Everybody in this category sells "signals" and almost nobody defines one. The word has been stretched to cover a page visit, a topic surge inferred from anonymised browsing, a job posting, a funding round, a competitor's outage, and a person typing "does anyone know a good X" into Reddit at midnight. Those are not the same object and they do not support the same action.
This post is the definition we actually build against. It is not a neutral industry standard (there isn't one), but it is testable, which is more than the category usually offers.
The test
A buying signal is a piece of public evidence, attributable to a specific party, dated, that a problem you can solve is currently costing somebody enough that they are doing something about it.
Five clauses, and each one is load-bearing. Drop any of them and you have something else.
- Public evidence. Something a person wrote, that you can read and quote. If you cannot show it to a colleague, it is an inference, not evidence.
- Attributable. To a named company, a named business, or at minimum an account you can reply to. "An account in the financial services vertical" is not attributable.
- Dated. Intent decays. A complaint from fourteen months ago is history.
- A problem you can solve. Not a problem. Your problem. This is the clause that makes a signal specific to a seller rather than a property of the post.
- Doing something about it. The difference between someone who is annoyed and someone who is shopping.
Run that test on the things vendors call signals and most fail on one clause, predictably:
| Called a signal | Fails on | What it actually is |
|---|---|---|
| Topic surge from an intent co-op | Public evidence, attributable | A prior, useful for prioritisation |
| Website visit | Public evidence | First-party analytics |
| Mention of your brand | A problem you can solve | Brand monitoring |
| Funding round | Doing something about it | A budget event, one hop from a signal |
| Job posting | Nothing, it passes | A signal, if the role implies your problem |
| "Our current vendor keeps going down" | Nothing, it passes | The cleanest signal there is |
The last two are worth dwelling on, because they are the two ends of the useful range.
The four parts of a signal that can be acted on
A signal that survives the test still has to carry enough structure to be worth a person's time. In practice that is four things.
1. The evidence itself. The full post or review text, not a search snippet. This matters more than it sounds: a classifier judging a 160-character search preview is the single largest cause of a wrong label in this kind of system. The snippet contains the keyword (that is why it was returned), so it looks relevant by construction, and the context that would have disqualified it is in the part that was cut off. Where a platform allows full text, take it. Where it does not, label the source honestly as snippet or best-effort so a human knows the confidence is lower.
2. A classification, in a vocabulary that fits the situation. "Relevant" is not a decision. What a salesperson needs is which kind of relevant. The vocabulary has to differ by what you are watching, because the same word means different things in different contexts:
| Watching | The strongest label | The rest of the vocabulary |
|---|---|---|
| Your market | high_intent | evaluating, relevant, competitor, content |
| A named rival | switching | frustrated, comparing, praise |
| Local businesses | pain_confirmed | at_risk, healthy |
praise and healthy are in there deliberately. A rival's happy customer is not a lead, but it is the objection list you will meet on every call, and discarding it loses information you paid to retrieve.
3. A score, and the reasoning behind it. A number alone is a black box. The useful artefact is the number plus the sentence explaining what in the post produced it, because that sentence is what a rep reads before deciding whether to open the thread. It is also what makes the system auditable: when a score looks wrong, you can see whether the model misread the post or read it correctly and you disagree about the threshold.
4. A date and a window. Different evidence stays fresh for different lengths of time. A conversation about a problem is a 30-day artefact. A review of a local business is more like 90 days, because the underlying operational problem it describes moves more slowly than a thread does. Fixing one window for everything is how you end up with a feed full of stale complaints or one that throws away signals that were still good.
The one-hop rule
The hardest part of this is not detection. It is knowing when to stop inferring.
A job ad for three support agents is not a person saying they need help-desk software. It is a company saying it has a support-volume problem and has budgeted for it. Getting from the first to the second is one inferential hop, and it is a legitimate one: the evidence for it is in the posting, which names the volume, the tooling, or the pain.
Two hops is where it stops being evidence. "They are hiring support agents, therefore they are growing, therefore they are probably replacing their CRM" is a story you told yourself. Nothing in the posting says anything about a CRM.
So: one hop, and the evidence for the hop must be in the source document. If you have to reach outside the post to justify the inference, you have manufactured a lead rather than found one. This rule is what separates a derived signal from an invented one, and it is the discipline that makes the difference between a list a rep trusts after two weeks and one they stop opening. The job postings as buying signals post works through the hiring case in full.
Why most candidates are not signals, and why that is the product
Here is the part that is counterintuitive to anyone who has bought a monitoring tool: the value is overwhelmingly in what gets thrown away.
A run that retrieves a few hundred candidates and keeps a handful has not failed. It has done the job. The reason mention-metered tools feel productive and then quietly stop getting opened is that they keep everything, and a human becomes the filter. Humans are expensive filters and they burn out in about three weeks.
What makes a discard defensible is recording why. Every candidate Openpulse drops keeps its reason, from a fixed set of ten: wrong shape of page, matched an exclude term, no lexical overlap, blocked domain, outside the window, duplicate, put out of contention by the reranker, a review above your star ceiling, scored below the bar, or classified as noise. When a listener goes quiet, that record is the difference between "the searches returned nothing" and "the searches returned plenty and the gates threw it all away": two problems with opposite fixes. The false positive taxonomy post takes all ten in turn.
What this means when you are buying
You can evaluate any tool in this category by asking what it does with the four parts above.
- Does it show me the full text, or a snippet it judged from?
- Does it give me a classification I can act on, or a relevance percentage?
- Can I see why it scored something the way it did, and why it dropped the rest?
- Does it know that different evidence has different shelf lives?
A tool that cannot answer those is selling you retrieval and calling it intelligence. That may be exactly what you want (retrieval is a real product and some of the tools that do it are excellent), but you should know which one you are buying. The best social listening tools roundup sorts the market by that question.
See it on your own market
Paste your website, review the plan it proposes, and read what comes back tomorrow morning.
Questions about anything here? Email support@openpulse.cloud.