Methodology
How FDA Recall Radar works
Sources, matching, AI summarization, and limitations — explained plainly.
1. What FDA Recall Radar monitors
FDA Recall Radar monitors official public records and food-industry signals — not internal company data, not private databases. Each week it pulls public enforcement records and news signals and checks them against the companies a subscriber tracks: your own firm and any named competitors.
Each source type is treated separately and labeled in every digest.
2. Source taxonomy
FDA Recall Radar separates official public records from news signals.
FDA food enforcement record
Class I, II, and III recalls from FDA openFDA — public enforcement records.
USDA FSIS recall / public health alert
Recall notices and public health alerts from the USDA FSIS recall API for meat, poultry, and egg products.
USPTO trademark filingNot currently available
Not currently available. No verified public USPTO trademark search API exists yet (see /methodology and docs/decisions/0002-uspto-endpoint.md) — trademark filings are not included in digests until a verified, official data source is wired in.
News signal
Category context, competitor mentions, and food safety events sourced from GDELT — the open global news database.
News signals are never labeled as official government records. USPTO trademark monitoring is not currently active — we have not found a verified, official public USPTO trademark search API to build on. It is not included in any digest until a verified data source is wired in (see Limitations below).
3. How company matching works
Each week, all retrieved records are checked against the subscriber's watchlist — the company name, registered aliases, and named competitors. Matching uses normalized name comparison plus fuzzy and trigram similarity scoring across structured fields such as firm name, brand name, and product description.
Only records that meet a minimum confidence threshold are surfaced. A record that does not meet the threshold is not shown. Confidence tiers are designed to reduce false-attribution risk.
When matching is uncertain, the system does not attribute. We prefer false negatives over false positives: a missed record is recoverable; an incorrect attribution is not.
4. Confidence tiers
Every entity-specific digest item is labeled with one of three confidence tiers. The tier determines how the item is worded.
Confirmed
The firm name in the public record matches your company or competitor name exactly, or matches a registered alias. Attribution is allowed.
Attribution allowed: Yes — direct attribution used.
Probable
The firm name is a strong but not exact match. Hedged language is used. Always click through to the official source to verify before taking any action.
Attribution allowed: No — hedged language used. Never attributed as a confirmed company event.
Possible
A partial or low-confidence match. Shown as a category-level signal only. Not attributed to a specific company.
Attribution allowed: No — hedged language used. Never attributed as a confirmed company event.
A lower-confidence match is never worded as a confirmed company event. When uncertain, we do not attribute.
5. How AI summaries are generated
AI assists with summarizing retrieved source material — it does not generate or invent facts.
The model only ever sees records that were actually retrieved from the source APIs. It may not add, infer, or speculate any fact, number, date, or claim not present verbatim in those records.
A verification step checks that every proper noun, recall number, and company name in the output is present verbatim in the retrieved inputs. Outputs that fail verification are dropped, not shown.
AI assists with summarizing retrieved source material. Digest items are designed to be source-grounded and include direct links to the underlying records or articles used.
6. Quiet-week fill ladder
When there are no direct hits for your company in a given week, the digest follows a structured fill order:
- 1
If your company has confirmed recall hits in the past 7 days — show them.
- 2
If not, expand to 30-day watchlist context for your company.
- 3
If still quiet, show category-level recall and public health alert signals.
- 4
Also include any competitor signals and news mentions from the period.
- 5
Result: every digest contains meaningful intelligence, even in quiet weeks.
You never receive a blank digest.
7. Limitations
Automated matching may miss records (false negatives) or surface incorrect associations (false positives). Always verify by clicking through to the official source.
openFDA data typically lags the official FDA announcement by a few days to one week.
GDELT news signals cover a large volume of sources and may include articles of varying quality. News signals are labeled separately from government records.
USPTO trademark monitoring is not currently available. We evaluated it and found no verified, official public USPTO trademark search API to build on — we do not fabricate or estimate trademark data, so none is shown until a verified source exists.
We do not cover state-level recalls, import alerts, or enforcement actions not published to the monitored federal APIs.
8. Corrections and contact
If you believe a digest item is incorrectly attributed or contains an error, contact us at support@fdarecallradar.com. We review all correction requests and correct the record as appropriate.
9. Government affiliation disclaimer
FDA Recall Radar is an independent product that monitors public source records and food-industry signals. It is not affiliated with, endorsed by, sponsored by, or approved by the FDA, USDA, FSIS, USPTO, or any other government agency.
FDA Recall Radar is an independent product that monitors public source records and food-industry signals. It is not affiliated with, endorsed by, sponsored by, or approved by the FDA, USDA, FSIS, USPTO, or any other government agency.