A useful audit produces evidence, not a score
A Google Ads audit should answer three questions: Can the numbers be trusted? Where is profit being created or destroyed? What should change first? A thirty-item scorecard cannot answer those questions unless every finding is tied to account evidence and a commercial consequence.
Choose a review period long enough to cover the business’s conversion lag and seasonality. Export campaign, search-term, landing-page and conversion-action data before changing anything. Record the account time zone, attribution settings and the date of the audit so another person can reproduce the analysis.
Build the audit evidence pack
| Export or screenshot | Why it matters |
|---|---|
| Goals summary and conversion-action settings | Shows which outcomes feed the Conversions column and bidding |
| Change history for 90 days | Explains whether a performance shift followed a human, automated or system change |
| Campaign report segmented by network, device and time | Separates blended averages that hide loss-making segments |
| Search terms and search-term insights | Shows actual demand captured, subject to Google’s privacy thresholds |
| Landing pages and final URLs | Connects spend to page experience and conversion paths |
| CRM orders or qualified leads by click source | Tests whether platform conversions became business outcomes |
Google notes that the search terms report omits some low-volume queries for privacy, while search-term insights aggregate them into themes. “No visible bad queries” therefore does not prove that all matching is relevant.
1. Prove measurement before judging media
List every primary conversion action. For each, write the business event, trigger, count setting, value source, attribution window and owner. Then perform a test conversion. Confirm the event once in the site data layer or tag debugger, once in GA4 if used, and once in Google Ads after processing.
For ecommerce, reconcile a sample of order IDs and revenue against the commerce platform. Tax, shipping, discounts, refunds and currency handling must be intentional. For lead generation, a form submission is usually an intermediate event. Reconcile lead IDs through qualified, opportunity and won stages, then import the stage that reflects economic value.
Do not make a GA4 import and a native Google Ads tag primary for the same purchase unless you deliberately want both counted. Secondary actions can remain available for observation without directing bidding.
2. Translate the business model into bidding constraints
Before labelling a CPA or ROAS “good”, calculate the break-even point. Consider a hypothetical retailer with a £120 order, £54 product cost, £10 fulfilment and payment cost, and £12 expected returns allowance. Contribution before advertising is £44. Break-even first-order ROAS is therefore £120 ÷ £44 = 2.73, or 273%. If the business needs £12 contribution after ads, allowable ad cost is £32 and required ROAS becomes 375%.
For a lead business, suppose 40% of enquiries become qualified, 25% of qualified leads close, and a new customer is worth £1,200 contribution. Expected contribution per raw enquiry is 0.40 × 0.25 × £1,200 = £120. If the company wants £40 after acquisition cost, the maximum raw-lead CPA is £80. Replace these hypothetical numbers with finance and CRM data.
3. Segment before diagnosing
Split brand from non-brand, acquisition from remarketing, existing customers from new customers where observable, and Search from Shopping or Performance Max. Then compare campaign, query theme, product group, device, location and landing page. A blended account ROAS can rise while new-customer profit falls if branded demand or repeat buyers take a larger share.
Look for materiality, not arbitrary thresholds. A query spending £100 without a sale means something different when the expected CPA is £20 than when it is £250. Use a simple triage:
- High spend, low business value: investigate immediately.
- Low spend, clear irrelevance: exclude when the intent is unambiguously wrong.
- Low volume, plausible intent: collect more evidence or pool with a meaningful theme.
- Strong platform results, weak CRM results: fix the optimisation signal before scaling.
4. Review settings in the context of the strategy
Check location presence options, languages, networks, brand settings, ad schedule, audience modes, URL expansion, final URL exclusions, product-feed filters and account-level negatives. A setting is not wrong because it differs from a template. It is wrong when it exposes budget to demand the business did not intend to buy.
For bidding, verify that the chosen objective matches the imported outcome. Google’s bid-strategy guide distinguishes conversion-volume strategies from value strategies. Check whether targets are constraining traffic by comparing the target with achieved CPA or ROAS and reviewing bid-strategy status. Do not invent a universal minimum conversion count; data sufficiency depends on frequency, delay, variance and signal quality.
5. Follow the click from query to page
Take the ten highest-spend non-brand themes. For each, record the query, keyword or matching route, ad promise, final URL, page headline, offer and primary conversion. This exposes message breaks that account-level metrics conceal.
Review ad assets for accuracy and coverage, not merely Ad Strength. Inspect mobile pages on a real device. Submit forms, use checkout, trigger validation errors and check consent behaviour. PageSpeed data can identify performance problems, but a perfect score does not prove relevance or persuasion.
6. Convert findings into a controlled change plan
Give every recommendation an owner, evidence link, expected mechanism, risk and validation metric. Sequence dependencies first: measurement before bidding; query control before creative conclusions; offer and page fixes before buying more of the same traffic.
For example: “Exclude ‘free template’ queries” is incomplete. A testable recommendation is: “The query theme spent a hypothetical £640, produced 19 primary form events but zero qualified CRM leads. Add an exact or phrase negative at the appropriate scope, monitor lost qualified volume for two weeks, and keep a rollback list.”
Where the effect is uncertain and traffic permits, use a Google Ads experiment. Google advises setting a clear hypothesis and limiting simultaneous changes so the result can be interpreted. Its experiments guidance is a better standard than a before-and-after screenshot.
7. Audit Performance Max and Shopping as retail systems
Do not evaluate retail campaigns only at campaign level. Export item ID, product type, brand, custom labels, clicks, cost, conversions and value. Join this with Merchant Center status, stock, price, margin and refund rate. A profitable campaign can contain a large tail of products with no credible role, while a low-volume product may be strategically valuable because it acquires new customers.
Review asset groups against product sets and destination themes. Check final URL expansion and exclusions, brand settings, account-level negatives, audience signals and search-term insights. Audience signals are inputs, not hard audience restrictions. For a deeper product-data review, use the product feed optimisation guide.
8. Audit the full lead funnel
Create a cohort table by campaign and click month with raw leads, contactable leads, qualified leads, opportunities, wins, contribution and median days between stages. This prevents two common errors: judging recent campaigns before outcomes mature, and optimising to the channel with the easiest form rather than the best customers.
Hypothetically, Campaign A creates 80 forms at £40 but only 8 qualified leads; Campaign B creates 30 forms at £70 and 15 qualified leads. Raw CPA favours A. Qualified CPA is £400 for A and £140 for B. The audit recommendation should be to improve the imported optimisation signal and investigate intent, not to “scale the lowest CPA campaign”.
9. Reconstruct what changed
Overlay daily or weekly performance with change history, website releases, promotions, stock events, consent changes and sales-capacity changes. Classify changes as platform/system, automated rule, agency, internal team or unknown. The number of changes is not a quality metric: one well-designed experiment can be more valuable than hundreds of bid edits.
For each material shift, ask whether timing, mechanism and segment pattern are consistent with the proposed cause. If mobile conversion fell immediately after a form release while desktop and traffic mix remained stable, the release is a strong lead. It is still not proof until QA identifies the defect or a controlled reversal restores performance.
10. Write findings so someone can act
| Field | Example |
|---|---|
| Finding | Purchase is primary in both native Ads and GA4 import |
| Evidence | 17 of 20 sampled order IDs appear in both actions |
| Impact | Reported value and bidding input are overstated |
| Action | Choose one primary action; retain the other as secondary |
| Risk | Bid strategy adjusts to lower reported volume |
| Validation | Seven-day order-ID reconciliation and diagnostics review |
Rank work by confidence, commercial exposure, reversibility and dependency. “High impact” without evidence is an opinion. “£18,400 of the last 90 days was optimised to duplicated purchase values” is a prioritisation case.
Make the audit repeatable
Store the column definitions, filters, attribution basis and CRM joins used in every calculation. A future reviewer should be able to rerun the audit and distinguish a genuine improvement from a changed report. Schedule the first validation date before closing the audit.
State audit limitations
List inaccessible assets, privacy-thresholded search terms, incomplete CRM stages, immature cohorts and missing margin data. Explain how each limitation affects confidence. A transparent partial audit is more useful than precise-looking recommendations built on unavailable evidence.