Bots in your conversion data: why filtering before delivery matters

Bot traffic is usually discussed as an analytics nuisance: inflated sessions, odd spikes, bounce rates that make no sense. The bigger cost is less visible. When automated traffic triggers events, those events are sent to your ad platforms as real behaviour, and the platforms learn from them.

How bots end up in conversion data

Invalid traffic covers bots, scrapers, data-centre traffic and automation tools that reach your site with no person behind them. Industry estimates put a substantial share of programmatic ad traffic in this category. Some of it does more than load pages: it adds to cart, starts checkouts, submits forms. Every one of those events is a signal you send to Meta, Google, TikTok and GA4.

The effects compound:

• Bidding learns the wrong audience. Smart bidding optimises towards the signals it receives, including automated ones.
• ROAS is overstated. Bot conversions make campaigns look better than they are.
• Lead quality drops. Form-fill bots become "leads" that sales has to clear.
• Delivery capacity is wasted on events that should never have been sent.

Why browser tags cannot fix it

Client-side tags run in whatever browser loads them, including automated, remote-controlled browsers. The JavaScript fingerprint they see is genuine, so the tag has no reliable way to tell a bot from a person. Each vendor runs its own invalid-traffic detection afterwards, but by then the event has already been delivered and counted.

Filtering on the server

A server sees things a tag cannot: the connecting IP address and its network, request rates across many visitors, geography and request headers. Datafly Signal uses these server-observed signals, across several detection layers, to score every event at the core of the pipeline, before anything is delivered.

Each filtered event is logged with the rule that matched and the action taken: drop it, quarantine it for review, or tag it and let it through. You can see the filtered rate per pipeline in real time and inspect individual events. Quarantined events can be routed to a warehouse table for analysis instead of being discarded.

Trying rules without losing real conversions

The risk with any filter is blocking real customers. Every detection layer in Datafly Signal has an observe mode: the layer scores every event and reports what it would have filtered, but its score does not count towards the drop decision. New environments start in observe mode, so the first weeks measure bot traffic without touching delivery. When the numbers look right, you switch a layer to enforce.

What changes when you filter upstream

Reported conversions may fall, because bot conversions stop being counted. What remains is real, and that is what bidding and lookalike audiences should be learning from. Vendor-side detection still runs; filtering upstream means the bot event never reaches it in the first place.

Checklist

• Compare conversion rates by network and geography for patterns that do not look human.
• Start any new rule in observe mode and review what it would remove.
• Quarantine rather than drop while you build confidence.
• Filter once, before delivery, so every destination gets the same clean data.

See how server-side bot filtering works in Datafly Signal, or request a demo.