Resemble agent detection: check the false positives
Resemble’s agent detector is advisory. Its own human false-positive disclosure explains why merchants should measure traffic before blocking it.
An AI-agent label is a reason to investigate a visit, not proof that the visitor is a fraudster. For merchants considering behavioural detection, Resemble AI’s own disclosure makes that distinction particularly important: its default Balanced setting flags about 5–6% of human visits as agents.
Resemble introduced RAD-S1 on 29 October 2026. This is a closer look at an existing release, rather than a new November launch. The company says its internal tests detected at least 98.8% of disguised-agent visits. That result is vendor testing, not an independent assessment of performance on a Shopify store.
Read the benchmark alongside the mistakes
The test measured detection on agent visits; it was not itself a measurement of how often real customers would be misclassified. The separately disclosed human false-positive rate matters if a merchant plans to turn a classification into a checkout restriction.
There is also an inconsistency in the announcement: one section says the test covered 12 website stacks, while its limitations section says eight. We would not use that account to make a universal ranking of bot-detection products.
What the integration actually provides
According to Resemble’s documentation, its browser script evaluates interaction signals and returns a verdict such as human, agent or review. A short visit without enough evidence may receive no verdict. The service does not block anyone automatically; the site decides what action to take.
That leaves two separate questions: whether the classification is useful, and whether the response is appropriate. A customer’s shopping assistant and abusive automation can both be agents. Conversely, a human classification is not a guarantee that an order is legitimate.
The documentation describes the signal as advisory because browser-supplied evidence can be falsified. It also says keystrokes are recorded by category rather than as typed text, IP addresses are hashed before storage, and retained event streams have an integration-specific expiry. Review those collection and retention settings before a trial.
A sensible merchant pilot
Our recommendation is to begin with observation. Define a small set of journeys, record classifications alongside completed purchases and support reports, and review apparent mistakes before changing access rules. Include mobile shoppers and customers using different input methods in the test rather than assuming a desktop demonstration represents everyone.
Decide what should happen when there is no verdict, and give legitimate customers a recovery route if a later rule challenges them. Measure abandoned journeys and support workload alongside detected automation. A higher detection percentage is not a useful business result if the intervention also prevents customers from buying.
This is an evaluation checklist, not a verified Shopify installation guide. Confirm that the intended storefront or checkout surface supports the proposed integration before adding it to a live store.
More from News
See all →