How the index is built, and why you should be skeptical of it

The idea is usually traced to Cold War–era Washington, where journalists noticed that late-night food deliveries to government buildings sometimes preceded public announcements of a crisis. The most repeated example is the period before the 1991 Gulf War. The observation was folklore among reporters long before it was a dataset.

It resurfaced as an internet phenomenon when Google began publishing live crowd estimates, which made the anecdote continuously checkable by anyone. That checkability is what made it spread — and also what makes it easy to over-read.

Every 30 minutes we retrieve a busyness percentage for each tracked location. Those feed a composite score:

• A reading of 0% is treated as no data, not as an empty restaurant. Around 41% of all readings are exactly zero and they track opening hours — a venue that reads 0% at noon and 36% an hour later was not empty at lunch, it simply returned nothing. Zeros are excluded from the averages and counted separately as non-reporting.
• The average of the reporting pizza locations contributes 70% of the composite.
• The average bar reading is inverted (100 minus the value) and contributes the remaining 30% — but only between 5pm and 2am Eastern, when bars are actually trading. Outside those hours the term carries no information, so the pizza average is renormalised to carry the full weight instead.
• If at least five pizza locations are reporting and a majority of them are individually above 50% busy, the composite gains up to 20 points. A single outlier cannot move the index, and neither can a small handful of venues late at night when most are shut.
• With fewer than three venues reporting, the index is capped at GUARDED. Two data points cannot support a claim of elevated activity.

The gating on the bar term corrects a real error in an earlier version of this index. Applied around the clock, the inverted bar figure sat at 75–91 during the day, when bars are shut, and only 56–63 in the evening, when it actually means something — it contributed most where it mattered least, and held the index at HIGH on ordinary weekday mornings.

The composite is then compared against what is normal for that specific half-hour, not against a fixed number. We hold a baseline for every half-hour of the week — separately for weekdays and weekends — built from the median of all past readings in that slot, with the spread measured as a median absolute deviation. A reading has to clear two independent bars to register:

• Unusual for the time. Far enough above the baseline for that slot, measured in robust standard deviations.
• Busy in absolute terms. Overnight the baseline is near zero, so any activity looks enormous in relative terms. Without a floor, 2am would outscore everything.

The two together map to one of five levels:

This is the part of the methodology we think matters most, and it is the main reason this site exists rather than simply linking to an existing tracker.

A pizza-only index cannot tell the difference between “something is happening” and “it is Friday.” Both produce a spike. But the two scenarios differ in a measurable way: ordinary busy nights lift restaurants and bars, while a working night lifts food orders while bar traffic stays flat. Weighting bar activity in reverse partially separates the two.

“Partially” is doing real work in that sentence. Two bars is a small sample, and the control is far from complete — but a pizza-only signal has no control at all.

We would rather state this plainly than bury it in a footer. The index has several weaknesses that no amount of presentation fixes:

• There is no ground truth. The signal has never been validated against actual events, and it cannot be, because the events it purports to detect are not published on a schedule anyone can check.
• The historical cases are selected after the fact. Anecdotes where pizza traffic preceded a crisis are memorable; the far more numerous busy nights that preceded nothing are not recorded.
• The data is an estimate of an estimate.Google's crowd figures are modelled from sampled location data, not counted orders. They can be stale, missing, or simply wrong.
• The Pentagon has its own food service. There is no particular reason staff working late would order from a shop two miles away.
• Publicity contaminates the measurement. As trackers like this one attract attention, some share of the traffic they observe may be caused by the attention itself.

Treat a high reading as a prompt to go read actual news reporting, not as information in its own right.

Nine pizza restaurants and two bars in Arlington, Virginia, all within a few miles of the Pentagon. Every location has its own page with a full reading history — see the full list.

Pentagon Pizza Watch makes money from the companion iOS app and from affiliate links to Kalshi, which are labelled as such wherever they appear. We do not sell a token, run a prediction market of our own, or take positions on anything the index might relate to.

The reason is straightforward: an indicator whose publisher stands to profit from a particular reading is not worth reading. The methodology above is fixed and published so that it cannot be quietly tuned toward more alarming numbers.

What is the Pentagon Pizza Index?

It is an informal observation that unusually high late-night activity at restaurants near the Pentagon might coincide with periods of intense work inside the building. It is a folk indicator popularised on the internet, not a recognised analytical method, and it has never been validated against classified activity because no such validation is publicly possible.

Is the Pentagon Pizza Index real?

The underlying data is real: Google publishes crowd-level estimates for most businesses, and those estimates genuinely rise and fall. What is unproven is the inference. Nobody has demonstrated a reliable statistical link between restaurant traffic in Arlington and military decision-making, and the handful of widely-cited historical anecdotes are selected after the fact.

Where does the data come from?

Google Popular Times, retrieved every 30 minutes for 9 pizza locations and 2 bars in Arlington, Virginia. When a live crowd estimate is available we use it; otherwise we fall back to Google's typical estimate for that day and hour.

Why does this site track bars as well as pizza?

Because pizza traffic alone cannot distinguish a crisis from a Friday night. A busy weekend lifts restaurants and bars together. Sustained work lifts food orders while bar traffic stays flat or falls. Weighting bar activity in reverse is a partial control for the single largest confound in the theory.

How often does the index update?

Location readings refresh every 30 minutes and the composite signal is recalculated every 15 minutes. This page serves a cached copy that revalidates every 5 minutes.

Does a high reading mean something is happening?

No. A high reading means several tracked restaurants are busier than usual while bars are not. That is all it means. Sporting events, weather, holidays, road closures, a single large catering order, or a Google data glitch can all produce the same pattern.