How to audit a Google Ads account with Claude: 50 checks

A Google Ads audit is a read-only review of tracking, settings, keywords, search terms, ads and automation that ends in a list of fixes ranked by the money each one puts at risk. Start with conversion tracking, because every cost per conversion depends on it, then run the 50 checks below from CSV exports or let Claude run them through Google Ads Query Language (GAQL).

The checks below come from google-ads-account-audit, an MIT licensed skill in adwarp-dev/paid-ads-skills, down to each threshold and query. It never writes to the account, so the fixes are yours to make in Google Ads.

What a Google Ads audit covers#

The skill runs 50 checks in 11 areas, in the order below. It scores Search, Performance Max and AI Max campaigns, and lists Display, Video, Demand Gen, App and Shopping campaigns as out of scope.

Each check ends as pass, warn, fail, n/a or not checked. In the tables, T means your target cost per acquisition, and "judgment call" marks a threshold the skill chose that you can change. Google's help pages set the rest.

Google's help pages, API guides and access levels quoted in this post were checked on 2026-11-07.

Queries carry a leading zero, Q01 to Q21, so they don't get confused with the search term checks Q1 to Q4. A few IDs hold more than 1 query, and all 31 are in references/gaql-checks.md.

The report opens with a scorecard that grades each area as (passes + 0.5 x warns) / (passes + warns + fails) x 100. The checklist below does the same math on your marks and downloads all 50 checks as a CSV.

Before you start#

Pick the window and the target first, then decide where the data comes from: CSV exports with nothing connected, or GAQL through Claude, which also reads account settings that no export carries.

Pick a window and a target cost per acquisition#

The skill uses the last 30 days unless you name another window. With no target, it takes the account's own cost per conversion over the window and labels it "account average, used as target". Recent days undercount conversions, so for a long sales cycle end the window 7 days ago (a judgment call).

Option 1, CSV exports#

In Google Ads, open each page below, set the window, click the download icon above the table and pick Excel CSV (Google Ads Help).

Option 2, GAQL through Claude#

Claude runs each check's query through an MCP server, set up in Run the whole audit with Claude. The audit only reads, so Explorer access is enough: 2,880 operations a day on production accounts, per Google's access levels page.

The same page says developer tokens were sunset on September 9, 2026. Access now belongs to a Google Cloud project, where enabling the Google Ads API grants Test access and you apply for Explorer in the Cloud Console.

If you read the rows yourself, divide any field ending in _micros by 1,000,000 to get account currency. GAQL has no LAST_90_DAYS, so a longer window is written segments.date BETWEEN '2026-07-01' AND '2026-09-30'.

5 settings no report shows#

For these, screenshot each setting page or mark the check "not checked". The skill never guesses a setting.

• Auto-tagging (C9): Admin > Account settings. Q01 reads it.
• Enhanced conversions (C7): Goals > Settings. Q01 reads this one too.
• Consent Mode (C8): Goals > Conversions > Summary, a website action, then Diagnostics. No API field reports it.
• Auto-apply (H1): the Recommendations page. Q20 sees only the types the API supports.
• AI Max (X1 to X3): campaign settings, or queries Q04, Q12 and Q13.

Conversion tracking (C1 to C9)#

Smart Bidding learns from primary conversion actions, and every cost per conversion in this audit divides by them. Google's help page says primary actions are "used for bidding" when their goal is, and secondary ones are "for observation only" (Google Ads Help).

If C1 or C2 fails, the skill marks S1, S2, S5 to S9, K1, K5, Q1, Q3, Q4, P4 and X1 "unreliable: tracking" and still reports their raw spend.

C5 follows Google's counting guide, which says usually only 1 lead per ad click adds value while every sale does (Google Ads Help). For C8, give a new Consent Mode setup time: its status can take 48 hours to show in Diagnostics, and up to 2 weeks in some cases (Tag Manager Help).

Q01 reads 3 of these checks in 1 call:

SELECT customer.id, customer.descriptive_name, customer.currency_code, customer.time_zone,
  customer.auto_tagging_enabled,
  customer.conversion_tracking_setting.conversion_tracking_status,
  customer.conversion_tracking_setting.accepted_customer_data_terms,
  customer.conversion_tracking_setting.enablement_setting,
  customer.conversion_tracking_setting.effective_enablement_setting
FROM customer

That's the v26 form. Adwarp pins API v25 unless ADWARP_GOOGLE_API_VERSION says otherwise, and on v25 the last 2 fields become customer.conversion_tracking_setting.enhanced_conversions_for_leads_enabled. It covers leads only, so check web enhanced conversions in the UI.

Campaign settings (S1 to S9)#

Settings decide where the money can go before any keyword matches. Most of these checks compare a campaign with T, or its partner and Display traffic with its own Google Search traffic.

Q05 splits each Search campaign by network: SEARCH is Google Search, SEARCH_PARTNERS is search partners and CONTENT is the Display Network.

SELECT campaign.id, campaign.name, segments.ad_network_type,
  metrics.cost_micros, metrics.clicks, metrics.conversions
FROM campaign
WHERE campaign.advertising_channel_type = 'SEARCH' AND campaign.status = 'ENABLED'
  AND segments.date DURING LAST_30_DAYS

Google's own page on the Search Network warns that partner clicks "may not always reflect highly targeted traffic" (Google Ads Help).

Google is changing Display expansion. It added High and Low priority tiers in October 2026, and from February 2027 Search campaigns with Display Expansion move to channel prioritization automatically, which Google warns can shift spend between Google Search, search partners and Display (Google Ads Help). The skill found no API field for the tiers, so read them in campaign settings.

Presence or interest is Google's default and recommended location option (Google Ads Help), and it suits a business people travel to. S3 warns only when the data shows leakage.

For S5, Google lets target cost per acquisition bidding start with no history but recommends judging it over 30 days with at least 30 conversions (Google Ads Help), or 50 for target return on ad spend (Google Ads Help). S6 rests on 1 line in Google's match type guide: "It's critical to use Smart Bidding with broad match" (Google Ads Help).

Structure and keywords (T1, T2, K1 to K5)#

T2 looks for brand keywords inside non-brand campaigns. They convert cheaply and drag the blended cost per conversion down, so a shared target ends up overbidding on brand. The keyword checks after it hunt for cost without conversions, by match type and keyword by keyword.

Google's help page says Quality Score "is not an input in the ad auction" and "should not be optimized or aggregated with the rest of your data" (About Quality Score). K2 never averages it and reads the 3 components only on the keywords that carry the spend.

K3 has its own query, Q10, because a keyword with 0 impressions never appears in Q09's performance rows.

Search terms and negatives (Q1 to Q4, N1 to N4)#

The search terms report is where waste shows, though not all of it. Google says "some search terms that don't have enough query activity are omitted from the search terms report" to meet its privacy standards (Google Ads Help). So the waste Q1 reports is a floor, and Q2 measures how much of the cost shows up in the report at all.

SELECT campaign.id, campaign.name, ad_group.id, ad_group.name,
  search_term_view.search_term, search_term_view.status,
  segments.search_term_match_type,
  metrics.cost_micros, metrics.clicks, metrics.impressions,
  metrics.conversions, metrics.conversions_value
FROM search_term_view
WHERE segments.date DURING LAST_30_DAYS AND metrics.impressions > 0

Q3 groups these rows by segments.search_term_match_type and compares NEAR_EXACT, NEAR_PHRASE, BROAD and AI_MAX with EXACT.

N3 applies Google's negative matching rules (Google Ads Help). A negative broad keyword blocks a query holding all its words in any order, a negative phrase needs them in order, and a negative exact needs the same words with nothing added. Negatives don't match close variants.

The account-level list in N4 takes up to 1,000 negatives and covers Search and Shopping inventory in Search, Performance Max, App, Shopping, Smart and Local campaigns (Google Ads Help).

The audit sizes the waste. For the full negative list and an upload file, run google-ads-search-terms.

Responsive search ads and assets (R1 to R5, A1, A2)#

Google asks for "at least 2 responsive search ads (RSA) with 'Good' or 'Excellent' Ad Strength per ad group" and allows up to 15 headlines and 4 descriptions in each ad (Google Ads Help). An ad group takes 3 enabled responsive search ads at most (account limits).

R2 never fails. Google says Ad Strength "doesn't determine whether your ad is eligible to serve" and "isn't used to calculate Ad Rank, Quality Score, or auction wins" (About Ad Strength).

R4 takes Google's line that pinning "isn't recommended for most advertisers and can affect ad strength".

For A1, sitelinks need at least 2 to show, and Google's tip is at least 6 with descriptions (Google Ads Help).

Performance Max (P1 to P4)#

Performance Max can show on the same searches as your Search campaigns. P1 and P3 look at that overlap through brand and search themes. P2 grades the asset groups, and P4 prices the Performance Max search terms that never convert.

Brand exclusions in Performance Max apply to Search and Shopping inventory (Google Ads Help). An asset group takes up to 25 search themes, and they get "the same prioritization as your phrase match and broad match keywords in Search campaigns", while an exact keyword identical to the query still wins (Google Ads Help).

P4 reads campaign_search_term_view, because search_term_view has no Performance Max cost. In Google Ads, it's the "Search terms and landing pages for Performance Max" view of the search terms report (Google Ads Help).

AI Max for Search (X1 to X3)#

AI Max lets Google match searches beyond your keywords and adapt ad text and landing pages. Google's API guide says that under AI Max "all keywords are treated as broad match by default" (Google Ads API).

Q12 splits search terms by segments.search_term_match_source and compares the AI_MAX_KEYWORDLESS and AI_MAX_BROAD_MATCH rows with ADVERTISER_PROVIDED_KEYWORD (AI Max reporting). It runs only when campaign.ai_max_setting.enable_ai_max is true.

For X2, API v25 added broad_match_migration_date_time and aca_migration_date_time, which record when Google moved a campaign to AI Max. X3 matters if your ads carry required text, since Google says "pinned assets aren't used if a more relevant URL is chosen" (Google Ads Help).

Auto-applied recommendations and change history (H1 to H3)#

In the made-up Fernhill account, change history shows Google's recommendations adding the broad match keyword "scheduling software" on October 18, and switching on search partners in a competitor campaign on October 2. Nobody on the team made either change. H1 to H3 find which auto-apply types are on and what they did.

SELECT recommendation_subscription.type, recommendation_subscription.status,
  recommendation_subscription.create_date_time, recommendation_subscription.modify_date_time
FROM recommendation_subscription

H1 fails on an enabled USE_BROAD_MATCH_KEYWORD, SEARCH_PARTNERS_OPT_IN or DISPLAY_EXPANSION_OPT_IN subscription. The API doesn't cover every auto-apply type (Google Ads API), so also open the Recommendations page, where "Add broad match keywords" and "Use Display expansion" are among the options (Google Ads Help).

H2 and H3 read change_event, which needs a date filter inside the last 30 days and a LIMIT of 10,000 rows or fewer, and doesn't return Google Ads Editor changes (Google Ads API).

Turn findings into fixes ranked by spend at risk#

To order the fixes, the skill prices every finding in account currency for the window, with the rules below. The scorecard stays secondary: by the skill's own rule, the score never outranks this list.

The fix list then runs in this order:

1. Measurement findings, any fail or warn on C1 to C6, with the cost per acquisition as reported next to the one counting only the primary action that matches the business goal.
2. Quantified findings by spend at risk, largest first.
3. Upside from S7 and Q4.
4. Warnings with no dollar figure, ordered by the spend of the campaigns they touch.

Measurement goes first without a figure because a double-counted lead changes every cost per conversion in the report, and Smart Bidding learns from that count. A search partners finding measured on it can point the wrong way.

Thin data gets no verdict. A cost per conversion check needs at least 2x T of cost in the slice or it's marked "too little data", and Q3, K1 and X1 need 10 conversions in the comparison group (judgment call).

Run the whole audit with Claude#

The skill runs in Claude Code, Codex, Cursor and other agents that read SKILL.md files. It works from the CSV exports alone, and with an MCP server connected it runs the queries itself.

1. Install the skills with npx skills add adwarp-dev/paid-ads-skills.
2. Connect Adwarp or Google's MCP server, or skip this and use exports.
3. Paste the prompt.
4. Read the report, starting with "Do these 3 first".

npx skills add adwarp-dev/paid-ads-skills
claude mcp add adwarp -- npx -y adwarp mcp
npx adwarp google auth login

Adwarp is a free, MIT licensed MCP server and CLI for LinkedIn Ads and Google Ads that runs from your terminal (Google Ads setup). Claude runs each query through gads_search, which refuses anything but a SELECT, and pulls conversion actions with gads_list_conversion_actions and search terms with gads_performance. From a terminal, the same query is adwarp google query "<GAQL>". The Google tools run only on your machine, because the hosted endpoint serves LinkedIn Ads.

Google's Google Ads MCP server runs the same GAQL through its search tool. It also has tools that create assets and labels, and its README says to set labels: false and assets: false in tools_config.yaml for a strictly read-only setup.

The prompt#

Name the account, the window, the target and your brand terms, and say read only.

Audit Google Ads account 123-456-7890 for the last 30 days with the
google-ads-account-audit skill. Target cost per acquisition is 150 USD.
Our brand terms are "fernhill". Read only, change nothing.

For exports, start with "Audit the Google Ads exports in this folder" and put the 8 CSV files in it.

A sample report#

This is the skill's report on the made-up Fernhill exports, cut down to the scorecard, the top 5 fixes and "Do these 3 first". 19 checks show "not checked" because exports carry no negatives, assets, Performance Max detail or the 5 settings above. A connected run fills them in.

# Google Ads audit: Fernhill (made-up account)
Window: 2026-09-21 to 2026-10-20 | Currency: USD | Target CPA: $150.00 (user target)
Data: Export mode | Campaigns in scope: 4 Search, 1 Performance Max | Out of scope: none

## Scorecard
| Area | Pass | Warn | Fail | Not checked | Score |
| --- | --- | --- | --- | --- | --- |
| Conversion tracking | 2 | 3 | 0 | 3 (+1 n/a) | 70 |
| Campaign settings | 2 | 3 | 3 | 1 | 44 |
| Structure | 1 | 0 | 0 | 1 | 100 |
| Keywords | 0 | 4 | 1 | 0 | 40 |
| Search terms | 1 | 2 | 1 | 0 | 50 |
| Negatives | 0 | 0 | 0 | 4 | not scored |
| Responsive search ads | 0 | 4 | 1 | 0 | 40 |
| Assets | 0 | 0 | 0 | 2 | not scored |
| Performance Max | 0 | 0 | 0 | 4 | not scored |
| AI Max | 0 | 0 | 0 | 3 | not scored |
| Recommendations and history | 0 | 2 | 0 | 1 | 50 |
| Overall | 6 | 18 | 6 | 19 (+1 n/a) | 50 |

Spend at risk in the window: $5,366.19 (27.8% of $19,280.00 total cost), counting each campaign, keyword and search term once.

Measurement first: reported conversions are 135, so reported CPA is $142.81. Counting only "Demo request (Google tag)" (88), CPA is $219.09, above the $150.00 target.

## Fixes ranked by spend at risk
| # | Check | Finding | Spend at risk | Fix |
| --- | --- | --- | --- | --- |
| 1 | C4 Double counting | 2 primary "Submit lead form" actions both record conversions: "Demo request (Google tag)" (88) and "GA4 generate_lead" (31) | Blocks bidding ($11,040.00) | Goals > Conversions > Summary: set "GA4 generate_lead" to Secondary |
| 2 | C3 Micro action as primary | A booking link click (Outbound click, 16 conversions) is primary | Blocks bidding ($11,040.00) | Set it to Secondary in Action optimization |
| 3 | C5 Counting | "Demo request (Google tag)" counts every conversion | Blocks bidding ($11,040.00) | Edit "Demo request (Google tag)" > Count > One |
| 4 | S9 Spends without converting | Search - Competitors: $1,790.00, 350 clicks, 0 conversions | $1,790.00 | Pause the campaign, or rebuild it with exact terms and a CPA cap |
| 5 | Q1 Wasted search terms | 5 terms outside Search - Competitors, 0 conversions, each over $300 | $1,980.00 | Run google-ads-search-terms for the negative list |

## Do these 3 first
1. Fix conversion counting (C4, C3, C5): 1 primary lead action counted once per click.
2. Pause or rebuild Search - Competitors (S9): $1,790.00 in 30 days with 0 conversions.
3. Turn off search partners on Search - Crew Scheduling and Search - Field Service (S1, $1,032.19 excess cost) and add the negatives from google-ads-search-terms (Q1, $1,980.00).

Every threshold and the full report format are on the google-ads-account-audit docs page.

How often to audit#

Run the full audit when an account changes hands and once a quarter. When cost per acquisition jumps, run the 9 tracking checks before anyone touches bids or budgets.

The API's change history reaches back only 30 days, so read the rest of the quarter on the Change history page in Google Ads, which keeps 2 years (Google Ads Help). February 2027 belongs on the calendar too. That's when Google moves Search campaigns with Display expansion to channel prioritization, so review every campaign S2 flags before then.

Between audits, paid-media-health-check runs a daily check that stays quiet unless something is wrong. For more on running Google Ads from Claude, see Google Ads in Claude and Google Ads reporting.

By Stan Rymkiewicz, maintainer of adwarp. Published November 7, 2026.