While I'm not surprised with the findings. From a quick search:
The Pentagon features over 30 distinct food service locations, including more than 20-24 various restaurants, fast-food chains, and cafes catering to its 23,000+ employees. The facility includes three main food courts—most notably the 875-seat Concourse Food Court—along with numerous individual kiosks, branded vendors (e.g., Starbucks, Subway, Taco Bell, Popeyes), and the Center Court Café.
When asking specifically about delivery:
Food orders can be delivered to the Pentagon, but not directly to offices; they must be screened at a remote facility or picked up by personnel at designated secure areas like the Pentagon Metro. Perishable items are generally prohibited from being delivered directly to the building, but staff frequently order from local spots during long shifts.
Key details regarding food deliveries to the Pentagon:
- Delivery Procedures: All items must go through the Pentagon Remote Delivery Facility, where they are screened and inspected.
- Pickup Location: Employees often meet delivery drivers at secure, accessible points outside the main building, such as the Metro.
- Internal Options: The Pentagon contains its own food court with options like McDonald's, Five Guys, Starbucks, and Subway, which are accessible to employees.
- "Pentagon Pizza Theory": Sudden surges in local, late-night pizza orders to the Pentagon have historically been noted as a potential, unofficial indicator of increased, high-stakes military activity.
But you do have the mention of the theory you're debunking.
Seems like outside delivery is really complicated and time consuming (though I don't know if the internal food halls can deliver to office door but seem like it might still be much quicker than leaving the building to meet the delivery person). Plus, a lot more than pizza can be delivered these days (but perhaps the "Pizza" in the name should not be taken literally).
As a side note, years pass when I working in the Intelligence field, when I first started reading some of the CIA's classified documents and studies I was surprised by just how much of the source information was from general, publicly available and unclassified information.
I heard similar story to that of the Alchian story many years back. Ph.D. candidates dissertation was about the risks to the power grid (for get if it was just electric or if other distribution networks were considered) which pointed out a number of ways an adversary could disrupt and disable the gird. If got published as is usual and then got noticed, classified and perhaps is no longer even searchable in the dissertation archives (if so it's highly redacted I suspect).
I would lump such cases into the bucket of info hazards.
I like two stories of how the US intelligence located both uranium enrichment plants in operation in mid-1950s USSR: one near Tomsk they identified in 1956 by verifying rumors from a German tailor with an isotopic analysis of uranium traces on a fur hat, the other one near Sverdlovsk in 1958 by an analysis of a photo from an electricity dispatching office published in a Soviet magazine! OSINT at its finest in the latter case
Nice work; the pizza index has always been one of my favorite pieces of OISINT lore. This might be the first time I’ve seen any attempt to actually test it. If I remember correctly, the pizza index originated from a Dominos down the street from the White House.
I have a book somewhere that covers this but can’t remember which one. If I find it, I’ll update this.
Thanks for taking us with you down the rabbit hole! Truly fascinating,
I wonder if this may have been true a (couple of) decade(s) ago, when ordering food was less common and there were fewer pizza alternatives. It is possible that Pentagon guys indeed order more food at such events, but nowadays baseline is so high that it does not bump stats meaningfully.
"However, a Pentagon spokesperson has denied this, telling Newsweek, “There are many pizza options available inside the Pentagon, also sushi, sandwiches, donuts, coffee, etc.”"
In my experience food vendors within office buildings close by the end of official work hours and if you work late you have to order from outside.
I think the hypothesis on Freddie's is that activity would decrease when something interesting is about to happen, because employees wouldn't have any leisure time.
I think stripping Google Maps activity data would probably make for a better dataset, since, as OP points out, this is largely just a graph of the Twitter account's posting history. I think that such a dataset would have to be collected over time, by scraping Maps every few hours or so, unless Google surfaces past activity data through some API endpoint or another.
The big advantage of scraping the twitter account's posting history is that it lets you back test. Any clever analysis we do now could only be forward tested, even if google surfaced the relevant data.
I don't think this is right, but maybe I will calculate new indices to account for this. For one thing, I think the idea is that Pentagon employees are less likely to frequent gay bars than other DC residents. For another, there exist tweets in the dataset that look like this:
HIGH activity is being reported at the closest Papa Johns to the Pentagon.
Freddies Beach Bar is reporting abnormally low activity levels for a Saturday at 7:11pm ET.
Classic indicator for potential overtime at the Pentagon.
As soon as modern data analysis became a thing, the US government has had to deal with people trying to use open source data to uncover its secrets.
During the early Cold War days and America’s hydrogen bomb testing, there was an enormous amount of speculation about how the bombs actually worked. All nuclear technology involves refinement and purification of large amounts of raw substances into chemically pure substances. Armen Alchian was an economist working at RAND and reasoned that any US company working in such raw materials and supplying the government would have made a killing leading up to the tests.
After checking financial data that RAND maintained on such companies, Alchian deduced that the secret sauce in the early fusion bombs was lithium and the Lithium Corporation of America was supplying the USG. The company’s stock had skyrocketed leading up to the Castle Bravo test either by way of enormous unexpected revenue gains from government contracts, or more amusingly, maybe by government insiders buying up the stock trying to make a mushroom-cloud-sized fortune with the knowledge that lithium was the key ingredient.
When word of this work got out, this story naturally ends with the FBI coming in and confiscating Alchian’s data-driven research and the G-men giving him a stern lecture on national security, but he had just invented the first event study of the modern world.
Pizza is the new lithium
As you might have guessed, Alchian’s intellectual heir is the X account Pentagon Pizza Report, which tracks activity in pizzerias with proximity to the Pentagon as reported by Google. Started in 2024 and with over 300K followers, it may be the gold-standard OSINT meme account. Polymarket has even formalized it.
Before the X account, I had never heard of the pentagon pizza theory, but it has its own wikipedia page and amazingly, this idea goes all the way back to the 1990s. The chain of causation is clear. People only work late in the Pentagon if there’s a lot of military work to do. When there is a lot of work to do, they get hungry and order a bunch of pizza. The Pentagon is huge and might strain local pizza production capacity; this shows up in Google’s location/traffic data that the feed captures. The world is messy and pizza data is messier, but can someone predict large scale US military action from public data on pizzerias?
The Pentagon has apparently disputed this notion:
We’re going to backtest this idea in the rest of this poast and get to the bottom of this one.
The Data
The Twitter API used to be a cornucopia for interesting data projects. Now you have to pay a fortune for API access and it’s legitimately terrible. I actually could not get anything to work using the API or even using standard scraping tech to get the Pentagon Pizza Report’s tweet history; Musk has shut a lot of that shit down. I couldn’t even get ChatGPT, run by Musk’s rival Sam Altman, to help me improve the scraping against Twitter because it thought that was unethical. I even DM’ed the Pentagon Pizza account to see if he would share his data, but he did not get back to me.
I ended up getting Claude code to write me a sophisticated Selenium script that very slowly and annoyingly scrolled through the tweet history in a browser and recovered the data as it did so.
I ended up with 648 clean PPR tweets. Here’s a short excerpt.
The Backtest
Using this history, I wanted to check out US military adventures of the past year to see if they can be observed in the Pizza Report’s data. I could have selected others, but these are the most prominent that overlap with the Pentagon Pizza Report’s history and for the purposes of this study I preregistered them. I haven’t yet looked at any of the Pentagon Pizza Report’s data as of writing this section.
Fordow bombing
On June 22nd, 2025, Operation Midnight Hammer was a sophisticated collection of strikes targeting Iranian nuclear infrastructure. This involved a large military footprint and careful planning much of which doubtlessly took place with involvement from the Pentagon.
Maduro capture
On January 3rd, 2026, Operation Absolute Resolve was an operation to capture Venezuela’s leader Maduro and bring him to justice in the United States. Again, this involved sophisticated intelligence and military cooperation in the Caribbean.
The Houthi stuff
This one’s interesting and takes place over a longer time frame, March 15th going into May 2025. Operation Rough Rider consisted of strikes against Houthi positions obstructing shipping in the Red Sea. This choice potentially allows us to see any longer term pizza signal that exists in the X data.
Evaluating whether PPR activity is faithfully related to these three events is a complicated statistical problem. The PPR history is quite clearly not a random sampling or uniform record of pizza-related activity near the Pentagon; I couldn’t confirm this with the user, but it’s plausibly selecting to tweet during unusual pizza activity levels. Without knowing how it works, you can’t make many sound statistical assumptions about it.
Let’s look at a few basic series.
This dataset is a dog’s breakfast. Look at that giant gap in the volume! This won’t be useful at all for examining the Houthi strike test in March 2025, so we’re stuck with just the Maduro capture and the Fordow bombing. There is a pretty astonishing spike in tweet volumes leading up to the Fordow strikes. To my eye, this looks more like the hiatus ending than some sort of tweet storm preceding that, but we’ll return to that. First off, and most importantly, WTF is “Freddie’s”?
Freddie’s Beach Bar and Restaurant is a gay bar in Arlington near the Pentagon. As you can see from Figure 2, starting in June 2025, PPR began posting about the activity status there.
Maybe it was part of some account reboot after the long hiatus lasting into June 2025. It is obviously hilarious, and what’s interesting about the Freddie tweets is that the account author pretty clearly intends for this to act as a control for the pizza activity measure. If after hours military activity at the Pentagon is causing increased pizzeria patronage, Freddie’s won’t be affected by that. But if some exogenous thing is causing increased activity for both nearby pizzerias and Freddie’s, we’d see that and know to discount how meaningful the pizza signal is.
So bravo to the PPR account for this little statistical improvement. This means that the most meaningful parts of Figure 2 will be when the yellow line and the blue line diverge.
I created a pair of slightly differently specified “Freddie-controlled” pizza activity indices which track the divergence of the Pentagon Pizza market activity versus Freddie’s activity. I think both of these are reasonable measures of what we’re trying to capture.
Using these, I really don’t see it. Both of these raids took place during utterly unremarkable pizza activity index times. In fact, the Maduro capture occurred at a positive nadir. Also, you can see these indices are correlated with each other, suggesting there’s some existing robustness in our measurement.
Coda
There are probably more sophisticated ways of testing this on a cleaner more complete dataset, but as a first pass, I’m going to declare the Pentagon Pizza Theory doesn’t work. If it did we'd expect to see at least see correlation: the Freddie-controlled index spike in the time before major operations. We don't. Hopefully, this means the FBI won’t find this and raid my house.