Automating a Google Ads account, and the parts you should not

What can honestly be automated in a search account

Split the work in two. There is the work of reading: pulling reports, sorting thousands of search queries into groups, spotting the campaign that started spending differently last Tuesday, checking that every ad group has the right negatives. There is also the work of deciding: raise this budget, pause that campaign, change the offer.

The reading is a good fit for automation because it is high volume, rule-shaped and cheap to check. Nobody enjoys it and nobody does it as often as they should, which is why most accounts carry months of accumulated waste that a careful hour would find.

The deciding is a poor fit, not because a model cannot form a view but because the cost of a wrong view is immediate and paid in cash. The practical arrangement, and the one that works in real accounts, is that the machine reads everything and proposes, and a person approves in a batch once or twice a week.

The audit, and what it actually reads

An account audit is the first thing to automate because it is the same checklist every time and it is the thing most accounts have never had done properly.

Conversion tracking comes first and everything else waits on it. If the account counts a form page load as a conversion, or counts one enquiry three times, then every report is wrong, every automated recommendation is wrong, and the platform's own bidding has been optimising towards the wrong thing for as long as it has been running. Check this before you look at anything else.

• Conversion tracking: what counts as a conversion, whether duplicates are being counted, and whether the count matches what your sales records say arrived
• Match types: which keywords are set to broad, which means the platform decides what counts as a related search and shows your ad against it
• Search terms: the real phrases people typed, which is a different list from the keywords you chose
• Negative keywords: what is excluded, at what level, and whether the list is shared across campaigns or trapped in one
• Budget against demand: campaigns that run out of money before the day ends, and campaigns that never spend what they are given
• Ad assets: how many variants are live, whether the extra links and callouts are filled in, and whether anything has been disapproved quietly
• Message match: whether the search phrase, the ad headline and the landing page headline say the same thing
• Settings that leak money: locations set to include people merely interested in your area, ad schedules running when nobody answers the phone, and partner networks left switched on

Search term mining, the loop worth automating first

This is the highest value automation in a search account and it is almost entirely mechanical. Every week, pull the list of what people actually typed before clicking. Group those phrases by intent. Show the ones that spent money and produced nothing.

A good grouping does most of the work. Job seekers looking for employment with a company like yours. People wanting to do it themselves. Students and researchers wanting a definition. Existing customers looking for support. Bargain hunters attaching the word free to everything. Competitor names. None of these are buyers, and in an unmanaged account they quietly take a meaningful share of the budget.

Two lists come out. A general one you can reuse across every campaign and every client in the same trade, because plumbing courses are irrelevant to every plumber. And an account-specific one mined from your own report, which is where the money actually is, because it contains the odd phrases nobody would have predicted.

The automation drafts both lists with the spend against each phrase, and a person reads them before anything is excluded. That read matters more than it sounds, and the next section explains why.

Ads and landing pages that say the same thing

The most reliable way to lift a search campaign is to make the ad echo the search and the landing page echo the ad. Someone types a phrase. They see that phrase back in the ad. They land on a page whose headline repeats it again. Nothing clever is happening. The person is simply never made to wonder whether they are in the right place.

This is tedious to do by hand across an account, which is exactly why it usually is not done. Producing many ad variants, each with the relevant phrase in the first headline and the offers rotating behind it, is a good use of automation. So is generating a matching landing page per group rather than pointing everything at one homepage.

Two practical rules. Fix the leading headline to the phrase the group is targeting so it always appears, and let the platform test the rest. And do not run more variants than your traffic can settle: if a group gets a handful of clicks a week, a large test will never produce a readable answer, it will only produce noise you might act on.

The person should never be made to wonder whether they are in the right place.

Pacing, alerts and a report somebody reads

Pacing is arithmetic, so hand it over. Spend so far against the month, projected finish, which campaigns are ahead and which are behind. A short daily check that only speaks when something is off beats a dashboard nobody opens.

Alerts are worth writing carefully, because an alert that fires often gets ignored. Useful ones: a campaign that stopped spending entirely, spend that jumped sharply without a matching rise in enquiries, an ad or account-level disapproval, conversion tracking that has recorded nothing for a day when it normally records something, and a landing page returning an error.

For reporting, the useful output is short and in plain language. What was spent, how many enquiries arrived, what each one cost, what changed since last week, and what the account needs decided. The number that matters is not clicks and often is not even leads. It is the enquiries that became money, which means the report has to reach into your sales records rather than stopping at the ad platform.

What should never run without a person on it

Every item here is a real way accounts get damaged, and most of them look sensible in a settings screen.

Write access with no ceiling. Anything that can change budgets needs a hard cap it cannot exceed and a log a person reads. An agent with permission to raise spend and a goal of finding more conversions will find more conversions, and it will find them at a price nobody approved.

Auto-applying the platform's own recommendations. These are optimised for the platform's outcome as well as yours, and switching them all on is how an account quietly ends up on broad match with the budget spread across searches you would never have bought.

Acting on thin data. Pausing an ad after a few clicks, or declaring a winner from a handful of conversions, is not optimisation. Set a minimum volume before any automated recommendation is even generated, and make the agent state the volume it is judging on.

Excluding terms without a human read. Negative keyword lists cause the ugliest automated damage because the effect is invisible: nothing breaks, traffic simply stops arriving. A phrase that looks like research to an agent may be exactly how your best customers describe their problem.

Constant tinkering. Bidding systems need a stable period to learn. An agent that adjusts something every day keeps the account permanently in that unsettled state and then reports that performance is unstable, which it caused.

Finally, this whole approach is the wrong answer when the account is small enough that it can be read by hand in an afternoon, or when the thing holding the business back is what happens after the enquiry arrives. Automating an account that sends good leads into an inbox nobody checks makes the waste faster, not smaller.

How soon the reading half runs, and who keeps the account

The reading half is the fast half and the sequence should reflect it. The audit takes the first week and, because an account audit is the same checklist every time, it is also the first thing worth handing over. Search term mining is live inside a fortnight, running weekly against your own query report and needing nothing from anybody once it is set. Ad and landing page alignment arrives across month two, since it depends on the mining having produced a list worth acting on.

Every switch on the list of things that should never run alone waits for human approval, and the mechanism is worth naming rather than the principle. Bid strategy changes, budget moves, campaign types being switched on, geographic and audience changes and anything published in the account are drafted and then approved by a person. Where the system meets something it cannot classify, a query that might be a new intent or might be noise, it hands it to a person with the data attached rather than adding a negative keyword to be safe.

The account stays yours, and in paid search that is not a small point, because the learning inside an account is what every later change gets judged against. The account, the conversion tracking and the automations sit under your own logins, so you own the system and a change of supplier does not reset the history. Running the mining loop with your own team is realistic for a business with somebody who enjoys the work, and doing it by hand for a month teaches more about the account than the first automated report will.

The published version of this is Zavcom, a United States internet, television and telephone provider, where AI-powered Google Search and call-only campaigns held 89 conversions across ninety days at $1.54 a click and a 5.8 per cent click through rate, delivered with Wembly Studios, and Amari Sayulita, a boutique hotel in Mexico, on intent-segmented search. Ali owns marketing and sales at Wobble, which works from Karachi, bills month to month and carries 25 engagements across six countries.

Common questions

Can AI run my Google Ads account by itself?

It can do all the reading, sorting, drafting and reporting without help, which is most of the hours. It should not hold unsupervised control of budgets, bids or campaign status, because the cost of a wrong decision is paid immediately in cash and is hard to reverse. The arrangement that works is that the system proposes changes continuously and a person approves them in a batch.

What is the first thing to automate in a search account?

The weekly search term review and the negative keyword list that comes out of it. It is mechanical, it is the task most accounts skip, and it directly stops money going to searches that were never going to buy. Have the system produce the proposed exclusions with the spend against each one, and read the list before applying it.

Why does conversion tracking need checking first?

Because every other number depends on it. If a page load is being counted as an enquiry, or one enquiry is counted several times, then your cost per lead is fiction, the platform's automated bidding has been chasing the wrong signal, and any recommendation built on that data inherits the error. Compare what the account reports against what your sales records show actually arrived.

Should ad copy be generated automatically?

Producing variants is a good use of automation, particularly making sure the phrase somebody searched appears in the ad they see. Two limits apply. Anything that makes a claim about price, guarantees or results needs a person to confirm it is true. And do not run more variants than your traffic volume can meaningfully settle.

What can go wrong with automated negative keywords?

The damage is invisible, which is what makes it dangerous. Nothing breaks and no error appears, traffic simply stops arriving from phrases that were converting. Agents tend to exclude terms that look like research or curiosity, and in some trades that is exactly how buyers describe their problem before they know the industry word for it.

How often should an automated account be reviewed by a person?

Once a week for the approvals queue, and once a month for a proper look at structure and tracking. Daily intervention is usually counterproductive, because bidding systems need a stable stretch to learn and an account changed every day never gets one.