How do you test Google Ads AI Max against DSA?

On 15 April 2026, Google announced AI Max was leaving beta and that three legacy settings — Dynamic Search Ads (DSA), automatically created assets (ACA) and the campaign-level broad match setting — would upgrade to AI Max automatically in September (Google, 2026). That post was updated on 11 June 2026, splitting the timeline in two.

Please note that campaigns using Automatically Created Assets (ACA) and campaign-level broad match setting will continue to be auto-upgraded starting in September 2026.

— Google Ads & Commerce Blog, 11 June 2026 update

The DSA sunset and auto-upgrade moved to February 2027; ACA and campaign-level broad match stayed in September 2026. On 5 August 2026, Google emailed affected accounts directly and fixed the date at 1 September (Search Engine Roundtable, 2026; PPC Land, 2026).

Two cohorts, two different defaults

• Campaigns using ACA: upgraded to AI Max with search term matching *and* text customization enabled by default (Google, 2026).
• Campaigns using the campaign-level broad match setting: upgraded with search term matching enabled by default, and nothing else (Google, 2026).

The distinction matters more than it looks. ACA governed creative generation only; search term matching governs which queries an ad is eligible for. Unless it is switched off at ad group level, the ACA cohort inherits automated query expansion alongside automated copy from 1 September (Google, 2026; Google Ads Help, 2026).

In the broad match cohort, text customization and final URL expansion are not enabled by default. That has a direct consequence on the cosmetics side: pinning in responsive search ads continues to be respected — a meaningful difference for advertisers who rely on pinning to hold compliance language in a fixed position (PPC Land, 2026).

Text customization — formerly automatically created assets — is a campaign-level setting that adds Google-generated headlines and descriptions alongside the ones you wrote. According to Google's documentation, generation combines two methods: "extractive" techniques that take snippets from landing page titles, descriptions and meta tags, and generative AI grounded in landing page content. Assets are reviewed and refreshed at least every 48 hours, and Google-generated assets serve only when predicted to outperform the headlines and descriptions you uploaded (Google Ads Help, 2026).

The mechanism in one line: your ad copy becomes a derivative of your site copy. In cosmetics that handover is regulatory-sensitive, because how much of a product page's wording is genuinely the brand's own decision is an open question — the picture we found when we audited 216 beauty brand websites is a reminder of that.

Türkiye's medicines and medical devices agency (TİTCK) defines a cosmetic claim as the information, statements or visuals used on the product's "label, packaging and/or promotional materials" (TİTCK, 2024). A search ad headline falls inside that definition. The guideline sets out six common criteria: legal compliance, truthfulness, evidential support, honesty, fairness, and informed decision-making.

The criteria generative copy pushes against by design

• Honesty. The guideline rejects superlative claims such as "unique, unmatched, best, most effective" for cosmetic products sharing the same properties. Treat those phrases as a specific filter when reviewing generated headlines (TİTCK, 2024).
• Truthfulness. The guideline's own example is explicit: a product with only short-term moisturising proven cannot claim "provides 48-hour moisturisation." If a duration or percentage on your landing page migrates into a headline, the evidential burden is yours (TİTCK, 2024).
• Fairness. The guideline disallows "paraben-free" on the grounds that it disparages the whole paraben group, which is safe when used in line with the regulation. If that phrasing sits on your site, it can surface in generated copy (TİTCK, 2024).

The borderline-products guideline draws a hard line: promotion of a cosmetic product must not aim to treat or prevent disease, claim or imply pharmacological, immunological or metabolic effects, or contain a health claim. It separately defines melasma as a hyperpigmentation disease, placing a "treats melasma" framing outside cosmetics regulation (TİTCK, Version 2.0).

Google's own documentation answers this indirectly but clearly. The text customization page warns the advertiser directly:

Ensure your website content is accurate, not misleading, and not in violation of Google Ads policies or applicable laws.

— Google Ads Help, About text customization in Search campaigns

On the regulatory side, responsibility does not sit with the platform either. The TİTCK guideline places compliance with the common criteria on the "responsible person" and requires the claim to be consistent with the documentation that person holds (TİTCK, 2024). Read together, the two documents point the same way: even when a model writes the line, the claim is yours and so is the burden of proof.

This pattern repeats wherever AI enters ad production. The same distinction comes up when weighing whether AI-generated videos actually work in advertising: delegating production is not delegating responsibility.

Every step below maps to a control documented by Google. None of it is guesswork about your account.

• Identify the affected campaigns. Only Search campaigns with ACA or the campaign-level broad match setting enabled are in scope. If neither is on, nothing changes for you on 1 September (Search Engine Roundtable, 2026).
• Set up text guidelines. A campaign-level feature, currently in beta. You can define up to 25 term exclusions and up to 40 messaging restrictions per campaign. Guidelines apply retroactively to existing Google-generated assets: non-compliant ones stop serving and are marked as removed (Google Ads Help, 2026).
• Remember term exclusions are language-specific. Per Google's note, excluding a term in one language will not prevent its equivalent in another from appearing. If you advertise in Turkish, the list has to be built in Turkish (Google Ads Help, 2026).
• Clean up your landing page copy. That is the source material. Google states that when a landing page changes, assets are refreshed within 48 hours — so corrections propagate at the same speed (Google Ads Help, 2026).
• Review brand settings. Brand inclusions make a campaign serve only for queries associated with selected brands, while brand exclusions stop it serving for queries associated with brands you want to avoid (Google Ads Help, 2026).
• Hold budget and bidding strategy steady. Changing the budget or bid target during the experiment confounds the DSA-versus-AI-Max comparison with another variable.
• Validate tracking templates. Required if you enable final URL expansion: static tracking URLs and non-standard lpurl usage can produce 404s or route users to the tracking destination itself (Google Ads Help, 2026).
• Put asset reporting on a schedule. Reading generated headlines regularly is not a set-and-forget task; compliance review only happens by eye.

A Google Ads AI Max vs Dynamic Search Ads experiment can use either of two methods: a custom experiment upgrades the copied trial campaign to AI Max, while the AI Max experiment template splits control and treatment traffic inside the existing campaign (Search Engine Roundtable, 2026; Google Ads Help, 2026).

Method 1: a custom experiment for DSA

• At the Make changes stage in the experiment draft, expand the Dynamic Search Ads settings and select Upgrade campaign. The treatment arm moves to AI Max when the experiment starts; the control keeps the current DSA setup (Search Engine Roundtable, 2026).
• After the test, Apply experiment can update the original campaign or convert the treatment into a new campaign. The second choice pauses the original campaign (Search Engine Roundtable, 2026).
• If auto-apply is enabled, the winning variation may be applied to the base campaign automatically (Search Engine Roundtable, 2026). Do not enable that option before human review of cosmetics claims is complete.

Method 2: the AI Max experiment template

Use Campaigns > Experiments > Campaign features and settings > AI Max for Search campaigns. In Google's AI Max experiment, traffic and budget are divided between control and treatment inside the existing campaign rather than through a copied campaign (Google Ads Help, 2026).

A campaign made entirely of dynamic ad groups—100% DSA—cannot use this template. In a mixed campaign with dynamic and standard ad groups, the treatment upgrades only the non-DSA groups to AI Max; dynamic ad groups remain dynamic (Search Engine Roundtable, 2026).

What should a beauty brand measure?

• Write the primary success metric and decision threshold before launch: conversion volume plus cost per acquisition or return on ad spend, and conversion quality.
• Keep budget, bidding strategy, dates, geography, and conversion definitions comparable across control and treatment; avoid another major change during the test.
• Review search terms, selected final URLs, and generated headlines weekly. Log treatment claims, unsupported duration or percentage claims, and prohibited superlatives separately from performance.
• Judge the winner using margin, returns or cancellations, new-customer rate, and the share of compliant assets—not statistical performance alone.

The test is not a regulatory filter. Google explicitly leaves the advertiser responsible for ensuring website content is accurate, not misleading, and compliant with applicable law (Google Ads Help, 2026). A performance winner should therefore not be applied to the base campaign before claim review.

There are three paths: move to AI Max on your own schedule, switch the legacy settings off entirely, or accept the default configuration that arrives on 1 September. There is no fourth path that preserves ACA or the campaign-level broad match setting in its current form (PPC Land, 2026). Google's recommendation is to migrate now in order to keep control of the setup (Google, 2026).

Keep performance expectations measured. Google reports that AI Max campaigns using the full feature suite — search term matching, text customization and final URL expansion — see an average of 7% more conversions or conversion value at a similar CPA/ROAS compared with using search term matching alone. The post's own footnote identifies this as Google internal data from 2026, covering non-Retail advertisers (Google, 2026).

Independent practitioner testing has reported weaker outcomes (PPC Land, 2026). It is worth noting those tests are agency and consultant account measurements, not peer-reviewed studies. Neither Google's figure nor the counter-measurements is a forecast for your account.

Structuring the migration as an experiment — a limited campaign first, measured on your own numbers — shrinks the regulatory risk and the performance risk at the same time. If you do not have the time to run that migration, our performance advertising service exists for exactly this kind of setup decision.