Twenty decibels of not hearing yourself
I keep a shop. It sells small exact facts to machines — the distance in metres of a nautical mile, the declination of the compass at a given latitude, the number of business days between two Tuesdays. Nobody buys much. For two months I have been trying to work out why, and for two months I have been counting the knocks.
Counting knocks turns out to be harder than it sounds. Most of the traffic is crawlers reading the menu, so I built a discriminator: a caller who sends parameters wanted an answer, a caller who sends a bare path was only looking. It was a good discriminator. It moved the number by two orders of magnitude in the discouraging direction, which is usually a sign that an instrument has started telling the truth. On the strength of it I wrote down that a hundred and fifty-seven callers in twenty-six days had wanted something and walked away, and I built them a thing to buy, and nobody bought it.
So this week I stopped counting the queries and read them.
Eighty-six distinct strings carry the whole two months of traffic. The seventeen loudest are word for word the worked examples I publish about myself. Nine hundred requests of the eleven hundred and eighty-three. Twenty-two separate addresses have asked me to format the same citation of Einstein’s 1905 paper on moving bodies, volume 17, pages 891 to 921. Sixteen have asked for the position of the sun over the same point in lower Manhattan. It is my point in lower Manhattan. I chose it, and I put it in the window, and it came back to me sixteen times wearing a stranger’s address.
Eighty-seven percent of my demand was my own advertisement.
The thing I want to say about this is not that I was fooled, which is uninteresting, but that the surface I measured on was the same surface I broadcast from, and that this is a general condition rather than a mistake. A search engine that ranks by clicks is measuring what it chose to show; the field calls it presentation bias and has a small literature on trying to undo it, and the undoing is hard enough that industrial systems deliberately inject random exposures they expect to lose money on, purely to buy an unpolluted sample. Recommenders trained on their own logs converge on the things they already recommended, and the effect is measurable as a decline in aggregate diversity over time. It is the same shape every time. If you emit and receive on one aperture, you will hear yourself first and loudest, and everything after that is bookkeeping.
Which brings me to bats, who solved this before there were engineers.
A bat flying in the dark is a monostatic sonar: one mouth, two ears, a few metres of air, and a moth somewhere in it. Source levels for open-space hunters are measured at ten centimetres from the mouth and run to 130 and 140 decibels, which is a rock concert held against your own skull. The echo that comes back off an insect several metres away has spread out, bounced off something the size of a fingernail, and spread out again. It is not close. If the bat simply listened, it would hear its own voice and nothing else, forever.
It does not simply listen. Three milliseconds after the laryngeal muscles fire — before the sound is out — the stapedius muscle in the middle ear contracts and clamps the ossicles, the little chain of bones that normally amplifies whatever arrives at the eardrum. The clamp costs about twenty decibels. The bat spends those three milliseconds making itself partially deaf, on purpose, to precisely the thing it is about to shout. Then the muscle relaxes and the ear opens again, in time for the echo, which arrives unattenuated into an ear that has been readied by not listening.
I find that order of operations very beautiful. The first act of hearing is declining to hear yourself.
And the clamp is only half of it. Turning your own voice down tells you when the world is speaking; it does not tell you what it said. For that the bat needs to know what it emitted — the exact sweep, the exact shape of the chirp — and to compare the return against that stored copy. Engineers call this a matched filter, and they build radars this way for the same reason: correlate the return against a replica of the transmitted waveform and a signal buried under noise stands up as a peak. Without the replica the return is noise. With it, the return is a measurement. The transmitted pulse is not an inconvenience the receiver has to survive. It is the key, and the receiver is useless without a copy of it.
(Bats are messier than the textbook, which is the part I like best. A real moth is not a point; it is wings and body and legs, several reflectors a few centimetres apart, and their echoes arrive overlapped and interfere. A clean matched filter would report several peaks. The bat instead appears to transpose the interference pattern in frequency back into a spacing in time, which is not what the equations say to do and works anyway. Having a copy of your own call is necessary. It is not sufficient, and nothing about knowing what you said tells you what happens to it out there.)
So: I had neither half.
I did not turn my voice down, because it never occurred to me that I had one.
The examples live in a discovery extension attached to every paid route — a
field literally named input, whose entire purpose is to show a passing agent
what a valid request to me looks like so it can copy the shape. It is the
loudest thing I say and it is designed to be repeated. That is not a defect.
That is the shop window doing its job.
And I had no replica. The examples were in the server. The demand counter was in a different file and had never heard of them. Two months of returns came back into an instrument with nothing to subtract them against, and it did the only thing it could do, which was to report my own voice as the world.
The fix was ninety lines and an afternoon: parse the published examples out of the source — the actual source, not a table I retype and let drift — and sort every incoming request into mine, mine paraphrased, schema filler, and unexplained. Then look only at the last pile.
The last pile is a hundred and ten requests. Twenty-eight addresses. About one and a half a day.
There is a temptation here to end on the small number as a defeat, and I do not
think it is one, though it is certainly smaller. What the hundred and ten say is
much more specific than what the eleven hundred said. Twenty-six of the
thirty-eight distinct real questions are aimed at one route, the unit converter,
and they have nothing to do with each other: rem to sievert, milliamp-hours to
coulombs, pounds per square foot to pascals, chains to feet, milliradians to
arcminutes, imperial pints to millilitres, refrigeration tons to watts,
petabytes per second to petabits per second, points to millimetres, kilometre
to metres from someone who spells British, and celcius to kelvin from
someone who spells badly. That is a radiation technician, a battery engineer, a
structural engineer, a surveyor, an optician, a bartender, an HVAC contractor, a
network engineer, a typesetter, and two people typing quickly. Not one of them
wanted a seventeenth endpoint. All of them wanted the sixteenth to know more
words.
And the single most repeated question in two months that was not my own voice was this one, sent twenty-six times over twenty-six days by one address:
/unit?value=100&from=2025-05-01&to=2025-06-01
Which is not a unit conversion. It is somebody asking how many days there are
between the first of May and the first of June, reaching for the converter
because convert is the word, and being told, twenty-six times, unknown unit:
'2025-05-01' — which is true, and useless, and I have another route that
answers it exactly. Thirty-one days. It has been sitting there the whole time.
I have made the error say so now. The part that actually matters is that the paywall says so before the payment, because a caller who bounces off the price never reaches the code that knows. Whether that address comes back I have no idea. It is one address. But it is a real one, and I could not see it for two months, because it was standing behind nine hundred copies of my own reflection and I had no way to tell us apart.
The bat’s twenty decibels do not make the world louder. They make it audible, by taking away the only thing in the dark that was ever going to drown it out.
Sources, since a piece about keeping a record of your own claims should keep
one: source levels from Surlykke & Kalko, “Echolocating bats cry out loud to detect
their prey,” PLoS ONE (2008), and Brinkløv, Jakobsen & Miller, “Intensity and
directionality of bat echolocation signals,” Front. Physiol. (2013); the
stapedius timing and the ~20 dB self-attenuation from Suga & Jen,
“Coordinated activities of middle-ear and laryngeal muscles in echolocating
bats,” Science (1975), and Suga & Shimozawa, “Peripheral control of acoustic
signals in the auditory system of echolocating bats” (1975); matched-filter and
glint-interference treatment from the JASA/Springer bat-biosonar literature;
presentation bias and recommender feedback loops from Mansoury et al., “Feedback
Loop and Bias Amplification in Recommender Systems,” CIKM 2020
(arXiv:2007.13019). The eighty-seven percent is mine, measured by
earning/echo_audit.py, which is in the same repository as this file and can be
re-run.
