What this is, and what it has to do with the farm
Nothing, directly, and that is worth stating before anything else. The rest of this site is thirteen cores from a single weedy patch, and it is careful throughout about how little thirteen cores from one patch can settle. This page is a different question asked of different soil: not what grows where, but whether soils change together. It shares only the method - read the DNA, count who is there, compare.
The reason it sits on this site at all is that it answers the objection the farm study cannot: that a pattern seen once, in one place, might be that place. Here the same kind of measurement was repeated at independent sites over years, which is the scale at which a pattern stops being an anecdote. It says nothing about whether this farm's soil is good, and it is not evidence for any practice.
The observatory sampled about ten fixed plots at each site and returned to them each year, so a site was a repeated measurement rather than a fresh guess. Cores were compared only within a site, then sites compared to each other - so the enormous differences between an Alaskan bog and an Arizona desert can never masquerade as change over time.
The communities really did change from year to year
The first question was whether there was anything to explain. Soil taken from the same plots in two different years should look more different than soil taken from the same plots in the same year - if it does not, there is no change and nothing follows.
It did, at of the site-and-depth groups tested. The shift was small in absolute terms - about on a dissimilarity scale that runs from 0 to 1 - but it was consistent, it was measured against reshuffled years rather than against nothing, and it appeared in tundra, rainforest and desert alike.
So soil communities were not static between years. That is the premise everything below rests on, and it is the least controversial thing on this page.
Distant sites moved in the same direction
Change on its own could be local: each site drifting its own way in response to its own weather. The stronger claim is that the drift is shared - that when one site's soil shifts, distant ones shift the same way.
Testing that meant describing each site's change as a direction rather than an amount: for each of the bacterial genera common to every site, which way was it going, and how fast. Two sites agree if the genera rising at one are the genera rising at the other. They did, on pairs of sites: ρ = to , which is a positive correlation between places that share nothing but a calendar.
That is a range and not a single figure on purpose. The analysis included only sites with a measurable year effect, and exactly one site sat close enough to the cut that it went in or out depending on the random draw - , admitted in % of repeats, against sites that were always in and always out. It moved the size of the answer and never its direction: the correlation was positive and statistically clear in % of those repeats.
Reporting the range rather than the friendlier end is the point. Picking whichever repeat gave the strongest number would have been reporting the favorable end of a range as though it were the measurement.
It was a shared direction, not a shared drift
Five yearly readings cannot, by themselves, tell a one-way change from a slow swing back and forth. A community oscillating on a two-or-three-year cycle, with sites in step, would produce exactly the shared directions above. So we asked whether each year's step repeated the last one.
It did. Consecutive steps agreed at where shuffling the years gave (p = ), and the first half of the series agreed with the second at against (p = ). Both observed numbers look bad and are good: any two consecutive steps share a year between them, once as an endpoint and once as a starting point, so even pure noise scores about minus a half here. The comparison that means something is against the shuffle, not against zero.
But the change did not pile up. If sites were drifting steadily away from where they started, communities four years apart would be further apart than communities two years apart. They were barely different ( at two years, at four, on a scale where the year effect itself is many times larger).
So the supported claim is narrower than the word "drift" suggests: distant soils changed in the same direction in the same years. Whether those changes compound into long-term movement is a question five years cannot answer, and this page does not claim they do.
Most of it was not the machine
There is an obvious way to get a fake version of this result. Samples from one year tend to be sequenced together, so anything peculiar about a sequencing run would push every site's reading the same way that year and manufacture agreement between sites that share nothing.
That confound is usually just acknowledged. Here it could be measured, because the observatory sometimes split one year's samples from the same sites across two runs. Within one of those years at one of those sites the soil and the year are fixed, so whatever differs between the two runs is the machine. That gave measured artifact directions, which could then be subtracted from every site's trajectory.
% of the agreement survived the subtraction (p = ). Doing the same subtraction with deliberately scrambled run labels - to price what removing any four directions costs - put the real machine contribution at about %. Real, worth knowing, and a minority of the effect.
Chemistry predicted who moved together. Distance did not.
If shared weather drives the synchrony, then nearby sites should agree more than distant ones, because nearby places share weather. That is the textbook prediction, and it was written down before the test was run.
It failed. Across to kilometers there was no relationship between how far apart two sites were and how well they agreed (ρ = , p = ). Sites on opposite sides of the continent were among the closest-matching pairs.
Distance is only a stand-in for shared weather, though, and a poor one across a continent. So we measured the weather itself: daily temperature and rainfall at each site's coordinates, turned into how unusual each year was against a normal, then compared between sites. Two sites that had the same kind of year should, on this theory, move the same way. They did not (ρ = , p = across pairs, for the window picked in advance as the one that could actually have influenced the soil before it was sampled).
Soil pH did predict it. The more two sites differed in acidity, the less their communities moved together - ρ = to across the same repeats, over a pH range of to . Both axes were tested the same way, on the same pairs, in the same run. The pH axis was the only one of the three that pointed consistently in any direction at all.
This is the finding worth the page. Synchrony between distant soils was organized by what the soil is like, not by where it is. That makes sense of what is otherwise a strange result: pH is the master variable for soil bacteria, and pH is not smooth across a map - two neighboring fields can differ by more than two sites on opposite coasts.
The weather result comes with an honest complication. On the sites that measured themselves most reliably - a way of choosing sites that was decided after seeing the data, not before - shared temperature did track shared change (ρ = , p = , surviving correction for the weather measures tried). It is reported rather than buried because it points the predicted way and is the obvious thing for a larger study to chase. But it appeared under one choice of sites and not the other, and soil pH held under both. That asymmetry is the reason this page leads with chemistry.
The pH relationship was the fragile one, and it is worth being precise about how. Its direction never changed: negative in all repeats and in all re-runs where the reads were subsampled afresh. Its strength did - across those re-runs it moved between and and passed the usual significance bar in of . So the honest reading is a consistently downward relationship whose size this dataset does not pin down. A single site drove it: the arctic tundra plot was the most acidic ground in the set with nothing near it, so whether it is included decides which end of that range you get.
The agreement between sites had no such problem. Across the same re-runs it stayed between ρ = and and was statistically clear in % of them. It also flattened, like the pH result, when the sites that could not measure themselves were left in - over all groups the pH relationship read ρ = - which is what including sites that contribute no measurement does to any average.
A third of the sites could not measure their own change
Splitting one site's samples in half and computing its trajectory twice should give the same answer twice. At of the groups it did not: the two halves disagreed, so that site could not estimate its own direction of change, and it could not agree with anywhere else either.
The expected explanation was patchiness - soil varying so much from plot to plot within a site that the year-to-year signal is buried. That turned out to be wrong. Plot-to-plot variation was essentially the same at the sites that worked and the sites that did not ( against of total variation), and it barely predicted which was which (ρ = ). Blocking the split by plot changed the answer by , and an estimator that removes plot differences outright rescued of them.
What separated them was the signal itself. Year-to-year change accounted for of community variation at the sites that could measure themselves and only at those that could not - and that one quantity predicted the outcome almost perfectly (ρ = ).
The signal was not buried at those sites. It was absent. That is a practical finding about monitoring rather than a statistical footnote: a site can be sampled carefully for five years and still be unable to report whether its soil changed, and nothing in its sampling design - plot count, sample count, sequencing depth - says in advance which sites those will be. It has to be measured, and it is worth measuring before a monitoring program is designed around the assumption that it can be skipped.
What this does not show
It does not show a cause. Something continental moved these communities together and tracked soil chemistry; this analysis does not identify it, and a shared driver is not the only story that could produce the pattern.
It does not settle the depth question. The organic-horizon groups are too few and too noisy to compare against mineral soil, and the difference between them is not statistically distinguishable, so no claim is made in either direction.
It does not show that the change builds up - that is measured above and it does not. The word "drift" is doing less work on this page than it looks like it is, and the title is the last place it still overstates.
It is annual measurements from one observational program, read at the level of genus. Five points is a short basis for a trajectory, and no amount of care in the analysis converts observation into experiment.
The obvious check is to ask the same question of fungi, where soil pH matters much less to which species live where. If bacteria followed chemistry and fungi did not, that would point at biology rather than at something in the processing. The data exists - the same program sequenced fungi on % of the very same soil cores. The check still would not settle anything, and the reason is worth being plain about: a comparison can only be as sharp as the number it is compared against, and the chemistry result above is not sharp. Working through what the two possible answers would look like: if fungi really ignore chemistry, the check would say so about % of the time, and if they really follow it, about %. Both are close to a coin toss, so either answer would mean very little. It is left undone deliberately, rather than left out.
And it is not about the farm. It shares no soil, no samples and no site with the rest of this survey. If it earns anything for those thirteen cores, it is only the general point that soil communities are moving - not any conclusion about this ground. What the farm study can and cannot establish is set out on the front page, and the methods for both are in methods.