Enhanced Conversions Trains on Agent Purchases Unless You Stop It
Quick Answer: When an AI agent completes a WooCommerce purchase, Enhanced Conversions receives hashed customer data — name, email, address — and matches it at high confidence, feeding it to Smart Bidding as a human conversion signal. Agent purchases convert at 15–30% versus 2–3% human, so the algorithm trains on a cohort it cannot replicate in any buyable audience. The fix is server-side tracking at the PHP order hook that tags agent versus human purchases and conditionally excludes agent orders from the Enhanced Conversions stream — keeping Smart Bidding clean while agent revenue is tracked separately as a reporting-only conversion action.
In this article
- How do agent purchases reach Google Ads Enhanced Conversions?
- What happens when Smart Bidding trains on agent conversions?
- Did the April 2026 GA4 update make the problem worse?
- How do you stop agent purchases from reaching Enhanced Conversions?
- What are the symptoms of agent-contaminated Smart Bidding?
- How long does it take to clean Smart Bidding after removing agent data?
- Should you still track agent conversions in Google Ads for reporting?
- What is the business impact of clean vs contaminated Smart Bidding?
How do agent purchases reach Google Ads Enhanced Conversions?
When an AI agent completes a WooCommerce purchase, the same PHP order hooks fire Enhanced Conversions events with hashed customer data — Google Ads receives the purchase as a high-quality human conversion signal and trains Smart Bidding on it. Agent purchases include real customer data (name, email, address) — Enhanced Conversions matches them at high confidence, making the contamination worse, not better (Seresa, 2026). That’s the counterintuitive part: Enhanced Conversions is designed to improve match rates by sending first-party data. When the first-party data comes from an agent purchase, the match rate is excellent — which means Google is very confident about a conversion signal that shouldn’t be influencing your bidding at all.
The mechanism is mechanical. An MCP agent places an order. WooCommerce processes the checkout. The woocommerce_payment_complete hook fires. Whatever tracking plugin you’re running sends the purchase event to Google Ads, including the hashed email, hashed phone, and hashed address that Enhanced Conversions requires. Google matches the hashed data against its user graph and records a high-confidence conversion. Smart Bidding adds it to the training set. The system worked perfectly — it just learned the wrong lesson.
Agent purchases include real customer data — Enhanced Conversions matches them at high confidence, feeding a phantom conversion signal to Smart Bidding that the algorithm cannot replicate in any buyable audience (Seresa, 2026).
What happens when Smart Bidding trains on agent conversions?
Smart Bidding learns that users matching the agent-buyer profile convert at 15–30% versus 2–3% human — but no buyable audience segment matches that profile because the conversion was initiated by an AI agent, not by a human clicking an ad (Seresa / Presta Q1 2026). The algorithm sees a cluster of conversions with unusually high match quality, unusually high conversion rates, and perfectly clean customer data. It naturally tries to find more users who look like those converters. The problem is those “users” are AI systems, and Google can’t target an ad at ChatGPT’s shopping agent.
The result is a bidding model that’s optimising toward a ghost. It bids higher on signals that correlate with agent purchases — certain time-of-day patterns, certain device profiles, certain geographic clusters — because those signals historically preceded high-confidence conversions. But when real humans match those signals, they convert at 2–3%, not 15–30%. Your CPA rises. Your acquisition efficiency drops. And the gap between your reported ROAS and your actual human ROAS widens with every agent order the algorithm learns from.
Related: My GA4 Direct Traffic Doubled — Is AI Traffic Hiding Inside It?
Did the April 2026 GA4 update make the problem worse?
The April 2026 GA4-to-Google Ads update broke or partially disconnected conversion links for some advertisers — creating a window where conversion data was inconsistent, and agent purchases in the remaining connected data had outsized influence on bidding. Several advertisers reported GA4-to-Google Ads conversion links broke or partially disconnected after April 2026 (groas, 2026). The update changed how GA4 passes conversion data to Google Ads, and stores that relied exclusively on GA4-linked conversions found their bidding models suddenly training on a smaller, unrepresentative dataset.
For stores already sending agent-contaminated data, the April disruption compounded the problem. With fewer total conversions flowing through, each individual conversion carried more weight in the training set. Agent purchases — which match at high confidence and convert at artificial rates — became an even larger share of what Smart Bidding had to learn from. The stores that weathered the April update cleanest were those with server-side tracking that maintained a direct, uninterrupted conversion pipeline independent of the GA4 link.
How do you stop agent purchases from reaching Enhanced Conversions?
Server-side tracking at the WooCommerce order hook tags agent versus human purchases with a cohort flag, then conditionally excludes agent orders from the Enhanced Conversions event stream — Smart Bidding only sees human conversions. The conditional gate is: if human, send to Enhanced Conversions; if agent, route to separate reporting only (Seresa, 2026). The gate sits at the PHP layer where every order passes, regardless of whether it originated from a browser checkout, a REST API call, or an MCP agent transaction.
The implementation has three parts. First, the order hook inspects the incoming request for agent indicators — MCP origin headers, API authentication tokens, missing session cookies, known agent user-agent strings. Second, it writes a cohort flag (_order_cohort = human or agent) to the order metadata. Third, the server-side event pipeline reads that flag before dispatching: human orders fire Enhanced Conversions events with full hashed customer data; agent orders skip the bidding-eligible stream entirely and route to a secondary conversion action or a BigQuery table for separate reporting.
The fix is a conditional gate at the PHP order hook: human purchases fire Enhanced Conversions normally; agent purchases route to a secondary conversion action that reports without influencing Smart Bidding (Seresa, 2026).
| Scenario | Smart Bidding Signal | Reported ROAS | Actual Human ROAS |
|---|---|---|---|
| No agent separation | Mixed human + agent conversions | Inflated by 15–30% agent CR | Lower than reported |
| Agent excluded from bidding | Clean human conversions only | Matches actual performance | Same as reported |
| Agent as secondary action | Clean human conversions only | Human + agent visible separately | Accurately reported |
What are the symptoms of agent-contaminated Smart Bidding?
Rising CPAs despite stable conversion volumes, declining new customer quality, Smart Bidding spending more aggressively without proportional human conversion increases, and a gap between reported ROAS and actual human-only ROAS. The symptoms mirror many other Smart Bidding issues — the agent contamination is invisible without cohort-separated data to reveal it (Seresa, 2026). That’s what makes this problem dangerous: it looks like a dozen other things. You’ll troubleshoot audience targeting, creative fatigue, landing page experience, and bid strategy settings before you consider that the conversion data itself is compromised.
The tell is in the gap between your Google Ads reported conversions and your WooCommerce human-only orders. If Google Ads shows 100 conversions this month and your WooCommerce dashboard confirms 100 orders, the numbers match — but if 12 of those orders were agent purchases, Smart Bidding trained on 12 conversions that no ad placement could have influenced. Your effective human conversion count is 88, but Smart Bidding optimised as though all 100 were ad-attributed. The CPA it reports is lower than your actual human CPA, and the audience model it builds includes 12% signal that points nowhere.
How long does it take to clean Smart Bidding after removing agent data?
Google recommends allowing 2–4 weeks for Smart Bidding to recalibrate after any significant conversion data change — during this period performance may fluctuate before stabilising at the true human conversion rate (groas, 2026). That recalibration window is uncomfortable. Your CPAs may temporarily rise as the algorithm unlearns the agent-inflated conversion patterns it was optimising toward. Your reported ROAS may drop as the phantom performance disappears from the signal. Both are the algorithm correcting, not declining.
The practical advice is to prepare stakeholders. Before you flip the switch, document your current blended ROAS and your estimated human-only ROAS (using whatever agent detection data you have). When the 2–4 week fluctuation hits, the comparison shows what’s happening: the numbers aren’t getting worse, they’re getting honest. After recalibration, Smart Bidding targets an audience that actually exists, bids on signals that actually predict human conversions, and delivers a ROAS you can trust. The temporary discomfort is the cost of accuracy.
Should you still track agent conversions in Google Ads for reporting?
Yes — create a separate conversion action for agent purchases marked as secondary (not used for bidding) so you can see agent revenue in Google Ads reporting without it influencing the Smart Bidding algorithm. Google Ads supports primary (used for bidding) and secondary (reporting only) conversion actions — use this to separate signal from measurement (Seresa, 2026). The distinction is built into the platform. A primary conversion action feeds Smart Bidding. A secondary conversion action appears in your reports and columns but never touches the algorithm.
The setup mirrors the server-side gate. Your human conversion action stays primary — Enhanced Conversions fires for human purchases, Smart Bidding trains on them, and your bidding strategy reflects real acquisition performance. Your agent conversion action is secondary — it receives agent purchase events through a separate server-side path, appears in a custom column in Google Ads, and lets you track agent revenue growth without contaminating the model. You get visibility into both channels and clean optimisation for the one you’re paying to acquire.
What is the business impact of clean vs contaminated Smart Bidding?
Clean human-only Smart Bidding data drives accurate audience targeting, efficient bid optimisation, and predictable ROAS — contaminated data drives rising CPAs, audience mismatch, and budget decisions based on phantom performance that cannot be replicated. The cost of contamination compounds: every week of mixed data adds more wrong signal that takes 2–4 weeks to clear after correction (Seresa, 2026). Translation: the longer you wait to separate, the longer the recalibration takes when you finally do.
The math is unforgiving. If agent purchases are 10% of your conversions and you’ve been sending mixed data for three months, Smart Bidding has incorporated roughly 90 days of 10% wrong signal into its model. The 2–4 week recalibration window assumes a clean cut. If you only partially separate — filtering some agents but missing browser-based agents that mimic human sessions — the residual contamination extends the recalibration or prevents it from completing. The minimum viable fix is server-side tracking at the order hook, where every purchase passes and the cohort flag is set before any event fires.
A tool like Transmute Engine handles this at the WooCommerce order hook — tagging agent versus human at the PHP level, routing human conversions to Enhanced Conversions as primary events, and routing agent conversions to a secondary conversion action for reporting. The separation is clean from the first event, which means Smart Bidding starts recalibrating immediately on accurate data rather than a gradually improving mix. The 2–4 week window starts from day one, not from some future date when you’ve finished debugging a partial filter.
Key Takeaways
- Enhanced Conversions makes agent contamination worse, not better: agent purchases carry real customer data that matches at high confidence — Google trusts the signal completely.
- 15–30% agent CR versus 2–3% human: Smart Bidding trains on a conversion rate it can never replicate in any buyable audience segment.
- The symptoms are invisible without separated data: rising CPAs and declining quality look like a dozen other issues until you can see the agent share.
- Server-side gating at the PHP hook is the fix: tag agent vs human at the order level, then conditionally exclude agents from the Enhanced Conversions stream.
- Use secondary conversion actions for agent revenue: visible in reporting, invisible to Smart Bidding — both channels measured, neither contaminating the other.
- The contamination compounds weekly: every week of delay adds wrong signal that takes 2–4 weeks to clear, so the cost of waiting is nonlinear.
When an AI agent completes a WooCommerce purchase, the same PHP order hooks fire Enhanced Conversions events with hashed customer data — Google Ads receives the purchase as a high-quality human conversion signal and trains Smart Bidding on it.
Smart Bidding learns that users matching the agent-buyer profile convert at 15–30% — but no buyable audience segment matches that profile because the conversion was initiated by an AI agent, not by a human clicking an ad.
The April 2026 GA4-to-Google Ads update broke or partially disconnected conversion links for some advertisers — creating a window where conversion data was inconsistent, and agent purchases in the remaining connected data had outsized influence on bidding.
Server-side tracking at the WooCommerce order hook tags agent vs human purchases with a cohort flag, then conditionally excludes agent orders from the Enhanced Conversions event stream — Smart Bidding only sees human conversions.
Rising CPAs despite stable conversion volumes, declining new customer quality, Smart Bidding spending more aggressively without proportional human conversion increases, and a gap between reported ROAS and actual human-only ROAS.
Google recommends allowing 2–4 weeks for Smart Bidding to recalibrate after any significant conversion data change — during this period performance may fluctuate before stabilising at the true human conversion rate.
Yes — create a separate conversion action for agent purchases marked as secondary (not used for bidding) so you can see agent revenue in Google Ads reporting without it influencing the Smart Bidding algorithm.
Clean human-only Smart Bidding data drives accurate audience targeting, efficient bid optimisation, and predictable ROAS — contaminated data drives rising CPAs, audience mismatch, and budget decisions based on phantom performance that cannot be replicated.