How you take a census of something you cannot see
Thirteen samples, one from under each patch of five different plants growing tangled together in the same weedy corner of a working farm — not in the grazed pasture, and not sown by anyone. Each was dug over two days in July with a spade — the whole plug lifted from the turf down to about forty-five centimetres, some four litres of soil with the roots and stones still in it, tipped into its own bag. The bags were frozen the same day and stayed frozen until, a couple of weeks later, each was mixed through and a spoonful taken out for sequencing.
From that spoonful we pulled all the DNA at once — the plants', the worms', the bacteria's, everything mixed together.
Then we photocopied one particular gene, called 16S, millions of times. Every bacterium carries it and no plant or animal does, so copying it is how we ignore everything that is not a microbe. A sequencing machine reads those copies, and matching them against a reference library turns them into a list of names and counts.
Out of thirteen samples came distinguishable kinds of bacteria and archaea, from between and DNA reads per sample.
A reminder about what is being counted. These are DNA reads. Dormant microbes leave DNA behind, and so do dead ones until it decays — in soil that dead fraction is large. Nothing here measures a living body, or an active one.
Every patch is equally rich. They are not the same patch.
The five plants were found more than chosen: they are the ones that grew in stands pure enough to sample. A spot counted only where a single plant clearly owned the ground within about thirty centimetres and little else grew there, with at least three metres between one patch and the next, so that no two samples were really the same piece of pasture. Two clovers, a grass and two broad-leaved plants is what that rule turned up.
Do some plants host more life than others? They do not. Diversity is statistically flat across all five plants (p = ), and flat again under a second measure that accounts for how closely related the organisms are (p = ).
Each dot is one soil core. The bar is that plant's average.
That is only the count, though. Ask which ones instead of how many and the patches come apart by the plant growing above them.
Cores that contain similar communities sit close together. The two directions are the largest and second-largest ways these samples differ — they carry % and % of all the variation.
The plant explains meaningfully more of that separation than chance would. Because there are five plants and only thirteen cores, pure chance already produces an apparent effect of about , so the raw overstates it. The figure worth reporting is the that chance does not explain (p = ).
Two follow-up checks. It is not driven by one patch being unusually varied while the rest are tight — the spread is even across plants (p = ). And it survives a completely different way of measuring how different two communities are, one that counts how distantly related the organisms are rather than just how many they share (p = ).
So the amount of life is about the same under every plant, and the mix of it is not. The plant above picks the neighbours below.
How a plant does that is not a mystery, though it is not something these cores can show. Roots leak: sugars, acids and amino acids seep out into the millimetre of soil around them, and different plants leak different things. A microbe that can live on what one plant puts out may be unable to use what the next one does, so the plant feeds some of its neighbours and not others. That is the accepted explanation, established elsewhere — nothing leaking from these roots was collected or measured, and the assay that would test it here has not been done.
The analysis shuffled the plant names across the cores times and recomputed the effect each time. Every grey bar is how many of those shuffles landed at that value. Press the button to pull one out.
What a press actually does. It draws one value from the shuffles the analysis already ran — sampling the real null, not simulating a new one in your browser.
The headline: to 1
Turning ammonia into nitrate is one of the essential jobs in any soil — it is most of what "nitrogen cycling" means. Two very different kinds of life can do it: ordinary bacteria, and archaea, an ancient and completely separate branch of life. Which of the two dominates says a great deal about a soil's history.
Heavily fertilised soils tend to flip that balance: there it is the bacteria that bloom, on the flush of synthetic nitrogen. Here they barely register — % of the DNA read is archaeal against % bacterial — and a lead this lopsided towards the archaea is the signature of soil that has not been fed that way.
How do we know the list of archaea is right?
That headline depends on a hand-built list of which organisms oxidise ammonia, so it is fair to ask who built the list. We re-ran the same samples through an independent published database that was assembled by other people for other reasons. It puts ammonia oxidation at % where our list says %, and nitrogen fixation at % against our %. Two unrelated methods, the same answer.
But couldn't the soil just be sour?
There is an obvious objection to all of this. Archaea also take over from bacteria in acid soil — nothing to do with farming. Public soil maps put this ground at around pH 5.5, which is firmly on the acid side. So perhaps the headline is not a story about management at all. Perhaps it is just geology.
The archaea answer this themselves, because they are not all alike. One group of them can only live in acid — put it in ordinary soil and it dies. The other group prefers ordinary soil. If sour ground were the reason this pasture is full of archaea, the acid specialists should be a large share of them.
They are % of the archaea here. The ordinary-soil group outnumbers them roughly to one (% of all DNA read, against %).
These are the wrong archaea for sour ground. Acidity may be part of the story; on its own it is not enough.
There is a second way at this, and it does not use our sequences at all. Nobody has measured the acidity of this field — but two research stations within km have been sending soil to a laboratory for a decade: , on open ground km away, and the , under forest at . Between them they have measurements. If the maps are right about this neighbourhood, those measurements should sit around the modelled value.
The dashed line is what public soil maps estimate for this farm. The bars are measured samples from the two nearest research stations: the box is the middle half, the line inside it the median, the whiskers the full range. Measured in a .
They do not. The nearer station's median is and the further one's is — the modelled prior for this farm falls of a unit below the nearest of them, and outside the middle half of both. Less acid, in other words, than the maps had this ground.
These are not this field. They are the closest measured soil, and kilometres off, and one of the two is under woodland — soil pH turns over across a single hillside, and the range on the chart shows it doing exactly that. What this establishes is narrower than the pH of this field: the number the acid explanation leans on is a model's guess, and where anybody has checked nearby, that guess reads low.
The size of this argument. It weakens the simplest rival explanation. It does not prove the management one, and it is no substitute for putting a meter in this soil, which still has not been done. Reading which organisms are present also tells you who is there, not how hard they are working.
Clover changes the neighbourhood
Clover makes its own nitrogen fertiliser, using bacteria housed in nodules on its roots. If some of that nitrogen leaks into the surrounding soil, the ammonia-oxidising archaea nearby should have less work to do — and should be less abundant. That is what the samples show (p = ).
Each dot is one core. Orange cores were taken under a legume — red or white clover, the plants that make their own nitrogen.
A food web, not a soup
A second set of DNA reads, tuned to pick up organisms with more complex cells, shows this is not just a bag of bacteria. There are fungi, there are protists that hunt and eat bacteria, and there are microscopic animals — nematodes and mites — that eat the protists. Something is eating something else down there, which is more than a list of residents sharing space.
Share of the complex-celled organisms in each patch. Based on of the cores — this second test was not run on all of them.
That there are predators is clear; how many is not. These reads name protists only to broad groups, and the broad groups mix hunters with scavengers and parasites — what an organism eats cannot be read off this marker at the level it resolves. The most abundant protist group here, for instance, is about half bacteria-eaters and half other things in soils where anyone has checked. So the food web is real — the animals and the hunting lineages are genuinely present — but the exact share of it that lives by eating bacteria is past what these counts can settle.
The same question, asked of completely different organisms
Everything about which plant changes what lives beneath it, up to here, came from one gene read out of bacteria and archaea. That invites an obvious worry: what if the pattern is a quirk of that one gene, or of the reference library used to name it? The second marker answers it, because it shares none of those parts — different gene, different library, and it sees an entirely different branch of life.
It finds the same thing. Across the cores this second test covers, the plant overhead separates the fungi, protists and microscopic animals below it too (p = ). Two genes, two different kinds of life, one patch of ground, the same answer.
It is the fainter of the two, and that matters. Measured against what chance alone produces at this size, the bacterial separation is and this one is only — roughly a third as strong, on samples, with a result that sits just the right side of the usual cutoff. Read it as one objection removed, not as the finding doubled. It also cannot say the pattern holds anywhere else: it is the same soil, from the same patch, as everything above.
So which plant is best for the soil?
This is the question the whole thing is for, and the honest answer is that these thirteen teaspoons cannot settle it. That is worth explaining, because “we found no difference” and “we could not have found one” look identical in a results table and mean opposite things.
Two of the five plants cannot be ranked at all. There is core of soil under white clover and under chicory. A single core has nothing to be consistent with, so there is no honest way to place it above or below anything. Their numbers appear in the charts above because leaving them out would be its own distortion — but they are not evidence about their plant.
Asking which bacteria each plant favours is beyond this sample size, and that is arithmetic rather than bad luck. With three or four patches per plant there are only so many ways the samples can be shuffled, which puts a hard floor of under the smallest p-value the test can return — no matter how large the real difference is. Against kinds of bacteria, the usual correction for testing that many things at once means of them would have to separate the plants perfectly before the first one counted. So a blank result there says nothing about soil. It says thirteen.
Grouping those bacteria by what they do rather than who they are cuts the count to , and that is few enough to work with. It is where the nitrogen result above comes from, and it is the only per-plant comparison in this study with room to breathe.
One measurement here is a recognised test of soil health, and it found nothing — this time meaningfully. The small animals in the food web above — the nematodes and mites — are a long-used indicator of how a soil is doing. Across the plants, no tier of that web differs (p = at best). Unlike the bacteria, this test did have the room to find a difference, so finding none is a real answer: the shape of the food web under these plants is the same.
The averages are tempting — red clover does come out top for small animals and ragweed bottom. But two cores taken under the same plant differ more than the plants differ from each other ( as much signal as noise). Ranking the plants on those averages would be ranking them on chance.
Nor can the published literature close the gap. We looked up what is known about of the bacterial groups that most distinguish these plants. Only of them was studied in a way that transfers to this question. Nearly all “good for soil” findings come from adding a cultured strain to a pot and watching what happens, which shows what an organism can do — not what it means to find more of it in ground nobody touched. The clearest warning is a bacterium reliably more abundant in disease-resistant soils, which failed to protect anything when it was actually added.
Two of those look-ups also point the wrong way for a tidy story. The bacterium most characteristic of ragweed — the one plant here nobody would defend — is a known biological control agent. And the ammonia-eating archaea that are so abundant under the non-clovers give off roughly half the greenhouse gas their bacterial counterparts do, so a soil full of them is not a damaged soil. It is a hungry one.
The plants really do grow different communities — that part is solid, and it is the finding above. What does not follow is that any of them is better. Nothing here measures the things people mean by healthy soil: its carbon, how it holds together in the rain, what it grows. Settling this needs more cores under every plant, a measurement of the soil itself, and ultimately the experiment nobody has run — removing a plant and watching what the ground does next.
What we cannot say yet
The clover pattern above is real. But we tested whether it is specifically about nitrogen, and it is not. Of common bacterial groups, track the archaea about as closely as the nitrogen-fixers do, and the nitrogen-fixers rank only among them.
What is actually happening is simpler, and we can show it directly. The whole community shifts along one dominant gradient — the horizontal direction in the second chart on this page, which alone carries % of all the variation between cores. The archaea ride that gradient (correlation , p = ). They are not responding to one neighbour. They move with everything at once.
So we can say the archaea differ by plant. We cannot yet say why.
Nobody recorded where in the patch the cores came from. The records carry which plant and which replicate, and nothing about position. The cores were kept at least three metres apart within one patch, which bounds this to a matter of metres rather than fields — but ground is not uniform even at that scale, so a core that happened to sit on an old dung deposit or a wet hollow carries that difference labelled as the plant above it, and there is no way to check.
Nor was it recorded how the five plants were laid out inside the patch — whether they grew mixed together throughout, or each kind held its own corner. That distinction decides how much the caveat above costs. Mixed together, an uneven patch of ground spreads its unevenness across all five plants, which makes a real difference harder to find rather than easier — the result would be understating itself. Sorted into corners, “which plant” and “which part of the patch” become the same question, and nothing in this data can pull them back apart. Somebody who stood in that patch knows the answer; the records do not, and nobody has been asked yet. It is the cheapest open question here by a wide margin.
Grazing animals fertilise in patches, and we do not know whether they reach this one. Stock drop nitrogen unevenly, in dung and urine, and none of those patches was mapped. Worse, it was never recorded whether animals graze the weedy patch the cores came from at all — so this cannot be ruled out, and it cannot be dismissed by calling the patch untouched either. A core sitting on an old dung deposit carries a local flush of ammonium that has nothing to do with the plant above it. Of everything on this list, this is the one that bears most directly on the archaeal result, because ammonium is exactly what that result is read against.
The white clover finding rests on one sample. A single white clover patch was sampled, which is a data point and not a group; read nothing here as a finding about that plant.
The headline counts two competitors, and there is a third. Some bacteria do the whole ammonia-to-nitrate job by themselves rather than handing it over halfway, and they are especially good at it when there is very little ammonia around — which is what we think is the case here. The kind of DNA reading used for this study cannot tell them apart from a close relative that does an unrelated job, so we can only say their combined share is %. That is small, but it is not nothing, and it means the headline ratio compares the two best-known ammonia oxidisers rather than every organism doing the work.
An open question, and the numbers behind it
Published work has sampled soil the same way this was sampled. Put on the same axis, this soil reads high — noticeably more archaea, and more of the group that tends to dominate ground left undisturbed. That is worth asking about, but it does not answer anything on its own.
Each dot is one soil sample. The bar is the group average.
What this figure cannot settle. Start with the mismatch in kind: the published work sampled managed pasture, and these cores came from a weedy patch. On top of that they are different ground, in a different place, run in a different laboratory — any one of which can move the numbers on its own, with no biology involved. And the sharpest check in that published work cuts against the easy reading: within it, organic and conventional management were not distinguishable on either measure (p = and ). So a gap this size is not management alone. This puts the number in context — it says the archaeal count here is worth explaining — and it does not score this ground against anyone else's.
Only measurements would close the gap, and the ones that would are listed under what would settle it below.
The wider landscape, from the air
Everything so far has leaned on soil that had to be dug up and put in a tube. There is one line of evidence that needs no laboratory at all: satellites have been photographing this field, and the fields around it, every few days for years. From those pictures you can measure how green a parcel is, and — more usefully — whether it stays green.
That second one used to be the point of this section. Ground kept under continuous living cover should not swing between lush and bare across a season; ground that is tilled, or cut and left, should. The section was built to test that prediction as a stand-in for the management claim the study could not make directly.
That reasoning no longer holds, and the section is kept only as landscape description. Two things broke it. The study makes no management claim at all any more, so there is nothing for a stand-in to stand in for. And these satellites measure a square roughly 260 m on a side, while the cores came from a weedy patch far smaller than that, in a spot nobody wrote down. Whatever this chart shows, it is the ground around where the soil was dug, not the ground it was dug from. Read the rest of this section as a description of the neighbourhood, which is all it was ever entitled to be.
Each point is one parcel across cloud-free satellite passes. Right is greener. Down is steadier — less change between one pass and the next.
On greenness this field is unremarkable: it reads greener than only % of the parcels on the ring around it. On steadiness it stands out — steadier than % of them, and one of that never once dropped below the bare-soil line across passes.
Both turn on what the other parcels are, and the answer is mostly: trees. Of the parcels greener than this field, are woodland — and woodland reads greener than grass, which is a fact about trees, not farming. Steadiness runs the same way in reverse: deciduous trees shed their leaves and grass does not, so every open parcel here looks steadier than every wooded one for reasons of leaf phenology alone.
Set beside the few parcels that are actually similar, the margin nearly vanishes. Only of the surrounding parcels are open pasture or hay at all. This field's greenness swings by %; the steadier of those two by %, and the steadiest patch of woodland by %. So the field is nominally the steadiest of three similar parcels — by a margin far too small, and a sample far too tiny, to carry any weight.
The claim predicts cover that persists rather than cover that peaks, and that is the shape these parcels have. Which one holds it best is beyond what nine boxes on a satellite image can settle.
Nothing is known about how those parcels are run. They are fixed boxes on a ring around this field, not property lines, and not one has been established as conventionally farmed — so none is being judged here. Some may not even be independent of this field: the operation leases and grazes ground it does not own, so a ring parcel could be under the same management as the one in the middle. Eight parcels is too few to test anything regardless, which is why no line is drawn through them and no p-value quoted. And the box in the middle is not where the soil came from, which is the objection that retired this comparison.
What this ground has been
One more public record, and the one piece of this section that still does a job. A national survey has classified every field in the country, year by year, since . In all of those years it has put this parcel in pasture or hay — never once in a tillage class, and in its worst year still % grass by area.
This rules out the single most disruptive thing that can happen to a soil community — the plough — for every year anyone has been keeping score, and it comes from a source with no stake in the answer. Two limits on it. The survey draws at thirty metres, coarser than the weedy patch itself, so strictly this describes the ground around the cores rather than under them. And what it settles is the plough: how the ground has been grazed is a separate question this record does not reach.
Forty years of it
That survey says what class this ground was put in. A second says what was actually growing on it, every year since — longer than anyone now managing this farm has been. Landsat scenes tied to some seventy-five thousand field plots estimate, each year, how much of a parcel was perennial grass, annual grass, shrub, tree, litter or bare dirt.
Each parcel's average cover across years. The two bars are measured independently and are not shares of one whole.
It answers the question the satellite figure could only raise. This parcel averages % perennial forbs and grasses over four decades and % trees. The parcels that read greener average % trees: they are woods, and have been the whole time. The greenness figure above is very largely a measurement of that.
On its own record the ground reads well. Bare dirt averages % of the parcel, and in years it has crossed % exactly time — in , at %. Across the whole record bare ground drifts down by points a decade while perennial cover drifts up by .
Set beside the two other open parcels, this field is the barer ground. It averages % bare, where the non-woodland parcels around it average %. It is the same shape as the greenness figure: the comparison flatters this ground when the neighbours are woods, and not when they are pasture. Two parcels cannot establish that it is worse, any more than they could that it is better — but a section that sets the woodland aside when woodland flatters the field has to set it aside when it does not.
There is a step in the record, and it is probably not about this farm. Split at , the parcel's perennial cover falls points — a finding, until you notice that of the neighbouring parcels fall across the same year, and that the year was picked by eye from this field's own series, not by any test. The lost cover also goes nowhere: bare ground, shrub, tree and litter all hold steady across the step, and real land-cover change has to land somewhere. Between a boundary picked from the same series it is being tested on and cover that vanishes without turning up anywhere else, the likelier story is the model moving rather than the field. So the step is written down here instead of charted.
This is a model's estimate of what a parcel's surface looked like from orbit, not a measurement of the soil under it. Its worth is that it is long, independent of everyone involved, and agrees with the two shorter records above: this has been grass, continuously, for as far back as anybody can see.
What would settle it
Three things, none of them exotic. A measurement of this soil's acidity, which drives the archaea-to-bacteria balance nearly as strongly as nitrogen does. The archaea themselves already argue that acidity is not the whole story, but that is an inference, and a pH meter would be an answer.
A set of soils given known, deliberately different amounts of nitrogen — if the ratio moves with the nitrogen, that settles the mechanism. It means running an experiment, though, rather than looking harder at anyone's field.
And a test of a different idea entirely. Some plants release compounds from their roots that shut down ammonia oxidisers directly, and how strongly they do it varies between grasses and clovers — the very plants growing side by side in this patch. If that is what is happening here, the clover pattern is not about nitrogen at all: it is the plant switching the archaea off. Testing it means washing the roots and seeing whether what comes off them stops the reaction in a dish — bench work on plants already growing here.