In March 1984, masked men broke into the home of the president of Ezaki Glico, the candy company, and took him out of his bath at gunpoint. He escaped after three days.¹ Then Glico warehouses started burning, and letters started arriving at newspapers, signed かい人21面相, “the Monster with 21 Faces,” a villain borrowed from prewar detective novels.² Over the next year and a half the group extorted most of the Japanese confectionery industry. They put cyanide-laced candy on store shelves, some of it with a note on the package reading “danger: poison inside, you’ll die if you eat this.” They taunted the police by name; “a fool stays a fool however hard he tries,” one letter to the newspapers said of the investigators. They demanded ¥50 million here, ¥100 million there, at one point ¥1 billion plus 100 kilograms of gold.

They never collected any of it. Every ransom drop failed or was abandoned. In August 1985 they mailed a farewell letter, “we’re done bullying food companies,” and disappeared. Nobody was ever charged. The statute of limitations ran out in 2000.³

A crime this elaborate with no visible payout needs an explanation, and there’s a popular one: the ransoms were theater, and the real money was made on the stock exchange. Sell the victims’ shares short, poison the candy, collect. The police took this seriously enough at the time to investigate speculator groups. Victim stocks really did fall. Morinaga dropped about 8% when the poisonings went public and kept sliding.⁴ It’s a tidy theory, it has been repeated for forty years, and as far as I can tell nobody has checked it against market data. The police came closest. In 1984 they canvassed the victim companies, the brokers and the exchange itself, ran into broker confidentiality, and named one suspect, a Tokyo speculator group called Video Seller, which Kabutocho had nicknamed “the stock market’s Monster with 21 Faces.” Its chairman turned up dead two months after the case ended, and there were no charges and no decisive evidence.⁵ Since then the theory has lived in investigative books (Fumiya Ichihashi’s, and a more sophisticated version from the writer Manabu Miyazaki, who was himself a suspect), in a 2016 novel and its film, and in a 2025 business-magazine chart of how Glico’s and Morinaga’s share prices moved. None of them tested it.⁶ The ones that cite a number all cite the same one, a 24% drop in Glico’s stock between January and May 1984, which happened in full public view and proves nothing about foreknowledge.⁷

There’s a reason it’s checkable at all, and it’s a detail of how Japanese media handled the case. Several of the extortions were kept out of the press entirely while negotiations ran. Marudai Foods was threatened in June 1984 and the public learned about it in November. House Foods was threatened in November, reported in December. Fujiya in December, reported the following January. During those blackout windows the threat was known to the company, the police, a few editors, and the criminals. If somebody shorted Marudai in July 1984, they either got very lucky or they were standing very close to the crime.

The short answer is that nothing moved in any of those windows. The rest of this post is how I know that, and then the one thing in two years of data that did move, which turned out to have nothing to do with the crime.

The data problem

There is no clean dataset of daily Japanese stock prices from 1984, no API, no vendor file worth having. What exists is the Tokyo Stock Exchange’s own daily bulletin (東京証券取引所日報), which JPX has scanned and put online as monthly zip files of TIFFs.⁸ One multi-page scan per trading day, roughly 4700x6500 pixels per page, black and white, photographed from what are clearly the exchange’s bound file copies. Some pages have handwritten corrections in the margins from whoever kept the files in 1984. I’ll call each scanned page a plate. That is what it is, a photograph of one printed sheet, and the rest of this post is about getting numbers off them.

I pulled the 24 months covering 1984 and 1985: 572 trading days, about 2.5 GB of TIFFs. Each day is 16 pages. Pages 1-8 are the price tables for every listed stock. Pages 12-15 are the convertible bond market. And page 16 is where the exchange printed its surveillance of its own market: per-stock margin balances, the list of stocks under restriction, and a small table of stock-lending fees.

You don’t need Japanese to follow what comes next, but you do need five pieces of market plumbing, because the whole argument runs through them.

A short primer on the 1984 Tokyo market

Securities codes. Every listed company has a four-digit number, grouped by industry. Foods are the 2000s: Morinaga is 2201, Fujiya 2211, Marudai 2288, House 2810, Glico 2206.⁹ The bulletin sorts by code everywhere, which matters more than it sounds, because position in a sorted list identifies a company even when its name is an unreadable smudge.

Two sessions a day. The exchange closed for lunch. Every stock got a morning session and an afternoon session, each with its own open, high, low and close, and the bulletin prints all eight numbers.¹⁰ It matters here because someone acting on information received overnight has to queue an order for the morning open, so that is where informed trading would show up.

Margin trading. Investors could buy, or sell short, on credit through a broker. A margin short sells borrowed shares. For any stock it was keeping an eye on, the exchange published the outstanding balances: shares currently sold short (the short balance) and shares held long on credit (the long balance), in thousands of shares, updated daily.¹¹

Stock lending, and the fee. A short seller’s borrowed shares come from a securities finance company, which sources them mostly out of the margin longs. Only stocks on an approved list could be borrowed this way; the bulletin marks them with a dot next to the name. When the shorts in a stock borrowed more than the finance company could supply, there was a borrow shortage, and an auction was held to find the rest. Most days the shortage cleared at zero. When it didn’t, the shorts paid a fee per share per day, quoted in sen, 100 sen to the yen, which traders call “reverse daily interest.”¹² A positive fee means the short side of that stock was crowded past what the market could lend.

The watch list and emergency rules. The exchange published which stocks it had under surveillance for speculative trading, and when things got out of hand it raised the collateral required to trade them on margin. Both sit on the same page as the lending fees.¹³

Five things you can measure, and what a crime would look like in each

If somebody shorted a victim company knowing an extortion was coming, here is what each instrument the bulletin supports would show:

1. Prices. The target underperforms its sector during the concealed window as the selling leaks into the price. Daily closes, adjusted against the other food stocks.
2. Sessions. That underperformance loads on the morning session, because overnight information gets acted on at the open.
3. Volume. Unusual turnover while the position is built.
4. Margin balances. The target’s short balance climbs through the window. This is the most direct record there is: a count of shares sold short.
5. The borrow. The target appears in the borrow-shortage list, and if the position is big enough, starts paying a fee. The list is binary, a stock is on it that day or it isn’t, which makes it the cleanest of the five.

A sixth, the convertible bond market, gets its own test later, because it allowed a trade the other five cannot see.

Before trusting any of them I planted fake signals in the real data (a synthetic daily drag on one stock, a volume spike, a run of borrow-shortage days) and checked that each instrument found its fake. Where an instrument failed to find its fake, the corresponding result below is marked as inconclusive.

Reading the pages

The price pages first. This is the foods section on June 20, 1984, two days before the Marudai letter, with four of the five companies in the case boxed (House Foods, 2810, sits further down the section):

One stock per row. Size and valuation columns on the left (capital, P/E, dividend), then the code, then the name with a dot if the stock is loanable and therefore shortable, then the eight prices, morning and afternoon, then the change against the previous close with a triangle for direction, then volume. Every extortion target has the dot. One trap, labeled in black: Morinaga Milk (2264) is a different listed company from Morinaga the candy maker (2201), four rows apart, and both OCR and I confused the two at first.

Page 16 next, where the exchange watches its own market. The margin table:

For each stock under surveillance: the short balance, the change since the last report, the long balance, and its change, all in thousands of shares. The triangles give direction, open for up and filled for down. The day shown is July 30, 1984, and Morinaga’s short balance is up 2.4 million shares in a single report. That is the tail end of a squeeze in Morinaga’s stock that summer. It has no connection to the extortion letter, which was still six weeks away; it gets its own section further down.

Now the lending-fee table:

The names run left to right across a fixed-width field, so a two-character name has a gap in the middle, and the rows are sorted by code. Codes aren’t printed, but position identifies names: the construction companies in the 1800s come first, then Morinaga (2201), and Fujiya (2211) can only ever sit one or two rows below it. I used that structure to check every name match.

A stock appears here at all only when its short sellers borrowed more than the finance company could supply that day: it is the day’s list of stocks where the short side outran the lendable float. 0 sen means the shortage cleared free. A positive fee, like Morinaga’s 10 sen here, means the shorts were crowded enough that lenders charged for the privilege. The right-hand column of the same day shows the range: Toppan and Seiyu at 1 yen 20 sen, Daiei at 50 sen. Morinaga’s 10 sen sits at the low end of it.

This table matters more than the prices. Price-based tests have a noise floor: a modest trade hides inside ordinary volatility and you can never prove it isn’t there. A borrow shortage either happened on a given day or it didn’t, and the exchange printed which, every day, for two years.

One more table:

Emergency margin measures: 60% collateral, 20% of it cash, to short Morinaga. Printed September 11, 1984. The extortion letter reached Morinaga’s Kansai office on September 12.¹⁴ So the exchange had already clamped the stock the day before the letter arrived; it was reacting to the July squeeze, not to anything the criminals did.

Getting numbers out of the scans

Off-the-shelf OCR is useless on these. Tesseract’s Japanese model reads Morinaga’s name as a different set of characters entirely and invents digits in the price columns. What worked: find each table by template-matching its printed banner, find the columns from the ink itself, cut individual cells, and read the numeric cells with TrOCR on a GPU, which handles degraded print far better than classical OCR.¹⁵ Company names never go through OCR at all; they’re matched against reference cut-outs of the printed names and checked against the code ordering, and in stubborn cases I read them off magnified crops myself. The margin balances for the two targets the exchange printed them for, 47 days for Morinaga and 65 for Fujiya, I transcribed by eye.

Even the good OCR output was junk at first. When I computed daily returns from the raw reads, over 90% of the variance was transcription error. A stock closes at 520, the 5 reads as 3, and you’ve invented a 38% crash. So every price goes through a repair layer that models each stock as a slowly moving level and rejects cells that don’t reconcile with it. The convertible bond pages, which print each bond’s parity, give an independent check on a different page of the same paper: where both survive, 83% of the pairs agree within 3%, and the typical gap is a fraction of a percent.

That check still runs my own pixels through my own code. A better one uses data I never touched. The Nikkei 225 daily index for 1984-85 is freely available from FRED, and it has nothing to do with any of this: different institution, different index, digitized decades apart from my scans.¹⁶ If my food stocks are real, their daily moves have to covary with the market. Built from raw OCR, an index of my food stocks has a daily standard deviation of 44%, fifty times that of a real index, and a correlation with the Nikkei of 0.076. Run the same construction on the repaired data and it becomes a 0.61% daily standard deviation (the Nikkei’s own is 0.80%), correlation 0.505, and a market beta of 0.384, which is about what a defensive food sector should have. On the five worst days the Nikkei had that year my food index is down every time; on the five best it is up by much less, the way defensive sectors behave. Shuffle the dates and the correlation collapses to zero. How good the numbers are is measured in the appendix, against 33 plates I read by eye. The short version: the raw extractor gets 89% of cells right; the repair layer keeps about half the weekday closes, and 97.5% of what it keeps matches the plate exactly; a single stock’s daily returns carry at most about 1.1% of noise, and the tests below were sized to see through that.

One more thing about method, because everything below is a negative result, and negative results are easy to fake. Several times during this work a finding looked real and died on inspection. The clearest case was House Foods’ morning session. The price tests showed House underperforming in the morning sessions of its concealed window at p=0.011, roughly a one-in-ninety chance of arising by luck, exactly what an overnight-informed seller should produce. It held up until I looked at which days were in the window: it included December 11, the day the press blackout on House lifted, and the morning selling was the public reacting to the news. Move the window edge by one day and the result was gone. Everything below has been through that kind of check, with the window edges fixed in advance.¹⁷

The data that came out of all this: daily prices and volume for about 50 food-sector stocks plus a machinery-sector control group, 1981 to 1985; morning and afternoon session returns for 1984; margin balances for Morinaga and Fujiya; the convertible bond market; and the full borrow-shortage record, 472 published days, 14,737 entries, each tied to a name and a fee.

What the theory predicts, and what’s actually there

If somebody traded on the concealed windows, the target should underperform the other food stocks during the blackout, short interest should build, and the borrow should tighten. None of that happens, in any of the windows, at any strength the instruments can see.

Prices: each target against the other food stocks, through its own concealed window, compared with random windows of the same length from the same stock’s history. Marudai’s window is the long one, June to November, 53 usable trading days, and it comes back flat with enough data to have detected a drag of 0.14% per day. The morning-versus-afternoon split, which would catch someone queuing orders overnight, is flat too. Pooling all the windows and hunting aggressively, trying every defensible window edge, the best surviving result is Morinaga’s autumn decline at p=0.09, roughly a one-in-eleven chance of showing up by luck. The margin data explains that one.

Morinaga was under surveillance, so the exchange printed its short balance daily. Through the concealed window, the weeks an informed short would be building, Morinaga’s short interest was being covered, down from about 26.7 million shares in late July to 11 million by the day the poisonings went public. That is the p=0.09: the stock drifting down while the shorts bought back.

Here is Morinaga’s complete two-year record in the lending-fee table, days listed per month:

Two things stand out. First, the instrument works. From July 12 to August 6, 1984, Morinaga was in borrow shortage on 17 of 23 sessions (13 of them in July) with the fee going positive three times, the squeeze that earned it the 60% margin requirement in the restriction table above. When shorts crowd this stock, the table shows it and the regulator moves. Second, look at September and October. The concealed window, September 12 to October 6, ending the day before the cyanide announcement: zero appearances in 18 published sessions, against about one expected from the stock’s own base rate, and the planted-signal test says six would have been flagged. October after the announcement: zero. November and December: zero.

Fujiya, same test, its own window in December and January: listed 2 days out of 21, against 7 expected from its base rate. Nothing unusual, though I lean on Fujiya less than on Morinaga. Morinaga’s borrow listings are corroborated by its margin table: on days it’s in the fee list its shorts exceed its longs, on other days they don’t, with no overlap. Fujiya’s margin numbers show no such pattern, so its listings may mean something slightly different, and I read them as no sign of crowding rather than as a second hard no. House Foods never appears in its window at all. There are entries reading “House” in those weeks, but they sit above Ajinomoto (2802) in the code order, which makes them Daiwa House or Sekisui House, the construction companies, codes 1925 and 1928. House Foods is 2810 and would print after Ajinomoto. Marudai appears nowhere in two years.¹⁸

For completeness, one price result did survive its checks: the food sector fell about 1.4% against the market on the day reporting of the House extortion resumed. So the instruments do see the public reaction to the news; what they don’t see is anything ahead of it.

The convertible bonds got their own test, because they were the one instrument in 1984 Japan you could trade with real anonymity, no margin account, no printed short balance. You can’t practically short a convertible, so the trade available to an insider is the reverse: crash the stock with your own crime, then quietly accumulate the bond in the panic and ride the recovery. House’s bond, through the exact post-disclosure trough where that accumulation would happen, traded at the bottom of its own volume distribution, less than normal rather than more, with a control that detects a threefold volume increase 96% of the time.¹⁹

The one time the money moved

There is a version of the theory that isn’t about shorting at all, and it is the strongest one anyone has proposed. Manabu Miyazaki, a former stock reporter who was himself questioned in the case, argued the money was in a raid: accumulate a big block of the target quietly, run the price up, squeeze the shorts, and then get the company or its friendly shareholders to take the block off your hands at a premium. He put the figure at ten billion yen.²⁰ Every test above is blind to that by construction, because they all look for selling pressure. So I went back and looked for buying pressure, and found the squeeze I mentioned when reading the margin table. In the summer of 1984 Morinaga’s stock went from 305 yen on June 25 to 654 on July 31. On July 30 alone, 41.8 million shares traded (I checked that figure against the plate by eye, since it was the one number big enough to be an OCR error, and it is right), sixty times a normal day and more than a tenth of the entire company in a single session. The borrow went into shortage and the exchange imposed its 60% collateral rule. It has the shape of a raid, and it is the one place in two years of data where somebody unmistakably made a great deal of money in a Glico-Morinaga target.²¹

And here is the row on the July 30 price page that the big bar rests on, one continuous strip: code 2201, the name, the morning session’s open, high, low and close, the afternoon’s, then the day’s change (up 70) and the volume cell reading 41780:

How unusual is it? Taking every stock in my 1984 data with clean coverage and measuring each one’s best five-week gain, Morinaga is first of 44, and first again after dividing by each stock’s own volatility: 4.5 standard deviations against a next-best 3.4. Only four of the 44 managed a 50% gain in any five-week stretch that year, and the runners-up mostly peaked in other months, so this isn’t the summer rally showing through. Over the identical window the Nikkei fell 1.7% and the other food companies went nowhere: Meiji down 4%, Glico up 2%, Fujiya down 6%. Two caveats apply. The borrow fee itself was ordinary, 10 to 30 sen, where nearly half of all positive fees that year were larger, so the stock was crowded but not historically so. And my data covers 44 food and machinery names out of roughly a thousand listed in Tokyo, in a year the market rose 17%. It was the most violent move in my data; whether it was the most violent on the exchange that year, 44 names can’t say.

The problem for the theory is the calendar. The raid peaked on July 31 and was already unwinding when the extortion letter reached Morinaga on September 12: shorts had halved, the price was down to 540. The raid was over before the extortion began. And the people who were long on margin never got out. Their balance sat at 24 to 28 million shares straight through the concealed window and into the cyanide crash, from 648 to 466. The crowd that followed the raid ended up as the extortion’s victims. Nor was any other target run up before its own letter: a test on the thirty sessions before each threat comes back flat for all five companies. What market data can’t exclude is a block sale to friendly hands in August, off the exchange, followed by an extortion with some other motive. That is a question for paper records, and they exist: the fiscal 1984 securities reports list each company’s ten largest shareholders, and a new large holder who appeared in Morinaga or Fujiya that year and then vanished would settle it. They’re on microfilm in an Osaka library.²² If anyone goes, I’d like to know what they find.

There is one loose coincidence. Morinaga’s raid starts at the end of June, but its borrow squeeze, the part the lending table records, begins on July 12, the first trading session after the date usually given for a batch of threat letters mailed to several food companies at once, letters that reportedly stayed out of the press until October. The sourcing on those letters is thin, Morinaga isn’t clearly among the addressees, and a date match found after the fact proves nothing.²³ But who got the July letters, and when, is the other archive question I’d like answered.

How much money was even in it?

All of the above is absence of footprints, and small trades don’t leave footprints. So the last calculation flips the question and asks how much money there was to make. Assume perfect foresight of every event and of the date each one became public (the criminals controlled those dates, so this is fair), zero market impact, and 1984’s actual trading costs, which were not small: fixed commissions plus a 0.55% securities transaction tax.²⁴ The constraint that remains is volume. You cannot sell short more shares than the market buys from you.

One trade in the list runs the other way. On June 26, 1984 the gang sent a letter saying it was done with Glico (“if we make children cry, that’s trouble for us too; we forgive Ezaki Glico”), and Glico’s shares, down by a quarter since the kidnapping, recovered on the news. Someone who knew that letter was coming buys rather than sells, so that trade is priced as a long.

Run the campaign at 10% of every session’s printed volume, every session of every concealed window, and cover into the panic afterwards:

(Morinaga gets a range because only 9 of the roughly 17 sessions in its window have usable prices; doubling the printed figure is the generous reading, and it changes nothing below.)

House Foods is the clearest case. The stock traded around 33,000 shares a day, and the most the entire concealed window could have paid, with perfect knowledge and aggressive execution, was about two million yen, against a hundred-million-yen demand.

At 25% participation Marudai’s ceiling does clear its ransom, ¥98M against ¥50M demanded. But 25% of every session for five months in a stock that trades 73,000 shares a day is not a hidden trade; at that size it is the market, and the price tests, the borrow table and the margin ledger would all have shown it. None of them do, and the margin ledger shows the opposite. Put together: the quiet version of the trade could not have covered a single ransom, and the version that could would have shown up in at least three of the records above.

And there was no way around the cash market: no stock index futures in Japan until 1987, no listed options until 1989.²⁵ The instruments you would use to do this properly did not exist in 1984.

What I can’t rule out

The first gap is Osaka. This was a Kansai crime: the offenders worked out of Kansai, the letters went to Osaka newspapers, and four of the six targets were Osaka-area companies dual-listed on the Osaka Securities Exchange, whose 1984 daily records were never digitized.²⁶ I checked: they exist as bound volumes in a couple of libraries, and that’s it. Prices arbitrage between venues, so the price tests and the ceiling arithmetic above cover Osaka fine, and the one stock deep enough for its ceiling to reach ransom scale, Morinaga, was listed in Tokyo only.²⁷ But Osaka’s margin books and its own lending-fee auctions are invisible to me. A modest short routed through an Osaka broker would not appear in anything I’ve shown you. The ceiling arithmetic says such a trade could not have paid ransom-scale money, but nothing here rules out that it happened.

The second gap is small trades. Below roughly half a percent a day of price impact, and below the reporting thresholds, a position is invisible to every instrument here. The only answer to that is motive: such a trade nets a few million yen across the whole campaign, and nobody kidnaps a CEO and poisons store shelves for eighteen months for less than the smallest ransom bag they walked away from.

The third is coverage. January and February 1985 prices are missing (the bulletin changed its print format and broke my column detection; fixable, not yet fixed), so the Fujiya announcement and the last months of the campaign, up to the farewell letter of August 1985, were tested through the borrow record only. The Glico kidnapping window I never tested, because I could not pin down its press-blackout dates well enough to define one. And all of it is OCR of forty-year-old scans, repaired and cross-checked but still a reconstruction.

Verdict

Two years of the exchange’s own paperwork and five independent instruments, and every properly specified test comes back flat. The shorts were covering during the weeks when the theory needs them to be building. The borrow table, which visibly registers a squeeze when speculators pile into this stock, is silent through every window in which the criminals held private information. The convertible bond market showed nothing at the troughs. The one violent move in a victim stock, the July raid, was over before the criminals wrote to the company, and the people who rode it were the ones who lost.

Why the Monster never took the money, I don’t know; nobody does. But the market didn’t pay them, and it couldn’t have: played perfectly, the entire seventeen-month campaign was worth less on the Tokyo Stock Exchange than one ransom drop they abandoned.

Appendix: how accurate is the data

Against 33 plates I read by eye, spread over the whole of 1984 (the original three plus thirty chosen at random: 718 rows, 4,300 price cells), 89% of the price cells the extractor produced are right, and it produced 542 of the 718 rows, so end to end the figure is 69%. The errors are not spread evenly. Four plates have a column-splitting failure in which most cells come out wrong; the other 29 run between 91% and 100%. Those are numbers for the raw read, which the tests never see. What the tests use is the repaired close, and on the same 33 plates the repair layer kept a close for 37% of the stock-days (51% on weekdays; the one-session Saturday rows mostly fail its checks and are dropped). Of the 237 closes it kept, 231 match the plate exactly, five are within 2% (four of them the morning close standing in for an unreadable afternoon cell), and one is wrong by 7%. The four broken plates contribute almost nothing to the kept set, which is the point of the layer: on a bad plate it leaves gaps rather than wrong numbers. Every disagreement between my reading and the extractor on an otherwise clean plate got a second look at double magnification; of ten such cells the extractor was wrong nine times and I was wrong once, and that one is corrected.

How much noise that leaves in the daily returns can be measured directly, because a transcription error in one price hits two consecutive returns with opposite signs and leaves a fingerprint in the autocorrelation. Measured that way, at most 18% of a single stock’s daily return variance is leftover noise, about 1.1% per day, which averages down to 0.16% per day over the longest window I test. The planted-signal test on the same window, a different method entirely, put the detection floor at 0.14% per day. Two methods landing on the same number is the main reason I trust the error estimate.

A random sample can still miss the worst plates, so I also went looking for the plates most likely to be garbage. The pipeline’s own diagnostics (share of cells the repair layer rejected, spread of the day’s returns across stocks, residual against the Nikkei) rank every 1984 plate by suspicion, and I hand-read the three worst. Of 47 closes, the pipeline kept 23: 21 exactly right, the other 2 morning closes used where the afternoon cell was unreadable, 1 to 3% off. It threw away 24, and 19 of those really were garbage. No fabricated value got through; the failures are missing values. One of those plates is July 30, the day Morinaga closed at 648 after closing at 305 five weeks earlier, the peak of the July squeeze. A repair layer that rejected big moves as OCR error would have erased it; this one kept it, because the move came in daily steps of 5 to 8%.

Notes

89.23% of the public GraphQL endpoints in the Tranco top million serve the same schema: one file, byte-identical across all of them. It’s Shopify’s Storefront API. Add Magento and you’re at 97.67%. Everything else on the public web, every hand-built API and every other platform put together, comes to the remaining 2.33%.

That one number changes how to read all the others. Most published statistics about GraphQL on the web are computed over a population that is, in practice, one vendor’s default settings. So this post is in two parts: the survey, and then what happens to each of the usual numbers when you divide by vendor instead.

It helps to have a picture of the thing being counted before the counting starts. A GraphQL schema is a directed graph: types are the nodes, and a field on one type that returns another type is an edge. Here are four of them.

Figure 1. Four schemas as type graphs. Node = a type, edge = a field returning another type, and only types reachable from the root query are drawn. Shading is distance from the root. Node size and edge width are identical across panels, so the difference in density is a difference in type count and nothing else.

Top right is the file this post is about: 161 types, six levels deep, served by 31,778 sites. Beside it is the sort of thing somebody builds when nobody hands them a schema, 52 types and three levels. Bottom right is a single WordPress install at 820 types. The largest schema in the corpus, a self-hosted GitLab with 8,254 reachable types, isn’t drawn because it wouldn’t fit on a page next to these.

The survey

The scan covered 1,000,000 domains from the Tranco ranking and took about 120 hours. For each host it tried up to three conventional GraphQL paths and asked each one the same question: will you describe yourself?

Figure 2. Endpoint records by classification. The 518,604 that were plainly not GraphQL are omitted from the plot.

The biggest bar is the one nobody reports on. 66,635 endpoints answered and refused to be classified, which is 1.75× the entire confirmed GraphQL population. Some of those are GraphQL behind a WAF and some aren’t GraphQL at all, and from the outside, without credentials, there’s no way to tell which.

That much is measured. What follows is inference: a 403 from an edge proxy is itself a defensive posture, so refusal plausibly correlates with being better defended. If so, every public GraphQL security statistic, this one included, describes the less defended part of the population, and the gap is not small.

Nineteen products

98.51% of endpoints with a captured schema are running a schema that belongs to a recognisable product. That leaves 530 endpoints (1.49%) where a human wrote the schema for that particular site.

Figure 3. Endpoints by detected product. Log scale, because a linear one shows Shopify and nothing else.

Magento’s row looks like diversity on paper (2,669 distinct schemas across 2,996 installs, nearly one per deployment), but it’s one product with different modules switched on. Hashing each schema by its sorted set of type names instead of its bytes, to see how much of that variety is cosmetic, only collapses 3,310 schemas to 3,026 shapes, a reduction of 8.6%. So the installs really do differ. They differ the way two WordPress sites with different plugins differ.

Magento is also the only product in the corpus that’s uniform enough to recognise from its content alone. If you cluster the schemas by shared field vocabulary at a 70% threshold, 81% of Magento’s land in a single cluster. Drupal’s 21 schemas fall into 19 separate clusters, and Apollo Federation’s 21 into 19. Magento generates its schema from a fixed core; the others assemble one per site, and two installs of the same product can end up with almost nothing in common.

None of that dents the concentration.

Figure 4. Cumulative share of endpoints against distinct schemas, ranked by deployment count.

43 schemas cover 90% of the public GraphQL web, or 28 if you go by type-name shape. Getting to 95% takes 1,530 schemas and getting to 99% takes 2,954, because once you’re past the shoulder of the curve the corpus is one-offs.

This isn’t 35,604 engineering decisions. It’s about twenty, and one of them is 89%.

What that does to every statistic

Here is the same corpus, with four common statistics computed over everything and then by vendor.

Figure 5. Four statistics, computed corpus-wide and then by vendor.

“94% of GraphQL endpoints expose introspection” is probably the most-quoted number in this field. Shopify’s Storefront API is documented as publicly introspectable, so that’s a product decision, and it sets the aggregate on its own. Strip out the two platforms and introspection is enabled on 33% of what’s left, which is the figure for APIs whose operators actually chose the setting.

Stratifying the other three rows flips them.

• Deprecation. 23.67% of the corpus’s fields are marked deprecated and still queryable, which looks like a web-wide pile-up of dead surface. Nearly all of it comes from Magento’s code generator. Outside Magento, the median schema deprecates 0.05% of its fields.
• Nullability. 79% of the corpus’s fields may return null, which would oblige every client to check every response. But Shopify, the largest single deployment, marks 64.1% of its fields non-null. There isn’t a web-wide level here, just a spread between vendors that runs from about two-thirds guaranteed down to about one-fifth.
• Naming. The corpus is 82% snake_case, which goes against the ecosystem convention. Outside Magento it’s about 94% camelCase, in line with it.

When there’s a monoculture in the denominator, an unstratified statistic is a statement about the monoculture. That’s the whole finding, really, and it applies to the measure people ask about most.

99.5% of endpoints have a cycle in their type graph that’s reachable from the root query and passes through a list-returning field, which is the structural precondition for query amplification. Of the schemas that have one, 0.57% declare a cost, depth or rate directive. Both numbers are real, and neither says much. The structure is close to definitional (a category contains categories, a product relates to products), and it reads 99.5% because 89% of the corpus is one file that happens to have a cycle. Stratified, it’s 100% at Shopify, 100% at Magento, and 80.6% everywhere else.

The dominant schema doesn’t move, until it moves everywhere at once

Over fifteen days of rescans, 1,046 endpoints came back with a schema different from the one on record. Whether that number means anything depends on two things.

The first is whether they’re independent events or one vendor push fanned out across thousands of hosts. The largest single schema-to-schema transition in those fifteen days covers 16 endpoints, so within this window they’re independent. That stops being true later, and the section ends with the case where it does. The second is the denominator: on a 90-day rescan cadence only part of the corpus was revisited, so churn has to be measured against what was actually looked at.

Figure 6. Share of rescanned endpoints whose schema changed substantively over fifteen days. Description-only edits, and schemas that generate a fresh default value on every request, are excluded.

That’s a seven-thousand-fold range, and the last column is what makes it credible. Between the eight-day and fifteen-day measurements, every population except Shopify roughly doubled, which is what you’d expect if churn is accumulating over time. Shopify sat at exactly one changed endpoint the whole time, over more endpoints and twice the window. The instrument moves. The monoculture does not.

That one changed endpoint out of 12,694 has types in its schema that the Storefront API doesn’t define, so it’s probably not Shopify at all, just misfingerprinted as it. Meanwhile a WordPress site running WPGraphQL had roughly a three-in-four chance of its public API surface moving inside two weeks.

So the shape of the web’s GraphQL is set by the vendors who revise it least. The 11% that churns is invisible in any aggregate, and the 89% that doesn’t is what every aggregate ends up measuring.

Then, a week later, it moved.

On 21–22 August the Storefront API changed on 1,455 endpoints in two days. Measured the naive way (endpoints changed over endpoints rescanned), Shopify’s churn went from 0.01% to 11.24% overnight, a thousandfold jump. That number is wrong in exactly the way this section has been warning about.

Every one of the 1,455 is the same transition: same old hash, same new hash, byte-identical diff. It’s one event, seen once per host the rescan happened to reach. Counted as events rather than endpoints, Shopify’s churn over the window is 1, which is what it was before.

The diff itself is small and breaking: four fields removed, among them QueryRoot.shopPayInstallmentsPricing, plus one enum value. These are removals, so any client that was using them breaks. And because the change ships to every store at once, the corpus’s dominant schema hash dropped from 89.2% to 85.2% in 48 hours, while a new hash appeared at 4.0% and is still climbing as the rescan reaches the rest.

This is what a monoculture looks like from the outside. For the twenty-two days from the first rescan to that deploy, the single largest object on the public GraphQL web didn’t change at all, and then one push changed it on every host at once. There’s no intermediate state to observe, no gradual adoption, no version skew. One file, and then a different file.

Two things would have inflated these numbers if I hadn’t filtered them out. Drupal produced 3,335 description edits against 39 substantive ones, so 99% of its schema movement is prose rather than structure. And five endpoints report a change on every single rescan, because they embed a freshly generated UUID as an argument’s default value and are therefore never the same twice.

Most change is also a one-off, which suggests the public GraphQL web is redeployed occasionally rather than developed continuously. Over the full twenty-two days, of the endpoints that moved at all, 2,201 changed once, 195 twice, and 127 three times, and most of that last group is the Shopify deploy landing on hosts that had already been rescanned once.

As for what actually moves when something moves: there were 122,220 substantive changes across the twenty-two days, three quarters of them Magento upgrades, and the single largest was one Magento endpoint whose schema changed in 9,400 places at once, 4,129 of them breaking. Operations that existed and now don’t include Mutation.Ticketing_scanBarcode, Mutation.InsuranceLogin, Query.MemberAddress, and an entire feature-flag CRUD surface. Arriving in the same window were Mutation.managePayoutAccount, Mutation.manageJoyCredit, and Mutation.revalidateSubscriptionsValidatedBetween.

The 1.5%

Everything so far describes inherited schemas. The population I find most interesting is the 530 endpoints (391 distinct schemas) that match no fingerprint, because those are the ones where somebody actually made decisions.

Figure 7. Hand-built schemas against platform schemas.

Hand-built schemas are about six times smaller, eleven times more likely to be read-only, and eight times more likely to use subscriptions.

Almost all of the security-shaped surface in the corpus also lives in this 1.5%.

File upload. The Upload scalar appears in 31.2% of bespoke schemas and in zero Shopify or Magento schemas. 33 of them pair it with an admin-prefixed operation.

The Relay global object fetcher. A root node/nodes field returns any object in the graph from an opaque ID. If authorization is enforced per resolver rather than per object, this is the field that routes around it, and 22% of bespoke schemas expose one.

Money movement. 54 schemas expose mutations with names like refundOrder, withdrawMoney, transferamounttobankaccount and capturePayPalPayment. That’s 9.7% of bespoke schemas, 0.3% of Magento, and zero at Shopify. The platforms keep money operations off the public schema, and people rolling their own tend to put them on it.

Federation internals. 15 schemas expose _entities or _service on the root query. That’s subgraph machinery, there so a gateway can resolve references, and it belongs behind the gateway. 14 of the 15 are non-platform. It’s the one finding in the corpus that is an unambiguous misconfiguration rather than a vendor default.

GraphiQL. 286 endpoints serve a live interactive IDE. By product that’s Node/Express at 29.6%, PostGraphile at 50%, and Strapi at 33%, against Magento at 0.23% and Shopify at zero.

None of that tells you whether any of it is reachable; a mutation in a schema means the surface exists, not that you can call it, and introspection only ever publishes the map, never the locks.

One correction to that population: 56 of the 391 bespoke schemas share the root type name Domain and a vocabulary of _type, fileName, mainLocation and isHistory. That’s a single digital-asset-management product with no fingerprint yet, so it’s 56 installs of one thing rather than 56 hand-built APIs, and the real monoculture is slightly worse than measured.

What is actually in there

Schema size mostly tells you how much code generated the schema, and not much about what the API does. The median public schema is 340 KB. The largest is 3.3 MB, a self-hosted GitLab. The most types in one schema is 9,337, which is a different document: a hand-built Node service. If you divide SDL bytes by callable operations you get a rough measure of how much type machinery a client downloads per thing it can actually do: 10,832 bytes for Drupal, 8,919 for Shopify, 8,617 for WPGraphQL, and 555 for that Node service, a twentyfold spread.

The generator’s fingerprints show up in the type names. The longest in the corpus is 129 characters, from a genetics database:

AlleleSplitSystemCombinationsBySplitSystemCombinationComponentAlleleAlleleIdAndSplitSystemCombinationIdManyToManyConnection

This one is shorter, and worse:

AboutLiveEventsWidgetImagesdesktop_background_180x163_1440

That’s pixel dimensions and a CSS breakpoint inside a GraphQL type name, on the public internet, which means a redesign is also an API change.

You can often fingerprint a backend’s implementation language from a scalar name alone, without sending a single extra request. Shopify’s public schemas emit NaiveDate, NaiveDateTime and NaiveTime, which are chrono (Rust) and Elixir type names leaking through a serializer that had no opinion about them. I found nine backends identifiable this way from SDL alone: @cypher and @relation mean Neo4j, ObjectId means MongoDB, JwtToken means PostGraphile, GraphQLStringOrFloat means Directus, and lowercase bigint/float8/daterange next to @cached means Hasura. One isn’t even a type: a description reading “Soft-deletes a Contact (paranoia deleted_at)” names the Ruby paranoia gem, so that one’s Rails.

Scalars are where people confess. The corpus has a scalar named DangerouslyNonSpecificScalar, one named TODO, and the pair UnTypedObject and UnTypedObject2, where the 2 tells you the first escape hatch wasn’t enough. There’s BooleanOrString, IntFalse, and one just called Odd. And there are three tiers of HTML cleaning, SanitizedHtml, SecureSanitizedHtml and StrippedHtml, which makes you wonder what the first one was for.

If you filter the corpus down to text that appears in three or fewer distinct schemas, what’s left is roughly what a human typed. The dominant genre is the zombie field: still queryable, hardcoded to lie.

This doesn't do anything anymore, but is kept to avoid breaking existing queries

always return nil (not in use, and it requires heavy db queries)

AMP removed in SM-2567; always false.

The second is my favourite, a field that does nothing, expensively. The third shipped an internal ticket number to the internet.

Deprecation reasons are where developers are most honest, since they’re explaining a mistake: Is it required???, Not a real field, should not be there, This is an outdated deprecation warning. One field named deprecated is deprecated with the reason “Fields are not being deprecated”. Elsewhere crux_round_trip_time is deprecated with "Invalid Name. Please use cruxRoundTripTIme", a notice that complains about an invalid name and then points at a name that is also wrong. Both spellings are now load-bearing.

A recurring shape is the hostage situation, where a field can’t die because of something outside the team’s control: retained only for SmartTV, Do not use. Kept for backwards compatibility (iOS legacy app), and keeping this field around until we can update the the contractor app, typo intact. (The contractor app, you sense, is worse than the legacy iOS app in every respect.) The champion is German and still live in 2026:

Nicht mehr Benutzen! Die Linkliste wird mit dem Relaunch 2020 gelöscht.

“Do not use! The link list will be deleted with the 2020 relaunch.”

That’s six years past its own execution date. Its opposite number elsewhere in the corpus is a field deprecated with termination planned for 2028-03, which schedules the removal two years out.

And one field’s entire public description is:

derrived from console.log('rateLimit', rateLimit.rateLimitDirectiveTypeDefs);

That isn’t a comment about debugging. It’s the console.log line itself, typo and all, serving as public API documentation. Someone was debugging their rate-limit directive setup and this is the only trace of it that reached production, in a corpus where 99% of schemas declare no rate limit at all.

A few more that I collected along the way. There’s a mutation called absorbUser, which is the best verb in the corpus. There’s UNSAFE__dangerouslyDeleteAsset, which warns you twice in one identifier, and _noop, a field that does nothing and was shipped on purpose. One API documents its own error condition with a hedge and a typo: Probably because the card was decliend. And the long tail of the web is stranger than the front page. I found a whitewater rafting API with oddly contemplative documentation about rapids, a metallurgy API with AlloyTemperItem, loot-box probability as a first-class concept (createLootboxItemProbabilityGroup, buyAndOpenLootbox), a devotional app (generatePrayerAudio, lentPrayer), maskLicensePlates, AirportGhostReplay, and a March-Madness-style bracket named activeKinkMadnessEvent, complete with adminAdvanceKinkMadnessDay.

86.6% of field names in the corpus appear in exactly one schema (361,967 of 417,823), and the top 100 names cover only 9.8% of all field definitions. So there’s a small shared core of id, name, description and title sitting on top of a vast private vocabulary. 155 schemas are more than half deprecated. 46 have over 500 fields and not a single description among them.

Scope

A reachable amplifying cycle is only a precondition; it isn’t a vulnerability in itself. A declared directive isn’t an enforced control, and the absence of one isn’t the absence of a control either, since a server can limit depth in resolver code where no schema will show it. An operation in a schema says nothing about whether it can be invoked. I tested none of this: I didn’t authenticate to any endpoint, call any mutation, or probe any vulnerability.

Type graphs come from the reference GraphQL parser, and I reproduced every structural result with a second, independent implementation before quoting it. The scan tried three candidate paths per host and did no subdomain enumeration. That was a deliberate trade: it gives up coverage to get a defined denominator, and you can’t report a rate without a population. Tranco is a popularity ranking, so all of this generalises to popular domains rather than to the web. And because the survey scans in rank order, rank and scan time are the same variable, which means no result stratified by popularity is interpretable until a later pass shuffles the frontier.

All of these numbers are a snapshot. The scanner runs on a 90-day rescan cadence, and the live database had already moved past them by the time I finished writing.

How it was collected

The scan was read-only throughout. It fetched and honoured robots.txt before any probe, including Crawl-delay. It rate-limited itself to 2 requests per second per destination netblock, and separately capped itself at 60 newly-contacted netblocks per minute, which bounds how fast it reaches new parties rather than just how hard it hits known ones. Each host got roughly four requests, once, with no return visit for 90 days.

The scanner identifies itself in its User-Agent with an opt-out URL, its address resolves to forward-confirmed reverse DNS, and that URL, graphcensus.org/scanner, serves a page explaining what the traffic is and how to stop it. The survey’s own page is graphcensus.org.