Your Competitors Leak Their Strategy. Most Teams Don't Read It.

Most competitive intelligence reads like a news wire. "Acme launched a new dashboard." "RivalCo hired a VP of Sales." "CompetitorX raised their enterprise tier 15%." Each observation isolated, stripped of context, delivered weeks after anyone paying attention had already noticed.

The value is not in the signals. It is in the overlap. A hiring spike in healthcare compliance roles. A changelog entry for HIPAA audit logging. A pricing page that just grew a "Regulated Industries" toggle. Read alone, three news items. Read together, a roadmap.

That triangulation says someone is making a deliberate bet on healthcare verticals, probably timed to a Q3 compliance certification announcement. Caught six weeks before the press release, you have runway: accelerate your own healthcare roadmap, reposition before their sales team has new collateral, lock down the at-risk accounts in your pipeline. Caught the day of the announcement, you are reacting.

Job postings leak the most. Before a competitor announces a product, a region, or a segment, they hire for it. Rate of change matters more than absolute count^[3] — a company that jumps from 3 to 12 engineering postings in a single month is signaling something about the next 6-12 months. Not a guarantee. Hiring also reflects attrition or reorg. But the rate-of-change anomaly is where the read starts.

Departmental breakdown sharpens it. An engineering and product spike points to a build phase or platform overhaul. A sales cluster in one geography points to territory expansion. A marketing burst weighted toward demand generation points to a category-creation push. Each pattern is a hypothesis, not a verdict. Combine signals before acting. One channel is noise.

The system that produces summaries and the system that produces inferences differ by one design choice — the prompt. Most teams ask "what happened?" The right question is "what does this combination of events imply about where this company is heading?" The first generates a digest. The second generates a brief.

An inference prompt that holds up under real signals does three things. It groups signals by competitor and explicitly demands cross-channel convergence. It refuses to output an inference that does not cite a minimum number of supporting signals from different channels. It ties confidence to the diversity and strength of the evidence, not the volume — ten analyst quotes do not equal one hiring spike plus one pricing change.

The first version of this system at one of our portfolio companies weighted all five channels equally. Analyst quotes — largely aspirational, often ghost-written — carried the same weight as job postings. Inference accuracy at medium confidence sat near 40%, well below the 55-65% range we hit after dropping analyst quotes to a 0.5x multiplier and lifting job postings to 1.5x. The lesson is structural: the channels are not equally predictive, and your initial weights are almost certainly wrong. Calibrate against your own prediction log or accept that the radar is decoration.

Every intelligence program is tempted to measure output volume — briefs published, signals collected, competitors tracked. These metrics tell you the machine is running. They tell you nothing about whether the radar changes a decision.

Three metrics are load-bearing. Prediction accuracy over 90 days: what percent of medium-and-high confidence inferences proved correct against actual outcomes? Practitioner reports cluster around 55-65% accuracy at medium confidence and 70-80% at high confidence^[1]. Below those bands, the signal collection or the inference prompt needs recalibration — usually the prompt.

Time-to-action: when a high-confidence inference is published, how many days pass before a stakeholder takes a measurable action — adjusts a roadmap, modifies positioning, reaches into an at-risk account? Briefs sitting unread for two weeks are a distribution problem, not an intelligence problem. Fix the format and the recipient list before touching the model.

Competitive win-rate delta: across a rolling quarter, compare win rates on deals where the team had advance intelligence from the radar against deals where they did not. A working program produces a measurable improvement — practitioners cite 8-15 percentage points — but the number varies with deal complexity, team size, and how aggressively the inferences are operationalized. Win-rate delta is the metric that ends the budget conversation.

The fastest working radar is set up in a week. Pick the two competitors your sales team encounters most often. Map their career page, changelog, and pricing page. Set up weekly diff monitoring on each URL — Visualping, Changeflow, or a cron-driven headless browser snapshot all work and the choice is irrelevant.

Write one inference prompt using the template in this article. Feed it the first week's signals. The output will be flawed. It will also be more useful than a Slack channel full of "hey, did you see CompetitorX launched a new feature?" messages.

Four weeks of data is enough to set meaningful anomaly thresholds. Eight weeks is enough to see prediction accuracy stabilize. Twelve weeks is enough that the team stops asking why the radar exists.

The sustained edge in competitive markets is not better products. It is shorter latency between the competitor's move and your response. Six weeks of lead time, used, beats six months of strategic planning that lands the day after the press release.