Revisiting the Pizza Principle, Twelve Years Later

The bad news: a slice now costs more than a New York City subway ride. The good news: agents made it much easier to find out.

Twelve years ago I went to a talk about the history of Brooklyn pizza and came home with a scraping project. This is a fairly normal outcome for a statistician with a passion for pizza and no kids waiting at home (at the time).

My friend Scott Wiener of Scott’s Pizza Tours was speaking at the Brooklyn Historical Society when someone brought up the old idea that the price of a plain slice tends to move with the New York City subway fare. Scott wondered whether the spread of dollar-slice shops had broken the relationship and stared at me while joking that he needed a statistician in the audience. Naturally, I went home and tried to calculate it.

The 2014 analysis involved scraping more than 1,800 MenuPages listings with R, downloading individual menus, parsing inconsistent HTML and trying to isolate the price of a slice of cheese pizza. That last part required a long exclusion list containing pepperoni, mushroom, Sicilian, grandma, calzone, soda and a surprising number of misspellings. I also removed prices above $6 because they were usually whole pies or something else that had wandered into the data.

After all that, the mean price was $2.33, the median was $2.45 and the standard deviation was $0.52. The subway fare was $2.50, or about $2.38 after the MetroCard bonus. The Pizza Principle held up remarkably well.

When I came back to the analysis in 2026, my approach looked different. A small swarm of research agents could work through menu sites, local reporting and old fare references. Coding agents could parse the data and assemble the analysis. That combination would get you pretty far.

Then, of course, the research agents found SliceDex. JP Jaramillo had already built the dialer.

As of this writing, SliceDex reports an overall average of $3.42, which includes dollar-slice shops. Using its displayed price buckets (and excluding the 35 slices priced at $1 or $1.50 for consistency with my 2014 analysis), I get an approximate average of $3.56 with a standard deviation of $0.62. The subway fare is now $3.00. A plain slice is about 56 cents, or 19 percent, more expensive than a ride.

Anyone who has bought groceries or eaten at a restaurant recently could have guessed the direction. What blew me away was how little of the 2014 workflow I had to rebuild by hand.

When the prices separated

The Pizza Principle was always a loose observation. In 1980, Eric Bram noted that a slice and a subway ride had moved together for years and used pizza prices to anticipate a fare increase. It became a durable New York story because both prices were ordinary, visible and easy to remember.

My 2014 data fit that story almost too neatly. The mean slice price was within about five cents of the effective subway fare, while the median was within five cents of the base fare. Business Insider revisited the analysis after the fare increased in 2015 and reported another close match.

The evidence becomes less tidy after that because the later datasets were collected in different ways. Liam Quigley, who interviewed me for a segment on NPR’s WNYC not long after my original blog post, tracked 464 slices that he personally bought over eight years (excluding dollar slices). His plain-slice average rose from $2.52 in 2014 to $3.00 in 2022. In 2023, Quigley sampled 30 new shops and found an average of $3.24.

Between 2015 and 2022, the price of a New York City subway ticket stayed flat at $2.75. In 2023, it was raised to $2.90. Then in January of this year, the MTA raised the fare to $3.00.

There is no precise date when a folk principle stops working. The available evidence suggests the prices began separating in 2021 and were clearly out of alignment by 2022 or 2023. Pizza shops had to pass on increases in ingredients, labor and rent more quickly than the transit system raised its fare. The MTA also has subsidies and political pressure working against frequent fare increases. A pizzeria has a menu and a stack of bills.

In 2014, the scraper was most of the project

Looking back at the original code is a good reminder of what a modest data project required at the time. I first scraped 19 pages of restaurant listings to collect over 1,800 names and URLs. Then I downloaded each menu, searched the files for the word “slice,” parsed the menu tables and extracted anything that looked like a price. The code had to handle different layouts, missing fields, unreadable characters and menus where important information appeared in a category heading rather than the item row.

Then came the exclusion list. It is hundreds of terms long, including duplicates, abbreviations and partial words. That was the practical way to remove specialty slices, toppings, pizza rolls, drinks and entire pies. My code that parsed the XML was probably more complicated than it needed to be, which I also admitted in the original post. It worked well enough to produce 651 usable observations across Manhattan, Brooklyn and Queens. Including the code in the post gave readers an audit trail of what I downloaded, how I filtered it and where judgment was essential.

There were also clear limitations. MenuPages did not give me the Bronx or Staten Island. Posted prices could be stale. A name-based filter could remove a legitimate slice or keep something that did not belong. The final statistics described the restaurants for which the pipeline found a believable price rather than every pizzeria in New York City.

Rebuilding the analysis in 2026

I could not simply rerun the old scraper because MenuPages no longer offers the same stable universe of menus. I used research agents to find later studies and reconstruct the fare history, then put the reported prices, sample sizes and inclusion rules into a CSV. Coding agents wrote much of the comparison code. I checked the arithmetic, source attribution and assumptions separately.

The research agents in this analysis did not collect new shop-level prices. Conveniently, they found and reconciled work from people who had already collected them. The latest price data came from SliceDex, while our agent workflow handled the source discovery, normalization, coding, comparison and review.

SliceDex is itself an agent-assisted collection project. Its AI caller (named Heather) used ElevenLabs and Twilio to call 1,766 pizza shops and ask for the price of a plain slice. Shops answered 1,346 calls while 838 provided a price. Its current index lists 829 prices across all five boroughs, with Claude used to extract structured prices from the call transcripts.

The price distribution also lets us test the result more carefully. After removing the $1 and $1.50 slices, the displayed buckets contain 794 prices with an approximate mean of $3.56 and standard deviation of $0.62, comfortably above the $3 subway fare (though still within one standard deviation).

SliceDex has prices for 829 of the 2,252 shops it tracks, so the estimate still reflects the shops that answered and produced a usable quote. That is broader coverage than my original analysis, though it is not the same as observing every pizzeria in the city. Then again, sampling is a foundational principle of statistics.

The agents got me to a working dataset quickly. They found sources, pulled reported values into tables and wrote the first pass of the comparison code. Then I had to decide whether a posted menu price, a phone quote and a slice someone actually bought belonged on the same chart. That is where most of the review went.

This was a small project, but the workflow looked familiar to larger work at Lander Analytics. We have adjusted how we approach projects because the first useful version now comes together much faster, whether we are writing code, analyzing data, conducting research or building a new platform.

The updated principle

The Pizza Principle held up for a long time. Based on the available evidence, it stopped being a reliable estimate of the subway fare only over the last five years. A plain slice now averages over $3.50 compared to a $3 ride. At the moment, the subway may simply be the better deal (especially in this summer’s heat).

The two versions of this project also make a decent time capsule of applied data science. In 2014, the hard part was building a scraper and hand-coding enough rules to turn menus into a dataset. In 2026, agents handled much of the research. I spent more of the project deciding which prices belonged and how to tell this story I’ve been thinking about for over 12 years.

Jared P. Lander

Founder and Chief Data Scientist

Lander Analytics

---

Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.

Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com.

About the author: Jared P. Lander is Chief Data Scientist and founder of Lander Analytics, where he helps organizations build practical, measurable AI workflows grounded in strong data foundations.