AI in running PPC: what it already does well and where it still fails
Automation has taken over bidding, matching and ad assembly. Where that works, where it quietly loses money, and what work was left to people — written by people who had to build their own tool for it.
- Automation is good at optimising towards a given target. It is bad at deciding whether that target is right.
- Language models save the most time on copy, report reading and feed routine — not on decisions.
- The biggest risk is not a machine making a mistake, but nobody asking why it happened.
Over the past five years the PPC specialist who manually adjusted bids has become someone who sets targets, watches inputs and checks outputs. It is a useful change and it cannot be reversed. It has also accumulated a lot of imprecision in both directions — that “AI already does it by itself” and that “automation just spends money”.
This text is an attempt at a sober split: what automation already does better than a person, what it does worse, and what follows for work in the account.
What automation does better than a person
- Bids in real time. Estimating the probability of a conversion for a specific auction from dozens of signals is a task people never did well. That argument was settled long ago.
- Matching queries. Recognising that “trail running shoes” and “shoes for running in the woods” are the same thing is something the system does better than any keyword list.
- Combining assets. Testing which headline suits which query is combinatorics, not creativity.
- Watching for anomalies. Noticing that a campaign spent triple on Tuesday is exactly machine work — and the only way to catch a problem in time across more than a couple of accounts. We built ClientDeck for it, because a manual morning check of ten accounts does not work.
Where it fails
Every case below shares one denominator: the system optimises towards whatever you labelled as success and has no way of knowing it is not success.
- A badly defined conversion. Measure add-to-cart and you will get carts. The algorithm has no reason to think you wanted orders.
- Differing margins. The system cannot see that category A earns five times more than category B. With one target it sends money where things sell most easily. See margin by category.
- Unqualified enquiries. In lead generation it optimises for form fills, not for contracts — until you send back what happened to the enquiry. See offline conversions.
- Incrementality. The algorithm is happy to take credit for people who would have bought anyway. It has no reason not to; for it they are cheap conversions. See incrementality.
- Company capacity. It never knows you have run out of stock or that the salesperson cannot keep up with calls.
- Marginal return. It optimises to the target you set, not to maximum profit. The difference between the two is exactly what the article on budget is about.
Automation is excellent at answering “how”. It has no opinion on “what” and “why” — and those are what decide the money.
Language models in daily work
Different from in-platform automation are tools like ChatGPT or Claude used by a specialist at work. Here is our split based on what has worked for us:
| Task | How it goes | What to watch |
|---|---|---|
| Headline and description variants | Well; it will produce dozens worth testing | Invented product features and prohibited wording in regulated fields |
| Naming and tidying structure | Well; routine with no risk | Nothing material |
| Cleaning and enriching a feed | Well for titles and descriptions | Never for price, availability and identifiers |
| Reading a report and spotting patterns | Well as a first pass | Verify conclusions in the data; models readily confirm your hypothesis |
| Analysing search terms and suggesting negatives | Well for pre-sorting | The final call belongs to someone who knows the range |
| Decisions about budget | Badly | The model does not know your margin, capacity or season unless you tell it |
| Interpreting statistical significance | Badly | Frequent errors about what a test actually proved |
We never let a model make changes that take effect without a human approving them. Every change in an account has a written reason, an expected result, the name of whoever approved it and a review date. This is not distrust of machines — it is the same rule we apply to ourselves.
What to fear and what not to
Do not fear that automation will make an account expensive by itself. It becomes expensive when it gets the wrong target or the wrong data, and that is a human decision.
Do fear that over time nobody asks why something is happening. The most dangerous state of an account is not “the algorithm made a mistake” but “nobody knows what changed or who approved it”. With black boxes such as Performance Max and AI Max that risk is larger than ever, because the numbers look fine even when they mean nothing.
The practical defence is boring and it works:
- A change history you can roll back to.
- A regular check against real orders, not against the platform interface.
- Brand separated from the rest, so the report can be read at all.
- Once a year, an experiment that verifies how much of this is genuinely additional.
What was left to people
The complete list as we see it today: deciding what counts as success and what it is worth. Knowing the margin and the capacity of the business. Deciding what may be paid for a customer. Building measurement that can be trusted. Writing an offer worth advertising. And being able to say “do not do this” when the numbers look good.
No system does any of that yet, because they are not optimisation problems. They are decisions about a company’s money.
What to take away
Automation has moved the work in PPC from configuring to briefing. It is a change for the better, but it raises the cost of a mistake: a badly set target now shows up faster and at larger scale than before. The best investment in “AI for PPC” is therefore not new tools, but solid conversion data, knowledge of your own margin, and discipline in logging changes.
Common questions
Will AI replace the PPC specialist?
It has already replaced the part of the job that consisted of setting bids and manually matching keywords — that argument is settled. It has not replaced deciding what counts as success and what it may cost: automation optimises towards a given target but has no way of knowing the target is wrong. The work left to people is defining conversion value, knowing the margin and the capacity of the business, building trustworthy measurement, and being able to say something should not be done.
What can I safely use ChatGPT or a similar tool for in PPC?
It works well for headline and description variants, tidying naming and structure, cleaning titles and descriptions in a product feed, a first pass through a report, and pre-sorting search terms for negatives. We do not recommend it for budget decisions, because it does not know your margin, capacity or season, nor for interpreting statistical significance, where models frequently err. With feeds, never let a model near price, availability or product identifiers.
How do I keep control of automated campaigns?
With four things, none of them technical. Keep a change history with the reason, the expected result and the name of whoever approved it, so you can roll back. Check results against real orders in your own system, not against the platform interface. Separate brand queries, or the report cannot be read. And at least once a year verify by experiment how much of the result is genuinely additional — with black boxes like Performance Max the numbers look fine even when they mean nothing.
- Google Ads AI Max For Search In 2026: Setup, Real Performance Data, GroasSource of documented cases where automation without guardrails raised cost per conversion. An agency blog.
- Explaining incrementality, PPC LandBackground on why automation takes credit for conversions that would have happened without the ads.