The gap between what testing covers and what it claims
Over a career spanning retail, federal service, and consulting, Frank has reviewed hundreds of microbiological specifications submitted by food companies. Almost none describe a sampling plan. They specify the test and the threshold — not how many samples to take, from where, or across how much total material.
According to Marcel, a typical sampling plan tests roughly 0.00002% of a batch. A negative result confirms that the fraction tested was clean — nothing more.
“A negative result tells you the samples you tested were negative. It doesn't tell you the rest of the batch is free of the organism.” — Dr. Marcel Zwietering
A recent FDA review of the US powdered infant formula supply tested approximately 312 samples across 16 brands for heavy metals, PFAS, and pesticides. The FDA's public communication concluded the results affirmed the safety of the supply. Within days, a recall followed for cereulide, a heat-stable bacterial toxin that the chemical-focused review never tested for.
Similarly, a raw milk cheese producer in California recently cited its own negative tests to decline a recall despite strong epidemiological evidence linking its product to an E. coli outbreak. In both cases, the testing program was treated as confirmation of safety it was never designed to provide.
What a sampling plan has to get right
A plan aimed at the wrong stage of production catches nothing. US poultry testing once focused on whole carcasses, but contamination would occur in the parts after cutting. The Salmonella and Campylobacter performance standards for poultry parts corrected that, but only after the gap had been producing illnesses.
“Generally, in microbiology, you'll have heterogeneous contamination — it won't be very homogeneous. It's good to take a lot of samples, from different locations, and analyze all of them.” — Dr. Marcel Zwietering
Contamination doesn't distribute itself evenly. It clusters, which means a plan drawn from one location can miss what's sitting in another. The N60 beef plan draws sixty samples across the carcass for exactly that reason. European rules for sprouted seeds apply the same logic at the batch level — testing the seeds before sprouting, the irrigation water that contacts the full batch, and the finished product.
The design of the plan determines what it can find. A plan that doesn't account for where contamination concentrates, or which stage of production it appears at, produces clean results and false confidence in equal measure.
How repeat negatives build the conditions for failure
In a well-controlled chocolate facility with a low background level of Salmonella, routine sampling turns up a positive about once every five years. The same contamination level produces roughly ten illnesses a year, none showing up on a certificate of analysis or triggering a recall.
“At a certain point, if you keep not finding positives, you get a little more negligent.” — Dr. Marcel Zwietering
The resulting negligence is self-reinforcing — fewer positives lead to less testing, which leads to more illness the program still isn't built to detect. Salmonella in chocolate has produced outbreaks on exactly this cycle since the 1960s — heightened attention, decline, then another outbreak.
Marcel picks not testing enough as the greater danger to food safety — a long run of clean results, he notes, is what typically produces it.
AI and the case for deliberately unbiased sampling
Risk-based sampling guided by AI can direct testing toward higher-probability failure points — suppliers with concerning histories, conditions that have preceded problems. Frank points to the FDA pilot, where he says machine learning applied to seafood import data dramatically increased the agency's chances of finding violative shipments.
“If you believe AI completely, that's not wise. If you don't use it at all, when it's a strong method for improving efficiency in certain cases, that's not wise either. What you should always do is use AI, and then critically assess what comes out.” — Dr. Marcel Zwietering
Any risk-based model is bounded by what it already knows. A contamination event that has never appeared in the training data will not appear in the model's output. Marcel believes half of one’s testing resources should be directed by risk-based sampling, and the other half intentionally kept unbiased, so the program can still catch what the model hasn't seen.
Validation gives your test results their meaning
Frank saw the consequences of skipping validation firsthand while visiting produce suppliers running identical tomato wash systems. Some achieved a two-log reduction in contamination. Others were adding a two-log increase — their systems had never been put to a controlled validation test. Without a test, there was no way to tell which systems were controlling contamination and which were making it worse.
“Always do your ultimate best — and then do a test.” — Dr. Marcel Zwietering
The industry's testing programs are only as good as the processes behind them. Marcel has spent a career making that case — and the ICMSF's books, sampling frameworks, and published research are where operators can start putting it into practice.
This episode of Smarter Food Safety is available now, wherever you get your podcasts.
To learn more about the host, Frank Yiannas, and why Ecolab is partnering on this show, read our profile here.
Episode Field Notes
Terms and resources worth bookmarking for food and beverage operators this year.
ICMSF Books: The definitive series on microbiological sampling, food safety management, and microbial specifications — developed over more than six decades by the International Commission on Microbiological Specifications for Foods and essential reading for anyone designing testing programs or microbiological criteria: https://www.icmsf.org/publications/books/
Microbiological Criteria for Foods: The Codex Alimentarius principles and guidelines for establishing and applying microbiological criteria — connecting sampling plan design to food safety objectives, hazard severity, and conditions of use: https://www.fao.org/fao-who-codexalimentarius/
HACCP Principles and Process Validation: FDA guidance on the validation and verification distinction Marcel identifies as the most important step before any testing program can mean what operators need it to mean: https://www.fda.gov/food/hazard-analysis-critical-control-point-haccp/haccp-principles-application-guidelines
NCBI Pathogen Detection: The public genomic database through which contamination made invisible to routine sampling can still link a company's product to clinical illness clusters via whole genome sequencing: https://www.ncbi.nlm.nih.gov/pathogens/
