Everyone says watch the water.Nobody has published how.

Ask anyone in broiler production what to watch and water comes up first. It is cheap to meter, it responds fast, and 85 percent of surveyed houses already have a meter on the line. It is also, on the published evidence, much harder to turn into a reliable alarm than its reputation suggests. This note sets out what is actually known, because it is one of the few places where our reading of the literature supports our own product position rather than complicating it.

The one empirical test, and what it found

In the same body of work that produced usable mortality thresholds on eight commercial Dutch broiler farms, the authors also tested feed and water intake from the same production charts. Their conclusion on those two variables was that reliable thresholds for food or water intake could not be established.

The false alarm rates explain why. At comparable sensitivity, mortality produced 2.0 false alarms per 100 production days. Water produced 16.2. Feed produced 37.2.

A system generating sixteen false alarms every hundred days is generating one roughly every six days. On a farm running multiple houses, that is a nuisance rather than a tool. The reason is not that water carries no information — it is that water intake moves for many ordinary reasons. Bird age, house temperature, feed form, line pressure, nipple height, and management events all shift it. The disease signal is real but it sits inside a noisy band.

What the extension guidance actually says

Given that, it is worth looking at what the best available practical guidance recommends, because it is notably more modest than the vendor literature. Mississippi State University Extension’s guidance on water quality and broiler performance advises investigating a decline in water consumption if the decline lasts for more than one day.

The best published guidance on water as an early-warning signal is a duration rule, not a threshold: does the decline persist past one day.

That is qualitative, and we think it is qualitative for a good reason. It encodes the thing a numeric threshold gets wrong — that a single day’s dip is usually noise and persistence is what carries meaning. It is also, notably, a rule about the shape of a trend over time rather than about a value crossing a line, which is precisely the kind of reading a house controller does not do.

The water-to-feed ratio

The ratio of water to feed intake is more informative than either variable alone, because it partly normalizes out the growth curve. The commonly cited normal band is 1.6 to 2.0, with a value above 2.0 sometimes flagged as a possible sign of an impending health challenge.

We use that band as a sanity check and not as an alarm. We could not find a study establishing its sensitivity or false alarm rate, which means it is industry convention with a plausible physiological basis rather than a validated detector. That distinction is one we intend to keep making, including when it is inconvenient.

A related finding on temperature

While looking at which variables carry predictive weight, one result is worth recording because it runs against intuition. In a 2025 machine learning study in Poultry Science, temperature and humidity ranked last in feature importance — 1.238 and 0.765 respectively — against 20.391 for day of the growout.

The bird’s age dominated everything. That is not a claim that house climate does not matter physiologically; it obviously does, and it is the one thing the controller manages well. It is a claim that once you know how old the flock is, adding the current temperature reading tells a model comparatively little. For a product deciding what to put in front of a grower, that is a useful ordering.

The alarm fatigue constraint

All of the above matters because of a finding from the dairy sector that we treat as a hard design constraint. In a study of automated alert systems, only 21 percent of alerts triggered a farmer to actually check the animal, and farmers were willing to act on fewer than roughly twenty alerts per day.

Combine that with a 16.2 per 100 day false alarm rate and the outcome is predictable: the alerts stop being read, and once they stop being read the product is worthless regardless of how good the underlying model is. A false alarm is not a minor cost in this category. It is the mechanism by which these products die.

Where this leaves the claim

There is no published lead-time figure for water intake in broilers, and the one attempt to establish thresholds concluded it could not be done reliably. That is not a gap we can fill with a better model on the same data — it may simply be a property of the variable. Our position is that water is a supporting indicator whose value is contextual, and that anyone marketing water-based early warning with a specific hours-of-warning number should be asked which study it comes from.