Google Says AI Audience Expansion Finds Extra Buyers. Only a Holdout Can Prove It

The way Google Ads AI audience expansion reporting works today is that it never tells you whether all of the extra buyers it found would have bought anyway. This is exactly the gap behind Google's help page on optimized targeting, which says the conversions it brings in "are additional conversions that you wouldn't otherwise have received." In order to truly test the incrementality of that claim for your own ads, you would need to run a holdout, keeping one group of people from seeing the ads and comparing its conversions with the group that saw them. That is what Google has for this already, called Conversion Lift, but it has its own limitations. It only works with campaigns that have at least 1,000 conversions that Google observed, and also requires a minimum campaign budget of $5,000, according to Search Engine Land coverage from October 7, 2026.

Google's converter profile, according to its About optimized targeting page as it read on October 12, 2026, can draw on "Google searches for specific running shoe brands," and its Lookalike guidance asks for your top converting search terms. This query level truth is a better signal than the platform defaults, and your converting search terms are one you own outright.

Where is AI audience expansion in Google Ads on by default?

Optimized targeting, the setting that treats your audiences as starting points for Google Ads to serve likely converters beyond them, is by default enabled wherever Google offers it. See "About optimized targeting" in the Google Ads help center, which says: "Optimized targeting is automatically turned on for all campaigns." It's enabled for Display and Demand Gen campaigns as well as for Video campaigns with conversion or traffic goals. For Performance Max campaigns and for App campaigns, Google's setup page adds, it's always on. Where you can control it, you turn it on and off with a checkbox in the ad group's Settings tab under "Optimized targeting".

The way Google currently describes your audiences is as targeting signals, and it says optimized targeting "may reduce or stop serving traffic on your signals" when better traffic turns up elsewhere, while Performance Max calls its audience signals "audience suggestions" and says the campaign "may show ads to relevant audiences outside of your signals." So, for example, in Demand Gen campaigns, optimized targeting also switches on demographic expansion, and age, gender, household income and parental status are treated as targeting signals too, with restriction offered for age and gender.

The limit for Demand Gen audiences of how far you can widen a Lookalike audience became a suggestion by default, rather than a cap, from March 2026. Now, Google's help page says, the Lookalike segments for Demand Gen audiences "will evolve from being limited to similarity thresholds to audience suggestions". The settings for Narrow, Balanced and Broad, which used to cap reach at 2.5%, 5% and 10% of users in the target location, all now work as signals (as opposed to caps), and keeping the old behavior means filling in Google's opt out form. Note that Google says removal takes 3 days to take hold. The reader should also know that Display campaigns are moving into Demand Gen campaigns, using a migration tool that Google started to phase in during June 2026.

What does Google say expansion adds?

As an added note to its good practice for YouTube video campaigns, Google's Demand Gen best practices page, "The YouTube Performance Four," reports that advertisers who adopted at least three of the four recommended practices, optimized targeting among them, saw "over 40% more conversions and +30% higher conversion value" on average, per Google internal data covering April 2024 to December 2025. Note that this figure describes self selected adopters with no stated method, and the targeting setting is embedded within the bundle of practices, so it can't be separated out.

For Demand Gen Lookalike audience suggestions, Google states that the change is "designed to drive clear improvements and uplift in conversions compared to the original version." However, there are no conversion numbers to back up this claim.

The strongest case for enabling expanded reach of your audiences is found on Google's setup page, which says "The conversions that you get through optimized targeting are additional conversions that you wouldn't otherwise have received." That page also warns that your initial average CPA may fluctuate and tells you to wait for 50 conversions or 2 weeks before judging a new campaign, yet it gives no account of how the additional conversions claim was established.

Why can't the expansion row tell you whether those buyers are new?

The "Total: Expansion and optimized targeting" row, which Google's setup page sends you to in the Audiences report, only shows the conversions attributed to ads served through expansion, and Google's overview page says it holds that traffic "along with select traffic experiments run by Google." You cannot tell from this how many of those conversions happened because the ads were shown to these extra people. You need a holdout to tell you that.

Demand Gen Lookalike audience suggestions also get wrapped into the same reporting and muddy it further, since Google's FAQ says that with both features on, "you would find performance shift from Optimized Targeting towards Lookalike segment in the Campaign reporting table." This is indicated by a "signal" tag on the Lookalike row in that table, and a row from January 2026 and one from April 2026 can carry the same Audience name while describing different targeting settings.

Observational measurement tends to make ads look better than they really are. A study from 2019 by Northwestern University's Florian Zettelmeyer, Brett Gordon, and two Facebook researchers checked 15 US advertising experiments at Facebook, 1.6 billion impressions in all, and in Zettelmeyer's summary the common methods of measurement used by advertisers "overestimate ad effectiveness relative to what we found in our randomized tests," though in some cases they significantly underestimated it. The reason this matters is that Expansion looks for the people most likely to convert, and some of those people would have purchased anyway.

Google's new customer acquisition goal has a gap in how it determines whether a user is a new customer or an existing customer. For example, the way it can determine whether a user is a new customer or not is based on up to 540 days of past campaign activity and corresponding tracked purchases. Also, Google says that method "is less comprehensive because users can remove cookies and opt out," and it warns that an uploaded customer list of less than 1,000 members may limit the new customer acquisition goal.

What does independent research say about widening a lookalike?

In the 2022 nonprofit study K. Sudhir and Seung Yoon Lee of Yale (with Subroto Roy) ran, where performance meant donations and brand meant click through, the authors tested "audience expansion strategies by lowering match ranks between the seed and lookalikes to increase acquisition reach" for performance marketing and brand marketing. According to the findings of the study published as Cowles Foundation Discussion Paper 2302R in June 2022, "The drop in effectiveness with lower match rank range is much greater for performance marketing than for brand marketing." The authors moreover noted that third parties "can indeed identify factors unobserved to the advertiser" even in a cold start, and that seeding on people further along in the journey improved donations.

This is one discussion paper in one setting, so the findings do not transfer to your context directly, though they back lookalikes built with care, and the match rank result cautions against reaching further down the ranks, which suggestion mode now permits.

What is the strongest case for leaving expansion on?

The strongest case for expansion is the cold start. By reading the keywords on your landing page and in your creative assets, Google says, optimized targeting helps ad campaigns with no prior audience history still find converters. The manual segments are often little more than guesses at who might convert well too (in Google's rafting example, the system stopped serving on the signal "traveling families," which never performed, and instead added "white water rafting," which did).

There is also a cost associated with hard fences, as Google's Customer Match policy restricts the "Targeting" setting to accounts with 90 days of history and more than $50,000 in lifetime spend. List eligibility is also restricted to audiences with at least 100 members added or updated within the last 540 days.

The strongest case for a feature cannot include statistics that Google's help pages do not cite. They cite no holdout result for optimized targeting or Lookalike suggestion mode, so we have no idea how many of the conversions in "Total: Expansion and optimized targeting" would have occurred even without the ads.

Where does query data fit into audience expansion?

As noted above, Google's page on keeping Lookalikes as a targeting constraint recommends that the Lookalike be used in conjunction with a custom segment built from "10-15 of the top converting search terms." The Performance Max signals page offers similar advice, telling advertisers to "use search term based custom segments."

That is the queryDNA principle at work, and what Google reports back for your optimized targeting is audience insights on the top performing Google segments. Search inventory gets more than that. Google began rolling out full search terms reporting for Performance Max campaigns on April 30, 2025, with campaign level negative keywords and brand exclusions to act on it. Where Performance Max buys search traffic, that report shows which queries it paid for, as we reviewed in our Performance Max black box audit.

How do you check whether expansion found buyers or just reach?

Start by opening up the Audiences report for the individual campaigns (Display, Video and Demand Gen). Find the row for "Total: Expansion and optimized targeting" and set it beside your individual named segments. Then look at the "Cost per Conversion" column for both over at least the 2 weeks Google asks for, as a screen only. The conversion figures and the Cost per Conversion figures for both the expansion row and your named segments are attributed. The "Total: Expansion and optimized targeting" row also contains experiments run by Google. With Lookalikes in Suggestion Mode, read the "Lookalike" audience row as well, since Google says performance moves across to that row from optimized targeting.

Then check which conversion the row produces, and judge it on the one that you are paid for. If your CRM knows which leads closed, import that stage from your CRM, as we explained in conversion value from the CRM. If you set up a new customer acquisition goal, upload a customer list to it on top of the autodetection that Google does for you.

Make sure you are excluding obvious nonbuyers in "Where Ads Showed" for Demand Gen, and in "Search terms" for Performance Max campaigns.

Then run a holdout built around the expansion. Because Conversion Lift reports on participating campaigns, a study that mixes in other signals with the expansion can't split out the share of conversion lift due to expansion. So give the expansion a campaign of its own, as Google's gaming PC example does by running optimized targeting alone. Leave the audiences and rest of the creative unchanged for the period of the study. Also, be sure to hold off on the Display migration tool in the middle of the study, as Google advises. Note that Google says Conversion Lift is not available to all Google Ads accounts, and Google recommends a minimum of 14 days for Conversion Lift studies. But as PPC Land noted in a November 13, 2025 post on Google's documentation, shorter Conversion Lift studies saw up to 17% drops in absolute lift for businesses with long conversion lags. So, since that report also finds higher lift on page views, lead form submissions and add to carts than on purchases, be sure to keep the action that you get paid for in the study. Our piece on AI driven marketing mix models looks at where to place lift tests in relation to modeled attribution.

Should you turn AI audience expansion off?

Continue to use AI Audience Expansion in scenarios where a campaign has little to no history, pointed at the conversion you get paid for. However, the help page sentence calling every optimized targeting conversion additional has no test attached, so whoever is accountable for the budget tests it before scaling: accounts that clear the Conversion Lift bar run the holdout before raising spend, and below that bar read the expansion row only as a screen, on closed revenue past Google's 2 weeks. Compare the additional conversions to your existing named segments and test turning off the Optimized Targeting for scenarios where they perform worse. For Performance Max campaigns, where optimized targeting stays on, work the Search terms report instead. The Lookalikes for Demand Gen campaigns have begun to transition to Suggestion Mode by default as of March 2026. For these campaigns, the only way to opt out of the Suggestion Mode is a special opt out form. Because the system is optimized for the conversion that you handed it for, determining what constitutes a buyer and whether or not additional reach added value is up to the people who manage the account and answer for the spend. This is the same line we draw on the other automation problems that the dashboard hides.