%20invert(1)`%20at%20the%20call%20site%20-%20the%20ring%20already%20ships%20white;%20the%20filter%20would%20flatten%20the%20two%20brand-green%20dots%20to%20white%20with%20it.%20--%3e%3csvg%20xmlns='http://www.w3.org/2000/svg'%20viewBox='0%200%2064%2064'%20width='64'%20height='64'%3e%3cpath%20d='M31.25%2010.72%20A21.28%2021.28%200%201%200%2051.71%2037.87'%20fill='none'%20stroke='%23FAFAFA'%20stroke-width='11.95'%20stroke-linecap='butt'/%3e%3ccircle%20cx='51.71'%20cy='37.87'%20r='5.98'%20fill='%23FAFAFA'/%3e%3ccircle%20cx='54.03'%20cy='20.99'%20r='6.07'%20fill='%2354AE80'/%3e%3ccircle%20cx='42.64'%20cy='10.35'%20r='4.39'%20fill='%2354AE80'/%3e%3c/svg%3e)
OntoReviewWe find product safety and quality reports in public consumer reviews.
OntoReview analyzes public consumer reviews to identify reports of potential product safety and quality concerns. Each report is preserved with its source and date, giving your team evidence to investigate.
Find the signals. Keep the evidence. Support the investigation.
Nothing to install and no data to send us. The sample is a full report on a public case, sent by email.
Six public reviews described symptoms in pets. All are dated before the recall.
The recall was triggered by a state retail sample test, not by these reviews, and we do not claim reviews predict recalls. What this case shows is narrower: the reports existed, in public, with dates, and could be collected, classified and kept.
What this case does not show
It does not show that the difference from the controls is real, or that an alert built on these labels would have been useful.
Emerald marks measured or public-record data. Amber marks an estimate or an unverified claim. We keep the two apart here and everywhere in the product.
What this does not do.
Reviews are one surface, and they only show what a customer could notice. Here is where the method goes quiet.
It does not predict Prop 65 notices.
Those start with a plaintiff buying the product and sending it to a lab. Nothing about that appears in a review beforehand, so there is no signal for us to read.
It cannot see what a consumer cannot.
Heavy metals, pesticide residue, sub-clinical contamination. If no one can taste, see, or feel it, no one writes about it — and we read what people write.
What we read is what people experienced.
Illness, injury, foreign material, a label that did not match what was in the package. That is the readable surface: an unverified consumer claim, not proof that a product caused harm.
Regulatory records are compiled by hand.
Where relevant, public regulatory records are compiled by hand in the report and marked as such. Not finding a record in the sources checked does not mean none exists.
One case is one case.
The Victor comparison is one product against two controls, with 34 reviews in the key window. The same-season comparison was not significant and the alert rule raised a 12-week false alarm on a control. It is not a hit rate, and we do not present it as one.
From a public review to evidence your team can check.
Collect
Public reviews of your product are collected from agreed public sources. The source of each review is recorded.
Classify
Each review is read in full and placed in one class: a safety report, a reaction to news, a product change, or none of these. The classifier sees only the review text.
Add context
Where relevant, public regulatory records are added by hand and marked as such.
Keep
Every report keeps its review ID, date, rating, source and the reason for its classification, so your team can check each one.
Other tools count negative keywords. We read each review and keep the ones worth investigating.
The figures below are from the worked case above: 814 public reviews of three dog foods.
Keywords are noisy
In our worked case a keyword list flagged 145 of 814 reviews. Read one by one, 53 of those were reports of a pet's symptoms. The rest were something else.
Reactions are not reports
Of the 92 keyword hits that were not safety reports, 36 were written in reaction to recall news and 30 described loose stool after switching foods. We sort these out instead of counting them.
Judged on the text alone
The classifier is not told the brand, the date or whether a recall happened, so it cannot flag a review because of what came later.
One report, fixed price.
A one-time report on what consumers have publicly written about your product. It is the only thing we sell today.
Additional products: $800 each
- ·Public reviews from agreed sources for the product you specify, collected and classified for analysis
- ·Every safety or quality report, with its review ID, date, rating and source
- ·Reports sorted by type: safety report, product change, reaction to news
- ·Public regulatory records for the company, compiled by hand, with their sources
- ·A section on false positives and the limits of the classification
- ·An audit CSV listing each review and how it was classified
- ·A one-hour session to walk through the findings
- ·Delivered manually within four weeks
A sample report on a public case is available on request, so you can judge the output before you buy.
Request a sample reportThe first three reports are $1,500 (one product each), in exchange for permission to publish the engagement as a case study.