Every vendor in this category publishes a percentage improvement — better cost per acquisition, higher return on ad spend, some multiple of results. None of it is verifiable, all of it is selected, and none of it tells you what will happen in your account. So rather than add another number to the pile, here is what can honestly be said about where AI advertising helps, where the claims outrun the evidence, and how to test it yourself.
Why vendor statistics are not evidence
Not because vendors are lying. Because the comparison is unsound in ways that are hard to avoid.
Compared against what?
Usually against what the customer was doing before, which was frequently boosting posts or nothing. Almost anything structured beats that, and it says little about the tool.
Survivorship
The customers quoted are the ones who stayed. The ones for whom it did not work are not in the sample, and there is no way for you to see them.
Confounded
A business adopting a new tool usually changes several things at once — budget, offer, attention. Attributing the result to the software is a guess.
💡 A useful signal: a vendor who leads with how the product works rather than with an improvement percentage is telling you which conversation they are comfortable having. Judge on mechanism, not on claimed outcomes.
Where AI demonstrably helps
These are claims about mechanism, which you can verify rather than take on trust.
It builds campaigns that are structurally correct
Most self-run campaigns have structural faults that produce no warning — an objective that does not match the goal, one flat audience instead of comparable segments, a budget below the platform floor, placements pinned for no reason. A tool that gets these right is not being clever; it is avoiding known errors consistently. That is a real and measurable advantage over most manual setups.
How much? On one account, three campaigns in the same week differed only by objective and cost ₹6.31, ₹9.39 and ₹33.13 per click. Objective selection alone moved cost per click five-fold. Getting that right is not a marginal gain.
It produces variation at a volume people do not
A person writes two headlines and becomes attached to them. A model writes twenty in a minute and you pick three to test. Creative testing is one of the few things in advertising with reliably positive returns, and the constraint has always been production capacity rather than knowledge.
It does not get bored
Checking search terms weekly, watching for creative fatigue, noticing delivery stopped, alerting you the moment a lead lands. None of this is intellectually difficult and almost nobody does it consistently by hand. Response time in particular predicts conversion more reliably than most targeting decisions, and a system that never sleeps beats a person who does.
Where the claims outrun the evidence
"Continuous optimisation"
Sold as an unambiguous good, and it is not. Every edit to a live ad set restarts the platform's evaluation of it, which means a system rewriting your campaigns daily can suppress the very delivery it claims to be improving — and the interface gives you no signal that this is happening.
Adjustment on real signal after a learning period is valuable. Rewriting live ad sets daily is interference. A vendor who cannot distinguish between the two in conversation probably does not distinguish between them in code.
"AI-powered targeting"
Means completely different things between products, and the difference is invisible. Meta exposes roughly 284 behaviours through its API and returns the same catalogue for every country. A tool that fetches this live and selects from it is doing something real. A tool shipping a hardcoded shortlist is not — and both describe themselves identically.
We shipped the second version by accident: a hand-maintained list of fifteen behaviours was discarding 161 of 177 model selections for having no matching ID. Every campaign launched. Nothing errored. The list had been written for one market, so that market looked fine and every other one ran degraded. If it can be invisible to the people who built it, it is certainly invisible to a buyer.
Anything about creative quality
Generated creative is good enough and occasionally very good. It is not reliably better than a competent designer, and the claim that AI creative outperforms human creative is not supported by anything you can inspect. What is supportable is that it produces twenty options where a designer produces two, at a fraction of the cost — which matters, but is a different claim.
The thing that determines the outcome more than the tool
Worth stating because it is uncomfortable and true: whether AI advertising works for you depends mostly on inputs no tool controls.
- 1Whether you described your business honestly, or typed something generic. This changes the output more than switching products does.
- 2Whether you know what a customer is worth to you. Without it, no result can be judged, and you will make decisions on cost per click — which is the number most likely to mislead you.
- 3Whether your conversion tracking fires. A campaign optimising toward an event that never happens is optimising toward nothing, however good the targeting is. We ran one to a 4.62% click-through rate that recorded zero conversions because the destination page did not fire the event.
- 4Whether anyone answers the enquiry. For service businesses this dominates everything. A lead contacted within minutes converts far better than one contacted an hour later, and you paid the same for both.
How to test it honestly on your own account
Fix measurement first
Verify the conversion event fires on the exact destination page. Any test run on broken tracking measures nothing.
One month, one campaign
Properly funded, correct objective, left alone. Not three experiments running simultaneously on a budget that supports one.
Judge on cost per customer
Cost per lead divided by close rate, held against customer value. Not cost per click, not reach, not engagement.
Against the honest baseline
Compare against a properly built manual campaign if you can build one — or against nothing, if that is the real alternative. Be clear which you are testing.
The short answer
Yes, for the mechanical work, and the gain is real and verifiable — structurally correct campaigns, creative variation at volume, and consistent attention that humans do not sustain. No, for the parts that decide whether the campaign was worth running: what you sell, to whom, at what price, and what a customer is worth. Those have not moved and are not about to.
Test it on one campaign, judge it on cost per customer
Adyft builds and launches Google and Meta campaigns into your own ad accounts, with everything previewed first. 14-day free trial, no card, 0% of ad spend.
Start free — no card needed