became the basis for the company's first specified play in Part 6.
The commercial team did not lose its instincts. It gained a place to write them down as hypotheses and a way to find out which ones were right.
Start this week
- Ask your team to list every signal they believe in, and write each one as a hypothesis that could be false.
- Run one backward look on your strongest-believed signal using the last six months of outcomes. Compare accounts with and without it.
- Add a rule that no new signal or data feed affects scoring or routing until it has been tested against your outcomes.
Next: Part 5, The lead score must explain itself. All parts are on the series page.
How CloudNala can help
CloudNala helps commercial teams turn beliefs about buyers into a tested signal catalogue. We write the hypotheses with your team, run backward looks on your own outcome data, design holdout tests for the signals that matter, and set up the quarterly review that keeps the catalogue honest. We do not sell data feeds, so we have no interest in which signals you keep.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book a Go-to-Market Systems Review or write to us at consult@cloudnala.co.za