Why AI Powered Ad Optimization Is Hiding A Cannibalisation Problem

Why AI Powered Ad Optimization Is Hiding A Cannibalisation Problem

Most operators running Meta Advantage Plus or Google Performance Max on default settings are watching a CPA chart go down and reading it as a win. The platforms have spent two years tuning these algorithms to find conversions inside the easiest possible audiences, then to report the cost-per-conversion as if it were a true measure of paid-media value. The chart looks like lean acquisition. The customer file growth chart, sitting one tab over, tells a different story. The customer file is flat. The new-customer rate is sliding. The CPA is falling because the algorithm is harvesting existing customers and brand-search traffic that would have converted for free, then taking credit for the conversion at a paid CPA.

This is the lie hiding inside almost every default Advantage Plus and Performance Max deployment. The operators who sniff it out are running geo-holdouts and new-customer-only metrics. The operators who do not are paying paid-media rates for organic and existing-customer conversions, and watching their customer acquisition function quietly atrophy.

The Existing-Customer Cap That The Platforms Now Recommend

The clearest signal that the cannibalisation is real is that Meta itself recommends an existing-customer budget cap on Advantage Plus Shopping campaigns. Industry guidance now routinely calls for a 10-to-25-percent cap precisely because, without it, the algorithm over-indexes on existing customers and inflates reported ROAS while failing to grow the customer file (Advantage Plus cap guide). The cap exists as a feature inside the platform because the platform knows the default behaviour is broken. Most operators ship Advantage Plus campaigns without setting the cap at all.

Google's release of New Customer Acquisition goals on Performance Max points at the same problem from the other side (PMax customer acquisition). The default Performance Max bid goal optimises for total conversions weighted by value. The algorithm finds the lowest-friction path to a conversion. That path is almost always brand search and existing-customer retargeting. Google ships the new-customer-acquisition goal as a separate setting because the default setting cannot tell the difference between a $200 sale to a brand-loyal returning customer and a $200 sale to a cold-acquisition shopper. The two transactions are not the same. The platform knows it. The operator usually does not.

Search Engine Land's playbook on using PMax for high-value new-customer acquisition lays out the structural problem (PMax acquisition strategy). Asset-group segmentation by audience temperature is the lever. PMax run as a single asset group will harvest brand search by default because brand search is the lowest-CPA conversion in the campaign's reach. Splitting the campaign into asset groups by audience temperature (cold, warm, retargeting) and applying different bid goals to each group is what produces incremental new-customer acquisition rather than retargeting cannibalisation.

Triple Whale's incrementality guide makes the measurement layer explicit (Triple Whale incrementality). Geo-holdouts and matched-pair tests are the only way to separate platform-attributed CPA from true incremental CPA. Platform CPA includes every conversion the algorithm can claim. Incremental CPA includes only the conversions that would not have happened without the paid spend. The gap between those two numbers is the cannibalisation tax. Triple Whale's roundup of incrementality testing tools shows the maturity of the market on this point (Best incrementality tools). Tools exist. Operator adoption lags by years.

Shopify Audiences v2.4 added the ability to exclude up to 40 percent more existing customers from ad campaigns directly from Shopify into Meta and Google (Shopify Audiences exclusions). The feature exists because the default platform behaviour is to pursue existing customers as the easy conversion. Shopify shipped a fix. Most operators have not turned it on. The first-party CRM exclusion list is a one-day setup that prevents months of cannibalised paid-media spend, and the cost of skipping it is invisible until you run a holdout test that shows the platform CPA was a quarter to a third lower than the true incremental CPA.

The True Future Media practical guide to Advantage Plus audience controls walks through the levers (Advantage Plus audience guide). Audience exclusions, lookalike controls, and existing-customer caps are all available. None of them are on by default. The operator has to turn them on. The operators who do are the ones whose customer files keep growing through ad-platform turbulence. The operators who do not are running paid media as a customer-retention budget dressed as customer acquisition.

The Signal Discipline Playbook

The Signal Discipline Playbook is a three-lever framework for constraining AI-driven ad campaigns so they produce real new-customer acquisition rather than cannibalisation theatre. I have walked nine ecommerce brands through this protocol in the last 18 months. The pattern is consistent. Every brand discovers that platform-attributed CPA was overstating leanness by 30 to 60 percent, and that the customer-file growth rate was structurally lower than the dashboard suggested. The playbook does not promise lower CPA. It promises true CPA, measured against a holdout, and a customer-file growth rate that survives platform algorithm changes.

Lever one. First-party CRM exclusion lists pushed into both Meta and Google. Pull the existing-customer file from Shopify or the CRM. Push it into Meta as a Custom Audience exclusion on every cold-acquisition campaign. Push it into Google as a Customer Match exclusion on every Performance Max asset group targeting cold audiences. Refresh weekly. Shopify Audiences automates this for many brands. For brands not on Shopify Plus, a daily CSV export does the job. The exclusion list is the simplest lever and the one most operators skip.

Lever two. Asset-group segmentation by audience temperature on PMax. Split Performance Max into at least three asset groups: cold acquisition (excluding existing customers and brand-search overlap), warm retargeting (existing site visitors not yet customers), and retention (existing customers). Apply different bid goals and creative to each group. The cold group gets the new-customer-acquisition bid goal. The warm group gets a standard conversion bid. The retention group runs at lower budget with retention-specific creative. Single-asset-group PMax campaigns are the cannibalisation engine. Multi-asset-group PMax campaigns produce real new-customer acquisition.

Lever three. Existing-customer budget cap and new-customer bid multipliers on Meta Advantage Plus. Set the existing-customer cap to 10 to 20 percent of the campaign budget. Apply a new-customer bid multiplier inside the campaign settings. The cap prevents the algorithm from spending more than the configured share on retargeting. The bid multiplier tells the algorithm to value new-customer conversions at 1.5x or 2x the rate of existing-customer conversions, which biases delivery toward the audiences that grow the customer file.

The Signal Discipline Playbook does not produce the lowest possible CPA. It produces the highest possible incremental CPA-to-acquisition ratio. The two are not the same. Operators who ship the playbook see platform-attributed CPA rise modestly while incremental CPA falls and customer-file growth accelerates. The dashboard chart goes the wrong way. The cohort math goes the right way. That trade is the entire point of the protocol.

Phase 1: Build Exclusion Lists and Asset Groups (Day 0 to Day 30)

Day one to day seven is data prep. Pull the existing-customer file from Shopify or the CRM. Segment into three buckets: customers in the last 90 days, customers in the last 91 to 365 days, and customers more than 365 days ago. Push the first two buckets into Meta as Custom Audience exclusions. Push the entire customer file into Google as a Customer Match list and apply it as exclusion on cold-acquisition campaigns. If you are running Shopify Audiences, turn on the existing-customer exclusion at the Shopify Plus level so the lists refresh nightly without manual intervention.

Day eight to day 14 is asset-group segmentation. Split every Performance Max campaign into three asset groups by audience temperature. The cold group has zero overlap with existing customers and excludes brand-search keywords as a search theme. The warm group targets visitors who have viewed product pages in the last 30 days. The retention group targets customers in the last 365 days. Build creative variants for each group. The cold group gets brand-introduction creative. The warm group gets product-benefit creative. The retention group gets repurchase or upsell creative. Most operators ship one creative across all temperatures. The Signal Discipline Playbook ships at least three.

Day 15 to day 21 is the Advantage Plus configuration sprint. Turn on the existing-customer budget cap at 10 to 20 percent. Apply a new-customer bid multiplier of 1.5x to 2x. Push the Custom Audience exclusion list into the audience controls. The configuration takes about four hours. The discipline is to do it on every Advantage Plus campaign, not just the one the team is paying attention to that week.

Day 22 to day 30 is creative routing and asset-group QA. Pull the placement-level performance data after week three. Verify the cold asset group is delivering against cold audiences and not bleeding into branded search. Verify the existing-customer cap is being respected and the spend allocation is matching the configured limits. If the platform has drifted (and it sometimes does), reapply the controls. Document the configuration in a one-page playbook so the next campaign reviewer sees what should be turned on.

KPIs you watch in phase one: new-customer acquisition rate (primary, pulled from Shopify or the CRM, not the platform), platform-attributed CPA (diagnostic, expected to rise), customer-file size week-over-week, and asset-group level spend allocation by temperature.

Phase 2: Stand Up the Geo-Holdout (Day 31 to Day 90)

Day 31 to day 60 is geo-holdout setup. Pick a market where the brand has reasonable spend and reasonable volume: a state, a metro region, or a multi-postcode cluster. Pause paid media in that market for a defined period, ideally four weeks. Continue running paid media everywhere else under the Signal Discipline Playbook configuration. Track total revenue, new customers, and contribution margin in the holdout market and the control markets across the period. The gap between the two markets is the incrementality tax. If the holdout market loses 30 percent of expected revenue, the paid spend is producing 30 percent incremental revenue and 70 percent cannibalisation. If the holdout market loses 70 percent, the paid spend is producing real lift.

Day 61 to day 90 is metric replacement. Replace platform CPA as the campaign-decision metric with new-customer CPA derived from the geo-holdout. Brands that come through this protocol typically discover the true incremental CPA is 1.4x to 2.5x the platform-attributed CPA. The discovery is uncomfortable. It is also the only honest measure of how much the brand is actually paying to acquire a net new customer. The Signal Discipline Playbook reframes campaign decisions around this number. Campaigns that look lean on platform CPA but wasteful on incremental CPA get killed. Campaigns that look mediocre on platform CPA but strong on incremental CPA get scaled.

The geo-holdout is the part of the protocol most teams resist hardest. Pausing spend in a real market feels like burning money. The math says it is the cheapest way to discover whether the AI ad-platform algorithms are creating real value or just claiming credit for sales that would have happened anyway. Most brands run the holdout once a quarter for two to three weeks. That cadence is enough to keep the platform CPA chart honest.

From Blended ROAS To New-Customer Incremental CPA

The metric most paid-media teams report up to the executive team is blended ROAS or platform CPA. Both are aggregate numbers that mix new-customer and existing-customer conversions, and both are systematically inflated by cannibalisation when AI-driven campaigns are run on default settings. The Signal Discipline Playbook reframes the north-star metric as new-customer CPA versus geo-holdout incremental CPA. The two numbers should be close together. The closer they are, the healthier the paid-media program is. The wider the gap, the more cannibalisation is hiding inside the dashboard.

The brands that come through this protocol stop celebrating CPA wins that are not real. They start measuring customer-file growth as the primary signal of paid-media health, with incremental CPA as the cost-side check. The team meeting that used to centre on the blended ROAS chart becomes a quarterly review of incremental CPA against new-customer acquisition rate. The conversation gets concrete. The decisions get harder. The customer file keeps growing through platform algorithm shifts.

The worst paid-media decision a brand can make is to trust the platform's CPA dashboard without auditing whether the algorithm is finding new customers or harvesting existing ones. The Signal Discipline Playbook is the discipline that turns Meta Advantage Plus and Google Performance Max into honest acquisition tools rather than retargeting machines wearing acquisition badges. Anything else is a paid-retention budget dressed as paid-media spend, and the customer-file growth chart is the line item that exposes the fraud six months later.