Reading an agritech yield claim

Agricultural technology is sold on numbers. Thirty percent less water. Twenty percent higher yield. Half the labour. The numbers are specific, confident, and frequently untraceable to anything.

They are not usually fabricated. They are more often real results, stripped of the conditions that produced them, at which point they stop being evidence and become decoration.

The four questions

Compared with what? This is the question that resolves most claims on its own. Every improvement figure is a comparison, and the comparison is chosen by whoever reports it. Sensor-based irrigation scheduling compared with a fixed weekly timer will show a large saving. The same system compared with an experienced grower irrigating on judgement usually shows a modest one. Both numbers are true; only one is relevant to someone who already irrigates competently.

Where, and in what conditions? Agricultural results are local. A water saving established under arid irrigated conditions with a sandy soil transfers poorly to a temperate rainfed clay. Any claim without a location and a season is missing the information needed to judge whether it applies.

Over what area and how many replicates? Small plots outperform fields, consistently and for structural reasons: they get more attention, edge effects are proportionally larger, and variability that would average out across a hectare shows up as a clean difference across a few square metres. A single-season, single-site result is a hypothesis, not a finding.

What happened to the results that were not reported? A trial that ran across eight sites and reports the two that worked is telling you something quite different from what it appears to be telling you. This is rarely detectable from the marketing material, which is precisely why the underlying report matters.

The patterns worth recognising

The recovered baseline. A claim measured against a deliberately poor starting point. Common in irrigation and nutrition products, where the control is a practice nobody competent still uses.

The best-case aggregation. Several trials run, the favourable ones combined into a headline figure. Detectable by asking how many sites were involved and whether all are included.

The proxy substitution. An improvement demonstrated in something that correlates with yield - canopy vigour, root mass, an index reading - and then reported as though yield itself had been measured. Vegetation indices are especially prone to this, since they saturate in dense canopies and can improve without any change in harvestable output.

The unattributed round number. Thirty percent, half, double. Numbers that have been passed between marketing documents until nobody can name a source. When a vendor cannot say which trial produced their headline figure, that is the answer.

What good evidence looks like

It exists, and it is not hard to recognise. It names the location, season and crop. It describes the control honestly. It reports variation, not just a mean - and the variation is often the most useful part, because it tells you how likely you are to get the average result rather than one of the disappointing ones. It reports the trials that did not work. And it is findable, because the underlying report is published rather than referenced.

A substantial share of agricultural research meets this standard and is freely readable, which is the genuinely useful fact here. Checking a claim is usually a twenty-minute exercise, and the method is set out in finding agricultural research.

The reason to bother is not scepticism for its own sake. It is that some of these products work well and are worth the money, and the only way to identify them is to apply the same standard to all of them.