The short answer
An AI Google Ads audit uses an assistant connected to the Google Ads API to read every search term, ad, keyword, budget and setting in one pass — seconds, not an afternoon — then surface the spend that earns nothing. Auditing only ever reads, so it changes nothing, and its output is a list of candidates for you to accept or reject.
What an AI audit checks
In roughly this order, because each item depends on the one above it being sound.
- Conversion tracking. Are conversions recording at all, are any counted twice, and is anything important not counted? Every other number in the account is downstream of this, so an audit that skips it is measuring against a broken ruler.
- Search terms. What did you actually pay for, as opposed to what you bid on? This is where most recoverable money is.
- Ad coverage. Ad groups with one ad, or none serving, or all ads disapproved. Cheap to find, embarrassing to leave.
- Budget and impression share. Campaigns capped while still winning; and the opposite, campaigns with headroom they are not using.
- Geography and schedule. Spend in places you do not serve and at hours nobody answers the phone.
- Structure. Keywords whose landing page does not mention them; ad groups that have become a dumping ground.
- Recent change. What moved lately, and did anyone mean it? Google's change history answers this, and it is the fastest explanation for a sudden shift.
Where the money actually goes
Wasted spend is rarely one dramatic mistake. It is usually a long tail of terms that each look defensible alone. The recognisable patterns, with illustrative examples rather than anyone's real data:
| Pattern | What it looks like | Why it costs you |
|---|---|---|
| Job intent | “<product> jobs”, “<product> manager salary”, “<product> internship” | Someone looking for employment, not software. Almost never converts, and often cheap enough to go unnoticed for months. |
| Existing customers | “<product> login”, “<product> support”, “<product> app download” | You are paying to reach people who already pay you, and often outbidding your own organic result to do it. |
| Free intent | “free <product>”, “<product> free forever”, “<product> crack” | Explicitly signalling they will not pay. Worth a negative even when the click is inexpensive. |
| Research intent | “what is <category>”, “<category> wikipedia”, “<product> tutorial pdf” | Top-of-funnel with no purchase signal. Can be worth having deliberately — but it should be a decision, not an accident. |
| Careers and study | “<category> certification”, “<category> course”, “<category> exam” | Students and career-changers. A large category for anything with a professional audience. |
| Wrong product | A competitor's product name, or a homonym of yours | Broad match drifting. The single most common source of quiet spend on a well-run account. |
The questions to ask
Paste these more or less verbatim once your assistant is connected. All of them are reads.
- “List search terms from the last 90 days with cost above [threshold] and zero conversions, sorted by cost.”
- “Group those terms by intent and tell me which look like job seekers, existing customers or free-tier hunters.”
- “Which conversion actions recorded nothing in the last 30 days?”
- “Which ad groups have fewer than two serving ads?”
- “Which campaigns are limited by budget, and what is their lost impression share to budget?”
- “Where are we spending outside our target locations, and how much last month?”
- “Which keywords do not appear anywhere on their landing page?”
- “What changed in this account in the last 14 days, and who changed it?”
Ask the last one first if something has moved unexpectedly. It is the cheapest question in the list and it explains more surprises than any analysis.
Telling a finding from noise
An assistant will happily produce a hundred observations. Three tests separate the ones worth acting on:
- Is it costing money now? A pattern that cost you nothing last month is an observation, not a finding.
- Is the sample big enough to mean anything? A keyword with eleven clicks and no conversions has told you almost nothing. This is where AI output most often overreaches, and where a human should push back.
- Can the fix be walked back? Prefer the ones that can. A negative keyword can be removed tomorrow; a restructure cannot be put back in an afternoon.
Asking “how confident are you, and on what sample?” is worth doing routinely. A good answer names the numbers; a bad one restates the claim.
What about Google's own recommendations?
The recommendations tab and the optimisation score are genuinely useful and genuinely not neutral. They optimise for a healthy account as Google defines it, which correlates with spending more. Some are almost always right — fix disapproved ads, add missing conversion tracking. Others deserve a second opinion: broadening match types, raising budgets, auto-applying changes.
The useful division of labour is that Google tells you what it wants, and an audit run for you tells you what your money did. Neither is the whole picture, and the disagreements are usually the interesting part.
How often to run it
- Search terms: weekly. This is the one with a compounding cost, and it takes minutes when it is a habit.
- Full audit: monthly on an account that is running steadily, or weekly for the first month after any significant change.
- Change history: whenever a number moves in a way nobody expected.
The reading half can run on a schedule you set in your own AI client — the daily routine is that recipe, and it is written to read and report, not to change things.
From finding to fix
An audit that produces a list nobody acts on has cost you an hour. The shortest path from finding to fix is to let the same assistant apply the ones that can be walked back — adding negative keywords, pausing an ad that has stopped serving — the moment you have read the list and said yes.
That needs the full tier, which a connector has unless you ask otherwise, and it is worth doing in the order described in earning autonomy: questions first, then one account, then one job. Deletion is never on the table — the removal tools that delete an ad, a keyword or an asset are not exposed at all.