The AI Max Field Guide: What Google Gets to Decide, and What It Has to Earn

AI Max is not one decision. It is three: whether Google may expand your queries, whether it may write your ad text, and whether it may pick your landing pages. Each is a separate grant of decision-making with its own controls and its own failure modes, and from September 1, 2026 Google makes the first grant for you on any Search campaign still running campaign-level broad match or legacy automatically created assets. We audited all 19 Google Ads accounts under our management on August 25: zero enabled campaigns use campaign-level broad match, 25 campaigns still carry the legacy asset setting, and exactly one enabled, spending campaign sits in the auto-conversion path. This guide covers the placement spectrum we use (A, expansion engine, through B, precision defense), the seven gates a campaign passes before it earns more automation, and what AI Max actually served on our own accounts. It closes with the four GAQL queries that reproduce our audit on any account and a machine-readable appendix your own AI agent can lift whole.

The mechanics of the September migration itself (cohorts, timeline, opt-outs, and six theories of Google's motivation, verified against 66 dated sources) live in the companion dossier, AI Max by Default: The Verified Dossier on Google's September Auto-Upgrade. This piece is the operating manual.

Evidence legend

• ● Verified fact. A documented product mechanic, confirmed against the cited Google page on the access date.
• ■ Company claim. Google's own performance number, published without disclosed methodology.
• ◆ Independent measurement. Third-party data with a named method and sample.
• ◇ First-party data. Measured on accounts under our management, pulled from the Google Ads API on August 25, 2026.
• △ Inference. Our reading of the evidence. Could be wrong.

One toggle, three grants of decision-making

● AI Max is an optimization layer inside existing Search campaigns, not a campaign type. Google's setup page states it plainly: "AI Max isn't a new campaign type but an optimization layer within existing Search campaigns." The layer bundles search term matching with asset optimization, and asset optimization contains text customization and final URL expansion. (Google AI Max setup, accessed August 25, 2026)

Each component is a distinct grant:

1. Search term matching: the query grant. ● With it on, "all keywords are treated as broad match by default unless search term matching is disabled at the ad group level," and Google adds keywordless expansion from your landing pages on top. (Google Ads API AI Max guide, updated August 19, 2026)
2. Text customization: the copy grant. Google generates headlines and descriptions from your site, existing ads, and keywords. ● Generated assets are marked "Google AI" in the Added by column and can be removed, though not in bulk inside the interface. (Google AI Max reporting, accessed August 25, 2026)
3. Final URL expansion: the destination grant. Google selects landing pages beyond your final URLs. This turns your website into a targeting feed, which is why the site audit gate below exists.

● The grants are configurable but not fully independent: "In order to use the Final URL expansion setting, Text customization must be selected. Deactivating text customization automatically disables Final URL expansion." (Google AI Max setup, accessed August 25, 2026)

● Google's own migration defaults treat the components separately, which is useful precedent for treating them as separate decisions: a campaign-level broad match campaign converts with search term matching on and text customization off, while a legacy ACA campaign converts with search term matching on and text customization on. Final URL expansion stays off for both. (Google Ads Developer Blog, published August 12, 2026)

The controls map

Every control has a level, and the level decides who can contain a problem when one appears. ● All rows sourced from Google's How AI Max works page and the API guide, accessed August 25, 2026.

● One flag almost nobody discusses: bundling_required. The API documents it as the way to "determine if AI Max for Search campaigns must be enabled to respect or modify text customization and brand list controls." On a campaign where it reads REQUIRED, you cannot hold those component settings without the umbrella. ◇ It reads REQUIRED on 13 of our 18 enabled Search campaigns, so for most of our fleet AI Max is a package deal at the campaign level and the real containment tools are the ad-group and list controls above.

September makes the first grant automatic

From September 1 through 30, Google converts every Search campaign still using campaign-level broad match or standalone automatically created assets into AI Max, in place, with the asymmetric defaults described above. Creation of new legacy configurations was already blocked on August 3. September 1 is the deadline to prevent the automatic mapping; under current documentation the AI Max settings themselves remain changeable after conversion. The full mechanics and opt-out paths are in the dossier.

Eligibility on real accounts, measured the week before the window, looks nothing like the headlines suggest. ◇ Across the 19 Google Ads accounts under our manager account, on August 25, 2026:

• Zero of 18 enabled Search campaigns use campaign-level broad match. The setting exists on 2 paused campaigns fleet-wide.
• The live trigger is the other one: legacy automatically created assets. 25 campaigns still carry ACA opted in, and most of them are paused historical campaigns nobody looks at.
• 14 campaigns already carry the AI Max flag, 2 of them enabled: our own agency campaign and our own wraps-shop campaign, both deliberate tests on our own money.
• The remaining auto-conversion path holds 3 campaigns, and exactly 1 is enabled and spending: a client location campaign that did $1,199.20, 284 clicks, and 22 conversions in the last 30 days. That one campaign is the entire September decision for this MCC, and we are making it deliberately before the window opens.

△ The migration language does not exempt paused campaigns, so the quiet risk in managed accounts is the paused campaign that converts silently in September and wakes up months later carrying AI Max settings nobody chose. Sweep everything, including what is not currently spending.

The spectrum: from expansion engine to precision defense

The strongest AI Max campaign is the one that gives Google the widest decision space the advertiser can safely measure, govern, and fund. Fit for expansion is what the spectrum measures, and enthusiasm for automation is what it deliberately ignores. We place every Search campaign at one of six positions:

Campaigns earn their way toward A through controlled component tests. Nothing starts at A because the umbrella toggle exists, and nothing is stranded at B forever if the gates below start passing.

The seven gates

A campaign passes gates to move up the spectrum. Record evidence at each one, not just a yes.

Gate 1: Measurement integrity. AI Max optimizes to the conversion signals it receives, and it cannot tell a valuable customer from a junk form fill unless your account sends that distinction back. Are qualified outcomes imported? Does value reach Google before most optimization decisions are made? Are spam, duplicate, and unserviceable leads excluded or devalued? Green means qualified value flows reliably: the campaign may move toward A. Yellow means partial or lagging feedback: stay in the middle positions. Red means bidding runs on raw form counts: stay near B until measurement is repaired.

Gate 2: Economic room. Expansion only pays when the business can absorb more qualified demand at similar margin. Hard capacity limits, mixed-margin catalogs, and defense-only objectives all argue for B-side positions regardless of how good the data looks.

Gate 3: Query opportunity. Expansion needs somewhere to expand into. If valuable demand hides in long-tail, conversational, or emerging queries your keyword set misses, search term matching has a job. If the existing keywords already cover the profitable demand, the grant buys risk without upside.

Gate 4: Creative tolerance. Could a generated headline misstate price, eligibility, location, or offer terms? Who removes an inaccurate asset, and how fast? The hard-stop question: if a generated ad could create legal, patient-safety, or severe brand risk, why is text customization on at all?

Gate 5: Site fitness. Final URL expansion treats the website as a feed, so audit it like one before enabling it. Old offers, thin pages, careers pages, staging content, and unserviceable location pages are all destinations Google can pick unless exclusions cover them.

Gate 6: Brand, geography, and policy. Decide own-brand treatment, competitor treatment, and jurisdiction limits as business rules, and encode them as brand lists, negatives, and location controls with a named owner.

Gate 7: Governance. Someone must check settings after launch and after migration, review queries, assets, and landing pages on a cadence, and hold rollback authority. If nobody can observe and act on the additional degrees of freedom, do not grant them.

The scorecard

Score every gate 0, 1, or 2: fails, partial, or passes with recorded evidence. Fourteen is the maximum, and the total maps to a starting position on the spectrum:

Two overrides beat any total. Gate 1 at 0 caps the campaign at B no matter what else passes, because every grant downstream would optimize to signals that are wrong. And a Gate 4 hard-stop keeps text customization off at any score, because no conversion lift prices a compliance incident correctly.

What the independent record shows

The public evidence splits, and it splits along lines the spectrum predicts.

• ■ Google's April 2026 material reports the full AI Max suite produced 7% more conversions or conversion value at similar CPA or ROAS than search term matching alone, from internal non-Retail data with no disclosed sample size, test length, or significance method. (Google, published April 15, 2026)
• ◆ Smarter Ecommerce measured a median 13% increase in conversion value, a median 16% increase in CPA, and a median ROAS difference of 0% across more than 250 retail Search campaigns. Scale went up; efficiency did not. (Smarter Ecommerce, published March 23, 2026)
• ◆ Location3 ran AI Max as Google Experiments across five locations of one home-services franchise brand from December 1, 2025 to January 31, 2026 and reported conversion lifts between +97% and +720%, with CPA flat to improved 63% depending on location. (Location3, published March 18, 2026)
• ◆ Brad Geddes of Adalysis found "AI text customization wasn't as effective as human management of assets for highly optimized campaigns," while "AI performed quite well for the long-tail campaign" where human optimization had been limited. His B2B test lost audience qualification badly enough that the company "stopped the tests and went back to pinning its ads" after three weeks. (Search Engine Land, published July 28, 2026)
• ◆ Lunio's classifier put AI Max invalid-traffic rates at 2.46% in Q4 2025, 4.01% in Q1 2026, and 5.28% in Q2 2026 across about 49.5 million AI Max clicks, while standard Search ended Q2 2026 at 3.07% in the same analysis. Traffic quality needs its own monitoring line, separate from CPA. (Lunio, updated August 14, 2026)

△ Read together: expansion helps where coverage was weak (franchise locations, neglected long-tail ad groups) and struggles where humans already optimized hard (mature retail portfolios, pinned B2B copy). That is a component-fit story, and it is why the placement question beats the adoption question.

What AI Max served on our own money

We run AI Max deliberately on two first-party campaigns, so we can watch the grants exercise themselves without a client's budget in the blast radius. ◇ Both readings are from the Google Ads API's ai_max_search_term_ad_combination_view, the report that shows exactly which search term, headline, and landing page AI Max combined, over the last 30 days as of August 25, 2026.

On our agency account's Search campaign (AI Max on, text customization on): 1,346 impressions, 36 clicks, $115.32 in spend, and the combination report recorded zero AI Max combinations. △ On a tightly themed campaign with full keyword coverage, the expansion engine found little room; low-volume rows can also sit below search-terms reporting thresholds, so we read this as minimal expansion.

On our vehicle-wraps campaign, the report recorded 4 combinations totalling 5 impressions and 0 clicks, and all four are instructive:

• The query "billboard sign" was served the headline "Serving Edmonton & Area" and landed on the fleet decals page. Billboard intent, decals destination: an off-catalog match a keyword plan would never have made.
• The query "nd graphics edmonton", a local competitor's brand, matched twice with headlines "Edmonton Vinyl Lettering | Edmonton Pros Since 2010" and "Vehicle Wraps & Vinyl Graphics | Edmonton Truck Wraps & Decals". Competitor conquesting happened by default, which means brand exclusions are a decision you make, not a default you inherit.
• The query "work van decals" matched cleanly to the truck-wraps page with "Free Wrap Consultation | Professional Vehicle Branding". This is the grant working as sold.

Five impressions decide nothing about performance. What they demonstrate is the operating surface: every expansion is inspectable down to the term, the headline, and the URL, and the review takes minutes when the campaign is small and hours of triage when it is not. Build the review habit before the volume arrives.

◇ For context on where our fleet earns its results today: across all 19 accounts in the last 30 days, keyword-attributed Search spend split $6,616.50 on phrase (55.3%), $3,887.76 on exact (32.5%), and $1,467.31 on broad (12.3%) of $11,971.57 total. △ Expansion has to beat a precision baseline that already converts, which is exactly why it gets granted per campaign, per component, on evidence.

The operating rhythm

Four reports carry the whole monitoring load, and all four exist today.

1. Search terms, filtered to match type "AI Max", with the Source column added. ● This is Google's documented path for inspecting expanded traffic. Log spend share, not just term counts, because low-volume queries stay hidden under the standard privacy threshold. (Google AI Max reporting, accessed August 25, 2026)
2. Assets, filtered to "Google AI" in the Added by column. Review generated text against your claims inventory; remove anything that misstates the offer.
3. Landing pages, with the Selected by column. ● The column indicates when final URL expansion picked the page. Compare chosen pages against your approved set.
4. Lead quality, from your own CRM. Volume without qualification is the failure mode the platform cannot see unless you send it back.

Cadence: check settings the day after launch or migration, review weekly through the first month, then settle to the cadence traffic justifies. Set rollback triggers in advance: a generated claim error, an unsafe landing page, brand leakage past an agreed share, or invalid-lead rates past an agreed line each trigger the relevant grant's revocation, not a debate.

Failure modes worth naming

• Irrelevant expansion: matched demand you cannot serve or do not want, caught in the Source-labeled terms report.
• Generated claim errors: copy that misstates price, eligibility, or geography, caught in the Added by report and removed by a named owner.
• Unsafe destinations: traffic sent to pages that were never meant to sell, prevented by URL exclusions and caught by Selected by.
• Brand leakage: brand and competitor traffic blurring nonbrand results, prevented by brand lists and separate brand reporting.
• Invalid traffic: the Lunio numbers above; watch it as its own line.
• False attribution: a campaign claiming conversions another campaign would have won anyway. Evaluate at account level, and treat reattribution as noise, not growth.

Graduation rules

• A campaign moves toward A only on evidence of account-level incremental qualified value from a controlled test, one component at a time.
• ● The testing tooling is improving on schedule for the migration: Google announced multi-campaign A/B experiments rolling out in September 2026, alongside Performance Planner support for previewing bidding and budget changes. (Google, published August 20, 2026)
• A campaign retreats toward B the moment a rollback trigger fires, and re-earns the grant with a fixed gate, not a memory of the incident.
• When a test is inconclusive, the component stays off. The burden of proof sits with the automation, because the precision baseline is already paying the bills.

The preflight: four queries that audit any account

Every field name in this guide is queryable, and the four requests below are the ones we ran for the fleet audit above, each verified against Google Ads API v22 on real accounts on August 25, 2026. All four are read-only. Run them through the API, a script, or the agent that helps manage your account; the interface shows the same facts, but the API names are exact and the output diffs cleanly week to week.

◇ 1. September eligibility and current posture, one row per Search campaign. BROAD under keyword_match_type marks the campaign-level broad match cohort. TEXT_ASSET_AUTOMATION at OPTED_IN marks the legacy ACA cohort. enable_ai_max shows what already converted or opted in, and bundling_required at REQUIRED tells you the component settings cannot hold without the umbrella.

SELECT campaign.id, campaign.name, campaign.status,
  campaign.keyword_match_type,
  campaign.asset_automation_settings,
  campaign.ai_max_setting.enable_ai_max,
  campaign.ai_max_setting.bundling_required
FROM campaign
WHERE campaign.status != 'REMOVED'
  AND campaign.advertising_channel_type = 'SEARCH'

2. Ad-group opt-outs on campaigns already running AI Max. An ad group whose row comes back without the flag has search term matching active: the API omits unset fields rather than returning false.

SELECT ad_group.id, ad_group.name,
  ad_group.ai_max_ad_group_setting.disable_search_term_matching
FROM ad_group
WHERE campaign.ai_max_setting.enable_ai_max = true

3. Every combination AI Max served: the term, the headline, and the landing page. The competitor-conquesting example above came out of this report. It is the single most useful monitoring surface AI Max has, and almost nobody pulls it.

SELECT ai_max_search_term_ad_combination_view.search_term,
  ai_max_search_term_ad_combination_view.headline,
  ai_max_search_term_ad_combination_view.landing_page,
  metrics.impressions, metrics.clicks
FROM ai_max_search_term_ad_combination_view
WHERE segments.date DURING LAST_30_DAYS

4. Who changed what during the window. client_type and user_email separate your team's changes from rule-based, API, and Google-initiated ones, which is how you confirm what the migration wrote to your account. Three traps, all hit live: change_event rejects relative ranges like DURING LAST_30_DAYS, requires a LIMIT, and refuses start dates more than 30 days back (the error is START_DATE_TOO_OLD). Schedule this audit inside the month it matters.

SELECT change_event.change_date_time, change_event.change_resource_type,
  change_event.changed_fields, change_event.client_type,
  change_event.user_email
FROM change_event
WHERE change_event.change_date_time >= '2026-09-01'
  AND change_event.change_date_time <= '2026-09-30'
ORDER BY change_event.change_date_time DESC
LIMIT 200

The September checklist

1. Inventory eligibility across every account you touch. Preflight query 1 returns both triggers and the current posture in one pass per account. Keep the output: it is the before picture your post-migration verification diffs against.
2. Decide per campaign before September 1. Three honest options: adopt deliberately with the settings you choose, remove the legacy setting to stay out of the mapping, or let the migration run with a calendar entry to verify the result. Choosing nothing is choosing the third option without the calendar entry.
3. Sweep paused campaigns too. They carry eligible settings into the window and convert without anyone noticing.
4. After conversion, verify the defaults landed as documented for your cohort, check ad-group search term matching flags, and run the first weekly review immediately.
5. Keep the change history. Preflight query 4 distinguishes changes made by your team, by rules, by the API, and by Google. The migration writes to your account; your audit trail should say so.

The bottom line

The wrong question is whether AI Max works. The useful question is which decisions Google may make for this campaign, under what constraints, and what evidence must exist before it gets more freedom. A-side campaigns turn trustworthy signals, deep sites, and scalable economics into incremental value. B-side campaigns protect precision, message control, and capacity limits that expansion would damage. Everything in between is component-level testing with named owners and rollback triggers: ordinary account management, run on evidence.

The appendix your agent can read

A growing share of Google Ads accounts are watched by AI agents, and a guide only humans can use is half a product. The block below encodes this guide's operating facts in a form an agent can lift whole: paste it, together with the four preflight queries, into the context of whatever tool helps run your account. It asserts nothing the sections above did not already source, and it separates verified product mechanics from our own operating framework, which your agent should feel free to disagree with.

# AI Max operating facts
# Source: https://choice.marketing/blog/google-ai-max-field-guide/
# Mechanics verified against Google documentation and Google Ads API v22, 2026-08-25.

# --- verified product mechanics ---
migration:
  window: 2026-09-01 to 2026-09-30
  triggers: [campaign_level_broad_match, legacy_automatically_created_assets]
  legacy_creation_blocked_since: 2026-08-03
  converted_defaults:
    broad_match_cohort: {search_term_matching: on, text_customization: off, final_url_expansion: off}
    aca_cohort:         {search_term_matching: on, text_customization: on,  final_url_expansion: off}
  paused_campaigns: no exemption documented (our inference), sweep them too
api_fields:
  umbrella: campaign.ai_max_setting.enable_ai_max            # campaign level
  bundling: campaign.ai_max_setting.bundling_required        # REQUIRED = components need the umbrella
  broad_trigger: campaign.keyword_match_type                 # BROAD = campaign-level broad match
  aca_trigger: campaign.asset_automation_settings            # TEXT_ASSET_AUTOMATION + OPTED_IN
  stm_opt_out: ad_group.ai_max_ad_group_setting.disable_search_term_matching
  served_combinations: ai_max_search_term_ad_combination_view  # search_term, headline, landing_page
  change_audit: change_event   # explicit date range + LIMIT required, 30-day lookback max
component_rules:
  final_url_expansion_requires: text_customization
  search_term_matching_default: on when the umbrella enables, disable per ad group
  negative_keywords: honored unchanged
monitoring_reports:
  - search terms filtered to match type "AI Max", with the Source column
  - assets filtered to "Google AI" in the Added by column
  - landing pages with the Selected by column
  - CRM lead quality compared against platform-reported conversions

# --- Choice OMG operating framework ---
placement_spectrum:
  A:  {stm: on,  tc: on,  fue: on,  use: full expansion engine}
  A2: {stm: on,  tc: on,  fue: off, use: controlled expansion}
  M1: {stm: on,  tc: off, fue: off, use: matching-led expansion}
  M2: {stm: off, tc: on,  fue: off, use: creative assist}
  B2: {stm: one ad group, tc: off, fue: off, use: contained experiment}
  B:  {stm: off, tc: off, fue: off, use: precision defense}
readiness_gates: [measurement_integrity, economic_room, query_opportunity,
  creative_tolerance, site_fitness, brand_geo_policy, governance]
scoring:
  scale: 0-2 per gate, 14 max
  bands: {12-14: trial A2, 8-11: M1 or M2, 4-7: B2, 0-3: B}
  overrides:
    - gate 1 (measurement) at 0 caps placement at B regardless of total
    - gate 4 (creative risk) hard-stop keeps text_customization off at any score
rollback_triggers: [generated_claim_error, unsafe_landing_page,
  brand_leakage_past_agreed_share, invalid_lead_rate_past_agreed_line]

If a field name in the appendix stops resolving, the API version moved; check the source date above against Google's current release notes before trusting the rest.

Sources checked for this guide

All Google Help Center pages accessed August 25, 2026.

• Google, AI Max setup
• Google, How AI Max works
• Google, AI Max reporting
• Google Ads API, AI Max for Search campaigns, updated August 19, 2026
• Google Ads Developer Blog, migration announcement, published August 12, 2026
• Google, AI Max testing and planning tools, published August 20, 2026
• Google, AI Max upgrade announcement, published April 15, 2026
• Smarter Ecommerce, AI Max guide, published March 23, 2026
• Location3, AI Max franchise results, published March 18, 2026
• Search Engine Land, Geddes AI Max copy test, published July 28, 2026
• Lunio, AI Max invalid traffic rates, updated August 14, 2026
• Choice OMG, AI Max auto-upgrade dossier, published August 25, 2026
• First-party account data: Google Ads API v22, MCC sweep of 19 accounts, August 25, 2026