Audience Engineering: How to Steer Google's AI Targeting in Performance Max and Demand Gen
Google's AI now treats audiences as suggestions. Use the Seed, Score, Shape, Fence loop to steer Performance Max and Demand Gen.

For most of Google Ads' history, an audience was a boundary. You chose who could see the ad, and the platform delivered inside that line. In 2026 that boundary finished dissolving. Performance Max has treated audiences as hints since launch, and Demand Gen now treats Lookalike segments the same way. What is left for the operator is not targeting. It is engineering: deciding what the model learns from, what it is rewarded for, and where it is not allowed to spend.
This article lays out a four-part loop for doing that work deliberately: Seed, Score, Shape, Fence. It also settles a question most paid teams are asking right now, which is how demand gen vs performance max differs once both run on suggestion-based audiences.
Targeting Became Steering: What Changed in 2026
The shift did not arrive as one announcement. It arrived as a series of small ones, and each changed what an operator can and cannot control.
- March: Demand Gen Lookalike segments moved from similarity thresholds to audience suggestions, with an opt-out available by request. The narrow, balanced and broad reach setting no longer caps delivery.
- Late March: Performance Max first-party audience exclusions began rolling out. You can tell Performance Max to leave customer lists alone, as a steer rather than a wall.
- April 1: The Google Ads API stopped accepting new Customer Match adopters, making the Data Manager API the primary route. Existing integrations can keep running while they migrate; new builds go through the Data Manager API.
- June: Performance Max added a report showing audience segments by spend; from June 15 the Google Ads API stopped accepting new offline conversion import adopters. Visibility into audiences improved while the route for conversion feeds changed.
- August 17: Google began adjusting targets upward for overperforming campaigns across five campaign types, including Demand Gen. Stated tCPA and tROAS targets behave differently when a campaign is outperforming them.
Standalone Display campaigns are also being folded into Demand Gen, with Google confirming a migration tool from June and full transition expected during 2027. Audience strategy for Demand Gen now has to cover inventory that used to live in a different campaign type.
The misconception that costs the most
Most accounts still operate on one wrong default: audience signals are targeting. They are not. Google's own documentation describes performance max audience signals as guidance that helps the system find the right customers, and in Demand Gen they seed Lookalike segments that now work as suggestions. The model starts where you point it and goes wherever the data leads.
Operators fall into the default because the interface still looks like targeting. There are lists, segments and a place to attach them. The labels did not change, so the mental model did not either. If you believe a signal is a boundary, you will read off-list delivery as a bug and start adding more lists to fix it. That is the wrong response, and it is where wasted budget begins.
The Audience Engineering Loop
If the platform decides who sees the ad, your leverage moves upstream and downstream of that decision. Four levers remain, and each maps to a decision you control.
- Seed. What does the model learn from? Which customers define "good".
- Score. What is the model rewarded for? Which conversions and values count.
- Shape. What does the model see and say? Assets, structure and channels that steer delivery.
- Fence. Where is the model not allowed to spend? Exclusions, negatives and brand lists.
The order is deliberate. Score comes before Shape because a perfectly curated audience still optimizes toward whatever you reward. If the reward is a low-quality form fill, a pristine customer list will not save you. Seed comes first because it is the cheapest to audit. And it is a loop, not a line: the audience reporting that Google added in June feeds the next round of Seed, so each pass starts from evidence instead of instinct.
Run it in this order the first time. After that, treat it as a monthly review.
Seed: Feed the Model Your Best Customers
Seed is the audience data you hand the system: Customer Match lists, remarketing lists and the first-party segments around them. The model does not copy a list. It infers a pattern from it and looks for people who resemble that pattern. That makes the quality of the list far more important than its size.

Consider an illustrative case: a B2B services firm whose CRM holds every lead it has ever created. Uploading all of them teaches the model that unqualified form fills look like customers. Uploading only closed-won accounts, refreshed on a schedule, teaches a very different pattern. Same platform, same budget, different lesson.
What to do:
- Segment by value, not by existence. A list of everyone who ever purchased is weaker than a list of repeat or high-value buyers. Build tiers and test the top tier first.
- Keep lists fresh. Stale data teaches the model what your customers looked like a year ago. Set a refresh cadence and stick to it.
- Mind the processing time. Google notes that new or updated lists can take 24 to 72 hours to populate, and that new signals can take up to two weeks to be fully integrated. Plan changes around that window instead of judging results early.
- Use the Data Manager API for Customer Match. Google now recommends the Data Manager API for Customer Match workflows and advises against building new ones on the Google Ads API. If an older pipeline still feeds your lists, confirm it is running and plan the migration.
The trade-off: a narrow, high-value seed gives the model a clearer pattern but can limit scale. A broad seed gives volume with a noisier pattern. Start narrow in accounts where lead quality matters more than lead count, and widen only when volume becomes the constraint.
For the customer match google ads workflow specifically, the work is less about the upload mechanics and more about deciding which customers deserve to define the pattern. That is a business decision, not a technical one.
Score: Fix the Conversion Signal Before the Audience
This is the stage most teams skip, and it is where the largest gains tend to sit. The audience problem is frequently a conversion problem in disguise. The model optimizes toward whatever you tell it counts as success. If the signal is thin, delayed or wrong, no audience strategy can compensate.

Three pieces matter.
1. Capture the conversion reliably. Enhanced conversions let you supply hashed first-party data alongside conversion tags so Google can match conversions to the ad interactions that drove them. For anyone researching enhanced conversions google ads, the practical question is whether your tags are passing clean data on every conversion path, including forms, calls and checkout. Our guide to pixel vs Conversion API tracking covers the server-side mechanics behind that signal.
2. Bring offline outcomes back in. For lead-generation businesses, the conversion that matters often happens in a CRM weeks later. Importing qualified leads and closed revenue gives the model a far better target than a raw form submission. The rules changed here: from June 15, 2026, the Google Ads API stopped accepting new adopters of offline conversion imports, including enhanced conversions for leads, and Google now treats the Data Manager API as the primary route. Teams already importing through the Ads API can keep going while they migrate, but treat the old path as transitional and verify that your import still runs.
3. Assign value, not just volume. Treating every conversion as equal tells the model that a ten-dollar lead and a ten-thousand-dollar account are the same thing. Value-based signals let it distinguish them.
The trade-off: the richer the signal, the later and thinner it arrives. Closed revenue is the best target and also the slowest and lowest-volume one. A common compromise is to optimize toward an intermediate milestone, such as a qualified lead, that arrives fast enough to train on and correlates with revenue. Pick the milestone by checking how well it predicts closed deals in your own data.
Shape: Steer With Assets, Structure and Channel Controls
If Seed is what the model learns from, Shape is what it sees and says. In a suggestion-based system, creative and structure carry targeting weight. An asset group built around one offer for one buyer type tells the system who the campaign is for more clearly than a list attached to a mixed bag of messages.
Where steering happens now:
- Asset groups as audience hints. Theme each asset group around a single audience and offer, and attach the matching signal. Mixed groups blur the instruction.
- Audience reporting. Performance Max now offers performance breakdowns by audience segment, including age range and gender, which lets you judge whether the steering is working and adjust assets accordingly.
- Brand lists. You can exclude selected brand traffic at the campaign level with a brand list, which shapes the search side of delivery.
- Channel selection in Demand Gen. Unlike Performance Max, Demand Gen retains channel selection, so you choose where ads run, which is a form of targeting in its own right.
The trade-off: every additional asset group splits learning across more pieces. Split when audiences and offers genuinely differ, and consolidate when they do not. An account with six near-identical groups is not more precise than one with two distinct ones. It is slower to learn.
Fence: Set Boundaries That Actually Hold
Fencing is where operators regain the control they feel they lost, with one important caveat: most fences are directional, not absolute.

The tools available:
- First-party audience exclusions. You can exclude customer lists from Performance Max, which helps acquisition-focused campaigns avoid paying to reach people you already have. Search Engine Journal's coverage frames this as directional steering rather than rigid targeting. Treat it as a strong nudge. List matching is also imperfect: Google may match only part of an uploaded list, so some spend can still reach people you meant to exclude.
- Account-level placement exclusions and brand lists. A single block list now spans Performance Max, Demand Gen, YouTube and Display, so one set of exclusions protects every campaign type.
- The Demand Gen Lookalike opt-out. Google gave advertisers a way to keep Lookalikes as a targeting constraint instead of a suggestion. This is a real decision, not a default to accept or reject by reflex.
On that last point, the choice is control versus reach. Keeping the constraint preserves a hard boundary and costs you the model's ability to find similar people outside the list. Allowing suggestions gives the system room to expand and costs you predictability. A brand with a tightly defined buyer and a limited budget may prefer the constraint. A brand that needs scale usually benefits from suggestions, paired with strong exclusions.
Expect a dip when exclusions go live. Removing existing customers from reach can reduce conversion volume at first, because the system loses some of its easiest conversions. That is not automatically a failure. It is the price of shifting spend toward new customers, and you should decide up front whether you are willing to pay it.
Performance Max vs Demand Gen: Where Each Lever Applies
Both campaign types now run on suggestion-based audiences, which makes the demand gen vs performance max question less about targeting and more about role. The two have also drifted closer together, with Demand Gen moving down the funnel and overlapping more with Performance Max.
- Inventory. Performance Max: Every Google surface, including Search and Shopping. Demand Gen: YouTube, Discover, Gmail and related surfaces; excludes Search.
- Channel control. Performance Max: Reporting and exclusions, no direct channel picker. Demand Gen: Channel selection retained.
- Audience behavior. Performance Max: Signals guide the system; not hard targeting. Demand Gen: Signals guide the system; Lookalikes are suggestions by default.
- Optimization scope. Performance Max: End to end across surfaces. Demand Gen: Creative-led reach and consideration, increasingly down-funnel.
- Where Score matters most. Performance Max: Conversion quality across all surfaces. Demand Gen: Conversion quality and view-through behavior.
- Where Fence matters most. Performance Max: Customer-list exclusions and negatives. Demand Gen: Lookalike constraint decision and placement exclusions.
Our position: do not treat this as either/or. Use Demand Gen where creative can create demand and channel control matters, and Performance Max to capture it across every surface. The risk is overlap, where both campaigns bid for the same people. Shared fences, consistent Seed lists and a common Score reduce that overlap. Confirm current exclusion availability by campaign type in your own account before building a plan around it, since rollouts have varied. If Google is only one channel in your mix, Diversify or Die: A Cross-Platform Paid Media Strategy Beyond Google and Meta covers where the rest of the budget belongs.
How to Tell If the Steering Worked
Steering is only useful if you can measure it. The audience-segments-by-spend report in Performance Max gives you the first direct view of where budget goes by segment, alongside the demographic breakdowns Google added.

A practical test:
- Set pass and fail criteria before changing anything. For example, define what share of spend should reach new customers and what conversion volume you will tolerate losing.
- Record a baseline. Capture spend by segment and conversion quality for the period before the change.
- Change one lever at a time. If you alter Seed, Score and Fence together, you cannot tell which one moved the result.
- Wait out the learning window. Google says new signals can take up to two weeks to integrate. Judging earlier measures noise.
- Compare share of spend with share of conversions. A segment taking a large slice of spend with a small slice of conversions is the first place to look.
Treat the report as directional evidence, not proof. It shows where money went, not why a customer converted. To judge results beyond platform reporting, pair it with an attribution model that fits your sales cycle; Marketing Attribution 101 walks through the options.
Mistakes That Quietly Waste Budget
- Stacking overlapping lists. Several near-identical lists do not add precision. They add overlap, which muddies the pattern the model learns.
- Excluding too broadly. Aggressive exclusions strangle volume, and the system has fewer people to find. Start specific and expand based on data.
- Changing signals mid-learning. Every edit restarts part of the clock. Constant tweaking keeps the campaign in permanent learning.
- Optimizing audiences on broken conversion data. This is the most common and most expensive error. Audit Score first.
- Treating the Lookalike opt-out as free. Keeping the constraint trades reach for control. Make that trade on purpose.
- Ignoring the Aug 17 change. If your tCPA or tROAS targets were conservative and the campaign was outperforming them, the behavior may have shifted. Review targets against actual performance.
A 30-Day Implementation Sequence
You do not need to rebuild the account. You need to move through the loop in order.
Week 1: Audit. List every audience signal, exclusion and conversion action in each Performance Max and Demand Gen campaign. Confirm whether Customer Match uploads and offline imports still run through the older Google Ads API path, and plan the move to the Data Manager API. Note anything that overlaps. New to how these campaign types fit together? Start with Paid Media 101: A Starter’s Guide to Advertising Success.
Week 2: Score. Fix the conversion signal. Verify enhanced conversions, restore or build the offline import path, and set value rules. Do not touch audiences yet.
Week 3: Seed and Shape. Rebuild Customer Match lists around value tiers. Reorganize asset groups so each pairs one audience with one offer. Make one change at a time and log it.
Week 4: Fence and measure. Add customer-list exclusions and shared brand and placement exclusions. Decide on the Lookalike constraint. Then read the audience-segments-by-spend report against your baseline and decide what the next loop should change.
Repeat monthly. Each pass should start from what the previous report showed.
Get a Second Set of Eyes on Your Audience Setup
If you want an outside review of how your accounts handle Seed, Score, Shape and Fence, Refinity.io runs a Performance Marketing Audit that covers audience signals, conversion tracking and exclusions across Performance Max and Demand Gen. Book a call to walk through your setup.
- Performance Max
- Demand Gen
- Google Ads
- Customer Match
- Enhanced Conversions
- Audience Targeting
- Paid Media




