AI in Paid Advertising Works Inside the Auction. The Loop Around It Decides the Gain.

AI in paid advertising already works, inside the auction. Google and Meta run models that price every impression, for every advertiser, continuously. What does not work is the business around that auction. The goal it is given, the result it gets back, the reply after the click, the tests and the memory all run at human speed. A loop moves only as fast as its slowest part, so most advertisers with all this AI still get very little from it.

• Meta reports that one new ads model, GEM, lifted ad conversions 5% on Instagram and 3% on Facebook Feed in a single quarter. The machine inside the platform is real.
• The platform optimises only the signal you send it. Google itself tells lead-generation advertisers to optimise for qualified or converted leads, and recommends at least 30 conversions in 30 days before target CPA bidding.
• We checked 156 live UAE Meta ads: of the 115 that showed a destination, 66 sent the click into a chat or a call, where the platform's reporting stops at the conversation.
• BCG found 60% of companies report minimal revenue and cost gains from AI, while 5% get real value. The 5% redesigned how the work is done.
• Electricity reached American factories in 1899, and productivity grew 0.2% a year until the 1920s redesign lifted it to 5.3% a year. Paid ads are at the same point now.

---

Factories had electricity for twenty years before they had the gain

Paul David, the Stanford economic historian, measured where electricity stood in American factories in 1899:

"less than 5 percent of factory mechanical drive" (Paul A. David, "The Dynamo and the Computer", 1990)

David described the early pattern. Owners swapped the steam engine for one large electric motor and kept the long line shafts, the belts and the building exactly as they were. The new power source sat inside the old way of working.

Productivity barely moved for two decades. David and Gavin Wright measured manufacturing productivity growth at 0.2% a year between 1909 and 1919. Between 1919 and 1929 it rose to 5.3% a year. The change in between was the redesign: a small motor on each machine, single-storey plants, work moving in one continuous line from raw material to finished product.

The electricity did not change in the 1920s. The layout did. David's 1990 paper, "The Dynamo and the Computer", made the same point about computers, and the point holds for AI in paid advertising today.

The AI inside Google and Meta ads is a parallel computing powerhouse

The AI inside the ad platforms is a massive parallel computing system. Meta Engineering reports that its Andromeda ads retrieval system raised model capacity 10,000x when it launched in December 2024. Meta Engineering also reports that its GEM ads model alone lifted ad conversions 5% on Instagram and 3% on Facebook Feed in the second quarter of 2025.

Google's Smart Bidding sets a bid at auction time, for each auction, using signals about the person and their context. These numbers are reported by the platforms themselves, so treat them as the platforms' claims. The direction is still clear. Inside the auction, the machine works per impression, for millions of advertisers at once, and it improves every quarter.

That machine has one limit, and it is the limit that decides your result. It learns only from what you send back.

Why do most advertisers get so little from AI in paid advertising?

Most advertisers get little from AI in paid advertising because the business around the auction still runs per month. The platform sees a click, a form or a purchase. It does not see the margin, the lead that sales rejected, the customer who came back, or the idea nobody tested. Five links stay slow, and the loop runs at the speed of the slowest one.

1. The goal. The bidder chases the event you defined. Google's own Performance Max guidance tells lead-generation advertisers to optimise for qualified or converted leads, because a form fill is a weak target. Schwartz, Bradlow and Fader tested this on a large bank's display ads across more than 750 million impressions: optimising for clicks instead of conversions would have cut new customer acquisition by about 10%.

2. The signal. The real outcome arrives late or never. A deal closes after six weeks. A job finishes after the invoice. A return comes back after the purchase was counted as a win.

   Google recommends at least 30 conversions in 30 days before target CPA bidding, and Meta's conversion leads optimisation asks for at least 200 leads a month with the lead's stage sent back within 28 days. A small advertiser with 40 real customers a month cannot feed the deeper signal at that volume.

3. After the click. The reply, the call, the checkout help and the follow-up sit with a person. A lead that waits two days for an answer is a different lead by the time anyone calls. The platform never learns that the problem was the reply, so it keeps buying more of the same leads.

4. The tests. One new idea a month is the usual pace in a small account. Optimizely's analysis of 127,000 experiments found that only 12% win on their primary metric.

   Koning, Hasan and Chatterji followed 35,000 start-ups and found that the ones that adopted A/B testing saw a persistent 11% rise in page views and launched new products at a 9% to 18% higher rate.

   With so few winners, the number of tests decides how many winners you find.

5. The memory. What last year's campaigns taught lives in a report, an inbox or an agency's head. When the agency or the marketer changes, the lesson goes with them, and the next team starts from the platform's defaults again.

We checked where live UAE ads send the click. On 6 November 2026 we read 156 active Meta ads from 109 advertisers in Meta's public Ad Library, across online stores, B2B services and local services. Of the 115 ads that showed their destination, 66 sent the click into a WhatsApp, Messenger or Instagram chat, or a phone call. For local services it was 33 of 44, and for B2B services 25 of 40.

In a chat, the platform's own reporting stops at "conversation started". Whether that chat became a customer reaches Meta only if the business sends the result back. The count is one day's snapshot, and the Ad Library's default order leans toward bigger advertisers.

Each of these links runs per week or per month. The auction runs per impression. That mismatch is where the gain from AI in paid advertising is lost.

Is the platform's own measurement enough to know what worked?

The platform's own measurement is not enough to know what worked, because it grades its own results. Blake, Nosko and Tadelis ran experiments at eBay and found that brand keyword ads had no measurable short-term benefit, since those buyers would have arrived without the ad. Gordon and colleagues compared 15 large Facebook advertising experiments with the usual observational methods and found those methods frequently gave the wrong answer.

The fix is not a better dashboard. It is a holdout: a group that does not see the ad, so the difference can be measured. A loop that never holds anything out keeps learning from its own claims.

---

A closed loop in paid media runs all the time, with all the context in place

A closed loop in paid media is one continuous system that runs all the time, with all the business context in place. It does five things without waiting for the monthly meeting:

• It finds the next opportunity from what customers search for, ask about and buy.
• It runs many experiments at the same time, each with a stated bet and a stop condition.
• It follows every lead or order after the click, through the reply, the sale and the repeat purchase.
• It sends that real result back to Google and Meta the same day, so their AI learns which customers are worth paying for.
• It judges each experiment against a holdout, keeps what won, records why, and starts the next round from that record.

The same loop looks different in three businesses. For an online store, the result sent back is the order value after returns and the second purchase, so the bidder stops paying for buyers who return everything.

For a B2B company, it is the qualified opportunity and the closed deal, so the bidder stops paying for cheap leads that sales never calls. For a service business, it is the confirmed booking and the finished job, and the reply to every enquiry happens within minutes.

At Ivanooo we call this the Growth Factory. Google and Meta cannot switch it on for you, because their AI stops at the auction. The rest of the loop runs inside your business.

Ivanooo practises and teaches Growth Engineering: engineering the context, direction, distinctiveness, experiments and evaluation above AI marketing execution, so a brand's output is distinct and aimed at a growth goal.

In paid media, three parts of that practice carry the loop. The Growth Graph is one living record of the business that every agent reads before it writes, decides or spends. The Experiments layer chooses the few moves worth making and proves them. The Evaluation layer proves whether what should have happened did happen, against a holdout. The Growth Engineering hub explains each part.

The five levels of AI in paid media

Level	What runs by itself	In paid media	Who provides it
1 · Assisted	AI helps with tasks	AI helps write ad copy and briefs	Any AI writing tool
2 · Delegated	AI makes the work, a person checks it	AI builds and places the ads; a person reviews each one	Google and Meta automated campaigns
3 · Self-checking	AI makes and checks against your rules	Written rules check every ad, lead and landing page	The business, with its own rules
4 · Self-optimising	Results against your goal change what runs	The real outcome, the paying customer, scales or stops campaigns	The business, sending results back
5 · Growth Factory	The whole loop runs itself	Finds opportunities, runs many bets, learns from customers after the click, starts again	The business and its system together

Google and Meta give most advertisers level two for free. Levels three to five depend on what the business sends back, writes down and tests, which is why the same platform produces very different results for two advertisers with the same budget.

Redesign is harder than switching on AI

Switching on Performance Max or Advantage+ takes an afternoon. Redesigning the loop changes who replies to leads, what the sales team records, how results reach the platform, how many ideas get tested and where lessons are kept. That is work across marketing, sales and operations, which is why most businesses stop at level two.

BCG's 2025 study of 1,250 executives found that 60% of companies get minimal revenue and cost gains from AI, while the 5% it calls future-built are reshaping their workflows around it.

Brynjolfsson, Rock and Syverson call this the productivity J-curve: a general-purpose technology needs large complementary investments in process before the gains show. The factories of 1919 and the ad accounts of today are the same case.

How do you find where your own loop stops?

Take your largest campaign and answer five questions in writing. Each answer is a time, a name or a number.

1. What event does the bidder chase today, and is it the result you are paid for?
2. How many days pass before the real outcome, the sale, the qualified lead or the finished job, reaches Google or Meta?
3. How many minutes pass between a lead arriving and a person or system replying?
4. How many separate ideas did you test last month, and how many had a holdout?
5. Where are last year's lessons written down, and would a new agency find them on day one?

Wherever the answer is "a person, later" or "nowhere", that is the slow link. Fix the slowest one first, because the loop cannot move faster than it. Our paid ads practice guide covers each link in more detail, and Agentic Marketing Platforms in 2026 scores where current tools stop.

---