.png)
Most "PPC automation innovation" is a new interface bolted onto a rules engine from 2017. If x, then pause y. If CPA exceeds target for 14 days, send an alert. That logic hasn't changed much in a decade; what changed is the marketing copy wrapped around it and the letters A and I on the pricing page. I've sat through demos where the headline feature was a Slack notification.
So it's worth being precise about what genuinely moved. In the last two years, one thing shifted in this category and everything else follows from it: the software stopped producing a queue of suggestions for a human to approve and started making the changes itself. That sounds like a small distinction until you count the labor. A recommendation engine reviewing an account with 40 campaigns can surface 60 changes a week. Somebody still has to read all 60, judge them against context the tool doesn't have, and click. The bottleneck was never the analysis. It was the person clicking, and every hour of that person's time was being billed to somebody.
Once execution moved into the software, the scope of what could be automated widened fast. Bids were always the easy target because they're numeric and reversible. The harder work stayed manual: writing the ad, matching it to a page that says the same thing, testing both against a control, killing the loser. All of that required generating something rather than adjusting a number. That's the part that opened up, and the interesting innovations of the last 18 months are almost all downstream of generation and execution arriving in the same system.
Fair warning before I go further: I work at groas, which sells autonomous Google Ads management, so I have an obvious stake in one side of this. I'll name where our approach is one of several defensible options and where I think the category is still pretending. What follows is the four shifts I'd pay for this year, the three that still fail in production, and how I'd sequence adoption if you're the one running the account.
The word "agent" got cheap fast. In an ads context it means something checkable: a model with write access to the account, a defined set of actions it can take, and a feedback loop that scores its past decisions against outcomes. A chat box that drafts headlines for you to paste in is not an agent. Neither is a rules engine with a language model narrating what it found. The test is boring and effective: ask what happens in the account if nobody logs in for three weeks. Real agents answer with a change log. Everything else answers with a backlog.
What makes this work now rather than in 2019 is that the actions got split up. Bidding, budget pacing, negative keyword mining, search intent classification, ad copy generation, and page mapping are different problems with different failure modes, and one model trying to do all six is worse at each than six models doing one. Our own build reflects that: separate agents for conversion copy, budgeting, search intent, opportunity discovery, and optimisation, with strategists supervising the accounts on top. The copy models are trained on over $500 billion in profitable search ad spend, which is less a bragging number than an explanation of why a generic language model writes ads that read well and convert badly. It has read the internet. It hasn't watched a million headline variants lose money.
The effect shows up in speed more than in cleverness. A human account manager working business hours across a book of clients touches an account a couple of times a week, usually Monday and Thursday. An execution layer running around the clock makes decisions on the hour, which matters most in the first 30 days of an account, when a wrong bid compounds daily. Across the accounts on our results page, the pattern is consistent: the biggest CPA moves land inside the first three to four weeks, not the first quarter. One US home services account spending $10,000 to $20,000 a month went from a $1,824 CPA to $364 in less than three weeks. That gap isn't superior strategy. It's the same decisions made far more often.
This is the shift I underrated longest. For most of my career the page was somebody else's department. I'd build 200 tightly themed ad groups, write copy matched to each one, then send all of that traffic to three landing pages because the web team had a sprint backlog and a redesign in flight. Every conversion rate problem I blamed on targeting for about four years was actually a promise mismatch: the ad said one thing, the page said something adjacent, and the visitor did the arithmetic in under two seconds.
What's new is that generating the page moved into the same system that writes the ad, so the mapping stops being a project. groas takes your existing page and deploys dynamic versions of it that adapt to each search intent, so someone searching "chest hair trimmer" reads a page about chest grooming rather than a general homepage. Mechanically it's the ambition behind dynamic keyword insertion, except it rewrites the headline, the body and the proof instead of jamming the query into an H1 and hoping. The reason it moves numbers is unglamorous: bid optimization changes what you pay per click, the page changes what fraction of those clicks become money, and the second number multiplies against every bid decision underneath it.
The tradeoff nobody puts in the deck: this only pays off if your offer is genuinely differentiable by intent. If you sell one product at one price to one buyer, generating 40 page variants gives you 40 ways to say the same thing, and the effort belongs in the offer instead. Regulated advertisers have a second problem, because a page that generates itself is a page your compliance reviewer hasn't read. Ask for an approval step before you point this at a legal or medical account. Adopt it if your keyword set spans distinct use cases. Skip it if your account is 30 keywords deep and all of them mean the same thing.
Ad testing used to be a calendar item. You'd build two variants, wait for significance, declare a winner in a monthly deck, and start again. That cadence made sense when a human wrote the variants and a human read the results. It stops making sense when the system generating the copy is also the one measuring it, because the loop closes in hours and runs on every ad group at once rather than the three you had time for. Our optimisation agents work like a team of data scientists running thousands of A/B tests simultaneously, around the clock, which is a description of throughput rather than genius.
The practical consequence is that your job changes from running tests to defining what winning means. If your conversion action is sloppy, continuous testing will find the copy that produces the most sloppy conversions, faster than any human ever could. I've watched an automated test converge beautifully on a headline that pulled in tire-kickers, because the account counted every form fill equally and nobody had told it otherwise. Decide what a good outcome is before you hand over the creative loop, and check what the loser variants looked like. If you can't see what was tested and what it cost, you don't have testing, you have churn.
The newest surface is the one nobody has ten years of playbooks for: paid placements inside AI assistants. We build for it alongside Google Ads, because a meaningful share of commercial queries now ends in a generated answer rather than a page of blue links, and the buying behaviour there breaks old habits. There's no result list to be third in. There's one answer, and either your offer is in it or the session ends.
What changes operationally is the unit of optimization. On search you optimize a keyword against a match type. On a generated answer you're optimizing for whether a model treats your product as a good fit for a described problem, which is closer to merchandising than to bidding. The advertisers who do well here will be the ones whose product data, page copy and offer are consistent enough that a model can describe them accurately without inventing a feature you don't ship. My advice for right now is not to chase the placement. It's to stop treating landing page copy as decoration and start treating it as structured input, because the clarity that helps a model recommend you also lifts conversion rate on the traffic you're already paying for. That's a bet with a floor.
First: fully automated cross-channel budget shifting. The pitch is that the system moves money to whichever channel is returning best this hour. The problem is attribution, not automation. Meta claims a conversion, Google claims the same conversion, GA4 splits it a third way, and a system reallocating on those numbers is optimizing a fiction with real money. Until your measurement is clean enough that you'd bet a quarter's budget on it, automated cross-channel shifting mostly launders your attribution bias into spend decisions. I'd automate hard inside Google, where the signal sits closest to the click, and keep channel-level allocation a monthly human call.
Second: chat-to-campaign builders. Type a sentence about your business, get a full account back. The generation part genuinely works now, and standing up structure from a landing page in minutes instead of two days is real time saved. What doesn't work is the implied promise that the build was the hard part. The hard part is the 90 days after launch, when the search terms report fills with garbage nobody anticipated and the conversion action turns out to be firing on the thank-you page and the contact page. A builder with no execution layer behind it hands you a tidy account and walks away at the moment the work starts. Judge these tools on day 45, not day one.
Third: predictive spend forecasting sold as a feature. Every dashboard now projects next quarter's CPA with a confidence band. I have yet to see one change a decision. A forecast is only useful if the system acts on it, and most of these are read-only charts sitting beside controls the tool can't touch. They also inherit whatever seasonality assumptions the vendor baked in, which is how you get a smooth December projected for an account that lives or dies on Black Friday week. If a forecast doesn't trigger a budget change you can inspect afterward, it's a slide, not a feature.
Sequence matters more than selection, because each of these changes the data the next one learns from. Fix measurement, then hand over execution, then let the system touch creative and pages. Do it in the other order and you'll optimize beautifully toward a conversion action that counts newsletter signups as sales. I've done that. It produced the best-looking report of my career and a client who couldn't understand why revenue was flat.
Test it the way you'd test anything else: with a control. Pick two or three campaigns worth roughly 30% of spend, hand those over, leave the rest running as they are, and give it 30 days against one primary metric agreed before you start. Not clicks, not impression share, not "efficiency." Cost per whatever actually pays you. Most accounts either move meaningfully in the first three to four weeks or expose a structural problem no automation can fix, and both of those are useful answers. The trap is a two-week read on a system still in a learning phase, which tells you close to nothing and usually ends with someone reverting changes at the worst possible moment.
Forget the feature grid. Two things predict whether a vendor is on the right side of this shift. The first is whether they own the execution layer or only recommend, because the ones that only recommend will keep shipping prettier recommendations while the cost of the human clicking apply stays exactly where it is. groas started as a recommendations tool that hundreds of operators ran on their own client accounts, and the thing that became obvious fast was that the human clicking the buttons was the bottleneck. The second is whether the vendor's commercial model survives the work getting cheaper. Charging a percentage of spend for automated execution is pricing 2015 labour, and that gap closes whether the vendor likes it or not.
So read the invoice as closely as the product. Ask what onboarding costs, how long the contract runs, and what happens in month seven when you want out. Ours is $0 to onboard, cancel anytime, with a 7-day free trial, and the weekly report lists every action taken rather than a summary of vibes. Set that against a $5k setup fee and a 6 to 12 month lock-in and you're no longer comparing features, you're comparing how confident each party is that results will keep you. Ask who the human is, too. We put a dedicated strategist on every account and keep the service fully managed below $25k/mo in spend, deliberately, because at that stage another login is not what anyone needs. Whatever the vendor's answer, make them name the person and the review cadence.
Here's where the category honestly stands: execution got automated, generation got good enough to trust with copy and pages, and measurement is still the weak link holding both back. That's why my adoption order starts with conversion tracking, and why I stay cautious about anything reallocating money across channels on numbers three platforms disagree about. If you do one thing after reading this, open your conversion actions and count how many feed your bid strategy. I'd bet at least one shouldn't. Fix that before you buy anything, and every shift above gets more valuable the same afternoon.
