The comms lead's stopping problem
A narrative about a supermarket's pricing practices breaks on regional radio at 06:40. By 09:00 it has been picked up by two trade newsletters. By 11:00 a hashtag is trending with 4,000 posts an hour. By 13:00 a national broadcaster runs it as the lead item. The comms lead is briefed at 14:00, ahead of a board meeting at 15:00, and treats the story as still forming — something that can be shaped, corrected, pre-empted. It cannot. It set at around 09:30, when trade coverage and social volume crossed a threshold past which retraction and correction notices no longer moved the story's shape, only its footnotes. The brief arrives four and a half hours after the decision it was meant to inform.
This is not a failure of speed alone. It is a failure of knowing when to stop watching and start acting — and that is exactly the problem search theory was built to solve.
Information has a price, and a rate
George Stigler's 1961 paper The Economics of Information asked why identical goods sell at different prices in the same market, and answered that buyers stop searching once the expected gain from one more look falls below its cost. John McCall's 1970 job-search model made this sequential: a searcher facing a stream of offers sets a reservation value and accepts the first offer that clears it. Diamond, Mortensen and Pissarides, working through the 1970s and 80s, showed that the reservation value is not fixed. It depends on the rate at which offers arrive. Slow the arrival rate and the threshold falls — you accept things you would once have rejected, because waiting for a better offer now costs more than it's worth. The three shared the Nobel in 2010 for this.
Applied to media monitoring, the "offer" is not a job or a price. It is a piece of evidence about where a narrative stands — a wire story, a broadcast segment, a spike in mentions, a correction notice from a source outlet. The comms lead's reservation value is the threshold of confidence at which they stop watching and issue a statement, brief a client, or recommend silence. That threshold should track the rate at which relevant evidence is arriving. A narrative accumulating three corroborating outlets an hour is a different object from one sitting untouched for three hours. Search theory says the correct stopping point depends on that rate, not on elapsed time or on a fixed evidence quota.
Three generations of watching
| Generation | What it observes | What happens to the arrival rate |
|---|---|---|
| Large Language Model | A frozen corpus of past coverage | Zero by construction — nothing new can arrive, so the optimal policy is to answer immediately from what is already there |
| Large World Model | A bounded monitoring window — a day's transcripts, a campaign's social pull | Positive while the window is open, but collapses to zero at the boundary; the threshold set at hour six is frozen even though the story kept moving at hour seven |
| Large Universe Model | Every stream still running — feeds, transcripts, social volume, correction notices — held as revisable belief with provenance | Positive, non-stationary, and measured continuously, so the threshold can move with it |
A monitoring tool built on a Large Language Model answers questions about coverage as it stood at training cutoff; it cannot tell you the hashtag is trending because it has never seen a hashtag trend. A tool built on a Large World Model can watch a live dashboard for the length of a shift, correctly detecting the story setting between 09:00 and 11:00 — but if the comms lead only checks it once, at the boundary of that shift, the read is only as fresh as the boundary. The claim behind the third generation is that arrival rate itself — how fast the story is moving, right now, across sources — should be a tracked quantity, not an assumption.
Two positions, honestly opposed
Position one. The comms lead's failure mode is structural, not personal: the brief happens after the story has set because nobody was pricing the arrival rate of corroborating coverage against the cost of waiting. If the monitoring layer had reported "mentions per hour crossed a stability threshold at 09:30, provenance: three independent outlets plus verified social volume," the reservation value would have been cleared four hours earlier than it was. On this view, continuous intake across feeds, transcripts, social streams and correction notices is not a nicety. It is the only way to compute the threshold correctly, because the threshold is a function of a rate, and a rate can only be observed by watching continuously.
Position two. Comms decisions run on deadlines set by other people — a board meeting, a market close, a print deadline for a rebuttal. Arrival rate is an academic nicety when the CEO wants an answer at 15:00 regardless of what the feeds are doing. Under a hard deadline, more monitoring does not change when you decide; it can only change what you know when the clock runs out, and a well-built snapshot an hour before the meeting may deliver 90% of the value of continuous tracking at a fraction of the tool cost and analyst hours.
Both are defensible. The first is right that the threshold — how much corroboration counts as "set" — depends on rate, and only continuous observation delivers rate. The second is right that the deadline is usually fixed externally and monitoring does not move it.
Search theory assumes the searcher knows the arrival rate and the offer distribution. In media monitoring neither is known in advance — you're inferring λ from noisy, multi-source counts in real time. Once λ itself has to be estimated, you're not in Stigler's clean optimal-stopping world any more, you're in a much harder partially observed control problem, and invoking Nobel-winning search theory to justify a monitoring budget is borrowing rigour it can't actually deliver.
This is correct, and it matters. With the arrival rate unknown, there is no closed-form reservation wage to compute; a comms team is doing Bayesian filtering on messy counts, not looking up a Gittins index. But the weaker claim survives the concession: whatever policy a team actually runs, its quality depends on how accurately it has estimated the rate at which the story is moving, and that estimate only improves with observed arrivals — more streams, longer history, better-attributed correction notices. Intractability of the exact optimum does not remove the need to observe the process that would define it.
The board meets at 15:00 whatever the feeds say. If the deadline is exogenous, arrival rate is irrelevant to when we stop watching — we stop when we're told to. A frozen snapshot taken at 14:45 does the job a live feed does, at a fraction of the monitoring spend.
Also correct, and also incomplete. The deadline fixes when you must decide. It does not fix what you should believe at that moment. McCall's finite-horizon extension shows the reservation value should decline as remaining time runs out — you accept weaker evidence near the deadline than you would with hours to spare, precisely because you expect fewer further arrivals before the clock stops. Getting that decline right still requires knowing how fast evidence has been arriving up to the deadline. A snapshot at 14:45 tells you what has already shown up; it gives no basis for weighing whether the silence since 14:30 means the story has stabilised or means the next three outlets are about to land at once, as they did with the supermarket story between 11:00 and 13:00.
Where this actually lands
There is a third objection worth taking seriously, and it cuts the other way from the first two: monitoring every stream continuously is not free. Analyst time, licensing for broadcast transcripts, the risk of contaminating a clean signal with unreliable social chatter — these are real costs, and Stigler's founding insight was that information is bought, not given. A comms team that monitors everything, always, may be spending more on search than the sharpened threshold is worth. This argues for bounded, chosen intake as good economics, not bad discipline.
But notice what that argument concedes. Deciding which streams are worth their cost — which outlets corroborate reliably, which social sources should be discounted, which correction notices actually retract rather than merely soften — is itself a search problem, and answering it requires observed arrival rates and observed reliability per source. The bounded, cost-disciplined monitoring plan a mature comms team runs is a conclusion reached from continuous measurement of stream quality over time. It is not an alternative to continuous intake; it is what continuous intake, done for a while, correctly recommends. The thesis narrows to this: the third position is terminal on the intake axis because it is the only one containing the arrival process at all — not because every comms team should watch every stream at every moment.