Homemade Pasta · Research

The machine runs the campaign now

On Meta and Google Search, about a third of ad spend now goes through campaigns where the platform picks the audience, the placement, the bid and more of the creative. Around those campaigns, AI agents now do much of the operator’s work. Ours build and change live campaigns on Google and Meta. The result now depends on a handful of inputs the advertiser still controls, on the people who approve the tools’ changes, and on tests the platform does not grade.

1 in 3 of Meta ad revenue and Google Search spend runs through fully automated campaigns
PublishedSeptember 2026
ReadAbout 22 minutes

The question

How much to hand over, and where people fit

Performance Max, Advantage+ and AI Max take a bigger slice of paid media budgets every year. Clients put two questions to Homemade Pasta about them. How much should we hand to the platform? And once it does the buying, what are we paying people to do?

To answer both we read the platforms’ disclosures, the spend data agencies and vendors publish, the economics of automated bidding, and the research on how work changes when software takes over tasks.

The page looks at two layers. The first is automated campaign types, where the platform decides audience, placement, bidding and budget inside the campaign and the advertiser supplies goals, assets and limits. The second is AI tools that do the operator’s work around the campaign: building it, launching it, writing the copy, checking it and reporting on it. Programmatic trading and automated bidding have been routine for a decade, so neither counts.

In short

A third of the budget, most of the decisions

30%+
of spend on the two biggest ad platforms sits in fully automated campaigns.

Meta reported a run-rate of over $75 billion through Advantage+ end-to-end campaigns in July 2026, about 32% of its annualized ad revenue. Alphabet says more than 30% of its customers’ Search spend goes through AI Max or Performance Max. Retail shopping runs higher. Across all US advertising, one analyst puts the share at 12%, heading for 27% by 2030.

2
settings decide the margin: the value the platform is told to chase, and the target.

When the target is the binding limit, the platform’s own measure of value and the target you set cap the spend between them. The advertiser controls both inputs. Evidence on automated against manual campaigns is thin and points both ways, and we found nothing independent showing that manual overrides beat automated bidding in general.

5%
was the measured lift where standard models estimated 24% to 64%.

That gap, the median for lower-funnel outcomes across 663 randomized Meta experiments, is why the platform’s own reporting cannot settle whether automated spend worked. Its A/B tests can also show each version to different people. The check has to come from outside.

50%
more ads per person when people worked with AI, in a field experiment with 11,024 ads.

Homemade Pasta’s agents build and change live campaigns on Google and Meta through the platforms’ APIs, with a senior person approving every change first. Meta and Amazon also opened ad accounts to AI agents in 2026. That takes much of the execution load that used to sit with junior staff. AI agents in general still fail often enough on multi-step work that anything touching live spend needs a person to approve it.

Share of spend in fully automated campaign typesLatest figure for each, different bases
Google shopping, retailers running both campaign types, Q4 202562%
Meta advertising revenue, Q2 202632%
Google Search spend, Q1 202630%+
All US advertising, 2026 estimate12%

One agency’s retail sample, our arithmetic on Meta’s run-rate, Alphabet’s statement and an analyst estimate. Part One takes each in turn.

People still own the work on either side of the machine: the value signal, targets and cross-channel budgets, creative, guardrails, approvals and testing. Those jobs cross disciplines and lean on judgment. Our reading is that they suit small senior teams whose people cover media, creative and measurement between them, with AI tools handling execution under their approval. That is how Homemade Pasta is set up, and Part Four lays out why.

Decisions

Eight decisions to keep

Give the platform the execution. Hold on to these.

  1. The conversion, and its value.

    Bidding optimizes whatever you report. In display research, clicks turned out to be a poor stand-in for purchases, and a validated long-term surrogate beat short-term proxies. Report the revenue or margin the business actually earns, even when a cheaper event is easier to track.

  2. The target, set from incremental return.

    A reported ROAS sitting on target overstates return by the share of conversions that would have happened anyway. The calculator in Part Two converts a target and a measured incremental share into the return the business actually gets.

  3. Budget across channels and campaign types.

    Every automated campaign optimizes inside its own walls. When Performance Max and a Search campaign chase the same query, Google’s serving rules pick the winner, whatever the plan intended.

  4. The creative the system assembles from.

    The platform can only choose among the assets it gets, and their message drives most of the outcome. Our creative research covers the evidence.

  5. Guardrails, set deliberately.

    Campaign-level negative keywords, search terms reporting and channel reporting all arrived in Performance Max in 2025. People used algorithms more, and forecast better, when allowed small adjustments. In one 15-firm hiring study, overriding the test made outcomes worse.

  6. An independent read on whether it worked.

    A test that randomizes who sees the ads, scored outside the platform. Sizing one is covered in our incrementality research.

  7. The AI tools’ permissions.

    Agents can now build campaigns and write copy; ours already do on Google and Meta. The platforms put an approval step in front of live spend, and we put one in front of every change. Hold your own tools to the same rule, and aim review at targeting, goals and claims, the kinds of error reviewers miss most.

  8. Who makes the other seven.

    Each decision above touches media, creative and measurement, and each goes better with someone who can tell when the machine is wrong. We think that favors a small team of senior people who cover all three, working with AI tools, over a large team of specialists.

Part One

How much runs on autopilot

Platform disclosures, independent spend data, and the direction of travel.

How much

About a third of the two largest platforms

Neither Meta nor Google states a share of spend outright. The figures here come from their earnings calls, and where the share is our arithmetic, the working is shown.

Meta

In October 2025 Meta said “the annual run rate of revenue running through our end-to-end automated solutions has now reached $60 billion.” In July 2026 it reported Advantage+ end-to-end campaigns at “over $75 billion in annual revenue run-rate.”1 Meta’s advertising revenue for that quarter was $59.4 billion, or about $237.5 billion annualized. $75 billion is 31.6% of that. The same calculation for October 2025 gives 30.0%.

Meta has not said how the run-rate is counted, and has said that many advertisers use these campaigns for only some of their campaigns. The figure may count the full revenue of every campaign with end-to-end automation switched on, which Meta made the default setting during 2025.

Google

In April 2026 Alphabet said “more than 30% of our customers’ Search spend now uses AI enabled campaigns AI Max or Performance Max.” In July it said AI Max was out of beta with half a million advertisers.2 Performance Max also serves on YouTube, Display, Gmail and Maps, so this is a share of Search spend only.

Independent data

Agencies that publish benchmarks from the accounts they run see a higher share in retail. Tinuiti reported that among its retail clients running both Performance Max and Standard Shopping, Performance Max took 62% of Google shopping investment in the fourth quarter of 2025. Advantage+ shopping campaigns took 27% of their Meta spend.3 Haus, a measurement vendor, put Advantage+ at roughly 39% of spend in the typical Meta portfolios of its clients, large advertisers, in 2025.4

Across all US advertising, Madison and Wall estimates that 12% of spend in 2026 runs where “AI controls targeting, bidding, budget allocation, placement, and ongoing optimization with minimal human intervention,” about $57 billion. The method is described only in outline, so it is best read as one analyst’s estimate.5

How fast

Steady at the platforms, still spreading across the market

Share of US ad spend in fully automated buyingAnalyst estimate, 2030 projected
2023202520262030 proj.

Madison and Wall, February and September 2026. The 2030 figure is a projection, revised up from 26% in the September update.

Across the market, the estimate has the share rising from 2% in 2023 to 12% this year and 27% by 2030, mostly by taking spend from other digital buying.5

Inside the two big platforms, the share has barely moved lately. Meta’s run-rate grew from $60 billion to $75 billion over three quarters, but its share of Meta’s advertising revenue barely moved, from 30.0% to 31.6%. Tinuiti’s Performance Max share of retail shopping spend ranged from 53% to 69% across the five quarters to the end of 2025, with no clear trend, though Tinuiti changed how it measured the share during the year.3 The same benchmark showed Advantage+ shopping campaigns falling from 38% of retail Meta spend in the first quarter of 2025 to 27% by the fourth, while Meta’s own figure kept rising. Meta made Advantage+ the default and merged its campaign types during that year, so part of the gap is probably how each side classifies a campaign.

Search keywords have drifted the same way. Across 30,000 Google accounts, Optmyzr found broad match had become the largest match type by budget, with exact match’s share down nearly 10 points between 2022 and 2026.6 AI Max extends that looser matching further.

Part Two

Telling the machine, and checking it

The two settings behind the margin, the automated-versus-manual evidence, and why platform reporting cannot settle it.

The levers

Two settings decide the margin

When the target, and not the budget, is the binding limit, target-based bidding spends until the return the platform reports falls to the target. At that point spend is capped at the value the platform reports, divided by the target. Google’s own researchers describe the same ceiling from the other side: platform revenue is capped by the total value advertisers can pay within their budgets and targets, and reaches the cap when each advertiser’s budget or target binds.7

That leaves the advertiser two settings with direct control over margin. One is the definition of value: which conversions count and what each is worth. The other is the target.

A study of display targeting found that “clicks are not good proxies for evaluation nor for optimization: buyers do not resemble clickers,” while visits to the brand’s site tracked purchases well.8 A 2024 study found that targeting on a validated surrogate for long-term outcomes beat short-term proxies, and at the Boston Globe was estimated to add $4 million to $5 million in net revenue over three years against the paper’s existing policy.9 Google’s own figure for moving from a cost-per-acquisition target to a return target is a median 14% increase in conversion value, from its internal data.10

An automated campaign pointed at the easiest event to track will get very good at producing that event. Whether the event is worth anything is a decision the platform leaves to you.

Your target

Reported return against real return

A platform hitting a 4× target is reporting $4 of revenue for every $1. Some of that revenue would have arrived without the ad. Enter your target, the share of reported conversions that are incremental, and your margin.

Incremental return at your target

Assumes the campaign is spending at its target, as target-based bidding does when the target binds.

Incremental share
4.0×
Reported revenue per $1 of spend
69%
From a holdout or geo test on your account
30%
Share of each sale left after cost of goods
Incremental ROAS 2.76× $0.83 gross profit per $1 spent
Close to break-even

The platform reports $4.00 for every $1. The ads created $2.76 of it, which returns $0.83 in gross profit. Breaking even at this incremental share takes a target of 4.8×.

Reported revenue per $1$4.00
Would have happened anyway$1.24
Break-even target4.8×
Incremental ROAS = target × incremental share. Gross profit per $1 = incremental ROAS × margin. Break-even target = 1 ÷ (margin × incremental share). The presets are the share of seven-day last-click conversions on Meta that proved incremental across 2,226 experiment and conversion pairs, from November 2019 to March 2020: 48% in travel, 69% across the full sample, 76% in e-commerce. Reported ROAS usually also counts view-through conversions, so against reported ROAS these presets likely overstate the incremental share.11 Your own share comes from a test on your own account.

Set the target from the incremental return the business needs, then confirm it with a test. A target set from reported return will be hit, and the account can still lose money doing it.

Versus manual

Automated against manual: thin evidence, both ways

The best comparison we found uses real holdouts but comes from a vendor. Across 640 Meta geo experiments run since 2024, Haus reported that 58% of brands saw higher incremental return from manual campaigns than from Advantage+ in head-to-head tests, and that Advantage+ delivered 12% lower incremental return on direct-to-consumer sales at 18% lower daily spend. Haus has not said how many of the 640 were head-to-head, and it sells incrementality testing.4

On Google, the available data measures overlap, not outcomes. Optmyzr found search-term overlap between Performance Max and Search in 97% of 511 accounts. On 424,820 overlapping terms, conversion rates were within 10% of each other 88% of the time, and return could not be compared because Performance Max does not report revenue by search term.12 Google’s rules decide who serves: a Search keyword that exactly matches the query is preferred over Performance Max, and since October 2024 Performance Max and Standard Shopping campaigns on the same products compete on Ad Rank.13

Where automated campaigns reach branded queries, the value depends on the competition. On eBay, brand-keyword ads had no measurable short-term benefit.14 On Bing, brand ads added 1% to 4% when no competitor was bidding. When competitors were bidding and the brand was not, competitors took 18% to 42% of the brand’s clicks.15

We found no independent evidence that manual management beats automated bidding in general, and no strong evidence the other way. The split that works for an account is found by testing that account.

Checking the work

The platform grades its own homework

Automated campaigns optimize toward conversions the platform can see and attribute. How much of that is incremental is a separate question, and the platform’s reporting is poorly placed to answer it.

Estimated lift against measured liftMedian across 663 Meta experiments, lower funnel
Randomized experiment5%
Double machine learning24%
Propensity matching64%

The same gap held further up the funnel: 29% measured against 83% and 173% at the top, 18% against 58% and 176% in the middle.

Across 663 randomized experiments on Meta, the usual observational methods, fed more than 5,000 features per user, overstated lift at every stage of the funnel.16 Platform A/B tests fail in another way. The delivery system decides who sees each version, so the two versions reach different people, and that can shrink, inflate or flip a result.17 When researchers checked A/B tests run on Meta’s tools, the two audiences differed by more than a common threshold in 22% of comparisons. In randomized lift tests the figure was 0.16%.18

Meta now offers “incremental attribution”, which it describes as machine learning models that predict whether a conversion was caused by an ad. We found no published validation of those predictions.19

Holdout and geo tests randomize exposure and leave the platform out of the scoring. Once the platform runs the campaign, it is the only check that belongs entirely to the advertiser. How to size one is in our incrementality research.

Part Three

The tools that do the work

How our agents build and change live campaigns, where the platforms stand, how good the output is, and why a person still signs off.

How we run it

Our agents build and change live campaigns on Google and Meta

At Homemade Pasta, AI agents work directly in client ad accounts on Google and Meta, with full write access through the platforms’ APIs. They make the changes themselves, and a senior person approves each one before it runs.

Google’s own AI connector is read-only, and Meta’s opened only this year. The underlying Google Ads API and Meta Marketing API have long accepted changes, and that is where our agents work. They create campaigns, ad groups and ad sets. They write and upload ads, set budgets, bids and targeting, build audiences and conversion actions, and pull the reporting afterward.

That takes the building, trafficking and first-draft work off people’s desks, much of what used to fill a junior buyer’s week. It does not take the decision. Nothing that touches live spend, or publishes copy under a client’s name, runs until a person has approved it.

  1. Propose

    The agent drafts the change from the plan: the campaign build, the copy, the budget, the bid limits.

  2. Approve

    A senior person on the account checks the targeting, goals, budget and claims, then approves it or sends it back.

  3. Execute

    Only then does the agent make the change in the live account.

It is the same rule the platforms have started writing into their own agent tools, applied to everything we run.

Where the platforms are

The platforms are opening their accounts to agents too

This second layer of automation arrived in 2025 and 2026, separate from the campaign types in Part One. Alongside API access, each major platform now offers its own route for AI tools into an ad account.

PlatformCapabilitiesStatus
GoogleAn official connector lets AI tools read and report on an account, and Google’s documentation says it cannot change bids, pause campaigns or create assets. Inside Google Ads, Ads Advisor can apply recommended changes, with the advertiser’s review and approval, and generate keywords and assets. Ask Advisor, a unified successor, is described as generating campaigns end to end.20Connector released October 2025, read-only. The Google Ads API itself accepts changes. Ads Advisor available to English-language accounts globally since April 2026. Ask Advisor announced May 2026, not yet broadly available.
MetaAds AI Connectors link an ad account to the advertiser’s own AI agent, which can create campaigns, ad sets and ads. New objects are created paused, and Meta’s documentation has the AI client ask for confirmation before anything starts spending.21Open beta since April 2026.
AmazonA connector that lets AI tools create, update and delete campaigns, and an Ads Agent that builds DSP campaign structure from a media plan. Amazon says the Ads Agent’s campaigns launch only after the advertiser reviews and approves.22Connector in beta for API partners since February 2026. Ads Agent since November 2025, US only for DSP.

Creative generation is further along. Meta says advertisers using its generative AI creative tools grew from more than 1 million in October 2024 to more than 4 million in January 2025 and more than 8 million in April 2026.23 Alphabet said advertisers used Gemini to create nearly 70 million text assets through text customization in AI Max and Performance Max in the fourth quarter of 2025 alone.24 Meta’s stated aim, in Zuckerberg’s words, is that “any business can basically tell us what objective they’re trying to achieve” and what it will pay for each result, “and then we just do the rest.”23

How good it is

Faster, and good enough to need an editor

The closest evidence to ad production comes from a field experiment in which 2,234 people made 11,024 ads that then ran on X across about 5 million impressions. People working with AI produced 50% more ads each, with higher-quality text. People working without it made better images. The AI-assisted ads also looked more alike.25 That last finding matters for rotation, which only works when the executions differ.

Other studies point the same way on speed. In a writing experiment with 453 professionals, ChatGPT cut time spent by 40% and raised quality by 18%.26 In field experiments on search content, machine drafts edited by people outranked content written by SEO experts, at much lower cost, and the authors concluded that the human editor “remains essential.”27

The platforms make their own claims for AI-made creative. Meta estimated a 7% increase in conversions for businesses using its image generation, and more than 3% higher conversion rates in tests of video generation.23 Meta has not published the method behind either figure.

Time savings are easy to overestimate. In a randomized trial with 16 experienced software developers, AI tools made them 19% slower, while the developers believed afterward that they had been 20% faster.28 The savings from AI tools are worth measuring in each team, since they vary more than the headline studies suggest.

Why a person signs off

Mostly right isn’t good enough for live spend

Agents are improving fast and are still unreliable on their own. On a benchmark of simulated workplace tasks, the best agent in the September 2025 results fully completed 30% of them.29 On a customer-service benchmark, the best model tested in 2024, GPT-4o, succeeded on under half the tasks, and on retail tasks it succeeded on all eight repeated attempts less than a quarter of the time.30 A tool that gets a task right once can get the same task wrong on the next try.

The frontier is moving. METR, which measures the length of task an agent can complete half the time, put the leading model at about five hours in January 2026, and longer since, with that length doubling every few months. METR also notes that tasks where errors are costly and hard to check may need a success rate of 98% or more before they are worth automating.31 Changes to live budgets, bids and targeting fall in that group.

Review needs to aim at the right errors. In an experiment with 2,784 people checking AI-extracted data, reviewers caught 82% of surface errors, such as mixed-up digits, but only 31% of errors that required understanding the rules, and those who viewed AI more favorably were less accurate.32 For ad work, the costly errors are usually the second kind: the wrong audience, the wrong conversion goal, a claim the brand cannot make.

The platforms build this in themselves. Google’s Ads Advisor applies changes with the advertiser’s approval, Meta’s connector creates campaigns paused, and Amazon’s Ads Agent waits for sign-off before launch. Wherever AI tools change live spend or publish copy under a brand’s name, the same rule should hold: the tool proposes, a person who knows the account approves, and only then does the change run.

Part Four

The job that’s left

How tasks divide between platform, tools and people, how to steer a system you do not operate, and the team that fits.

Who does what

A new division of labor

Economists describe automation as two effects at once: it takes over existing tasks, and it creates new ones where people have the advantage.33 Most of today’s US employment sits in job specialties that did not exist in 1940.34 Paid media is going through the same shift inside a single job.

TaskNow handled byWhy
Bids and budget pacing inside a campaignPlatformTarget-based bidding sets every bid against the platform’s own prediction of value.
Audience selection and placementPlatformFully automated campaign types choose both, inside whatever limits they are given.
Combining assets into adsPlatform, increasinglyThe platform assembles and rotates combinations from the assets supplied.
Building and launching campaigns from a planAI tools, with approvalAgents create the structure and settings; a person approves before anything spends. See AI tools, in Part Three.
First-draft copy and variantsAI tools, with editingFaster output and better text in field tests, with more sameness across ads.
Settings checks and reporting pullsAI tools, with reviewRoutine, repetitive and easy to verify, which suits automation.
Approving what the tools changePeopleReviewers miss errors of meaning more than surface errors. See oversight, in Part Three.
Defining value and conversionsPeopleThe system optimizes what it is told to value. See the levers, in Part Two.
Targets, and budgets across channelsPeopleEach campaign sees only its own results, and the target sets the margin. See the calculator.
Creative strategy and supplyPeopleThe platform chooses among the assets it has, and their message decides most of the result.
GuardrailsPeopleExclusions, negative keywords and new-customer rules are fixed before the auction starts. See guardrails, below.
Measuring incrementalityPeopleThe one read on performance the platform does not score. See checking the work.
Noticing when the system is wrongPeopleTakes a clear idea of what the result should have been. See experience, below.

This split is our reading, and no study has measured it in paid media. It does match the wider research, in which execution moves to software and the work that grows sits on either side of it.

Guardrails

Set the limits in advance, then leave it alone

People adopt an algorithm far more readily when they can adjust it a little. In a forecasting experiment, about seven in ten participants chose to use an imperfect algorithm when they could change its forecasts by as little as two points, and their forecasts improved as a result.35 Unstructured overrides have a worse record. Across 15 firms, managers who appeared to hire against the recommendation of a job test ended up with worse hires on average.36

For automated campaigns, the useful human input is a limit set in advance: an exclusion, a negative keyword list, a value rule, a target. Changing individual decisions week by week works against a system that learns from its own history.

Google added several of those limits to its automated campaigns in 2025.37

  • January. Campaign-level negative keywords in Performance Max, rolling out to all advertisers, with the limit raised to 10,000 in March.
  • April. Channel performance reporting announced for Performance Max, and search terms reporting rolling out.
  • May. AI Max for Search launched with brand and URL controls and search terms reporting. Google also announced a Performance Max beta, due later in the year, that excludes people who recently searched for or interacted with the brand.

Controls arrive and change on the platform’s schedule. Checking which ones exist, and setting each deliberately, is recurring work.

Experience

Experience counts most where the machine is wrong

Field evidence on AI at work finds it helps less experienced people most. Among 5,172 customer support agents, AI assistance raised productivity 15% on average and about 30% for less skilled and less experienced agents. Agents two months in, with AI, performed as well as agents more than six months in without it.38 In paid media, we read the closest equivalent as routine execution: bidding and pacing, which the platform now handles, and building campaigns, trafficking, first-draft copy, checks and reports, which AI tools increasingly handle. That is most of a junior’s week.

The remaining work sits closer to the tasks where AI help has been shown to fail. In an experiment with 758 consultants, AI users completed 12% more tasks, 25% faster and at over 40% higher quality on work inside the AI’s capabilities. On a task outside them, consultants using AI were 19 percentage points less likely to reach the right answer.39 Experience also buys some protection when the system is wrong. In a mammography study with 27 radiologists, inexperienced readers were almost 80% accurate when the AI’s suggestion was right and under 20% when it was wrong. The five most experienced readers, about 15 years in radiology on average, fell from 82% to 45.5%.40 Every group was affected. The most experienced lost about 37 points, against about 60 for the least.

Noticing when an automated campaign is wrong depends on knowing how a conversion should be set up, how a sound test is designed, and roughly where a plausible incremental return lands. That knowledge comes from having done the work before the platform did it.

The model

Small senior teams, each person working across the disciplines

The tasks that grow in the table above cross the usual departments. Setting a value signal is part analytics and part commercial judgment. A target depends on a measured incremental share. Guardrails depend on knowing both the account and the brand. Creative strategy supplies the material the platform works with. Handing each task to a separate specialist puts a handoff between every decision and the one it depends on.

One of the largest field experiments on AI and teamwork points the same way. In a preregistered study with 791 professionals at Procter & Gamble working on real product innovation problems, individuals using AI matched two-person teams working without it. Without AI, R&D staff leaned toward technical solutions and commercial staff toward commercial ones. With AI, both produced balanced solutions whatever their background. The pairs using AI were the most likely to produce top-tier solutions.41 We take that to mean narrowing the gap between functions lets one person cover more of them, and that a small team with AI does better still.

Taken together, we read the evidence as favoring a particular shape: fewer people, each senior enough to catch the machine when it is wrong, each working across media, creative strategy and measurement, with AI tools doing the building, trafficking and first drafts that used to need a junior layer, and a person approving their changes. Nobody has tested that shape in paid media. Homemade Pasta is built around it because this is where the evidence leads.

Appendix

Notes and sources

Caveats first, then all 41 sources.

Notes

Limits of the evidence

Every share figure in Part One is a platform statement, one agency’s sample or an analyst’s estimate. Meta has not explained how it counts its run-rate, and no independent panel weights spend across advertisers and platforms.

We found no independent controlled study of how value signals, offline conversion imports or first-party data change automated campaign results. The value evidence in Part Two comes from display targeting, one newspaper and Google’s own data.

The only holdout-based comparison of automated and manual campaigns is a vendor’s, and the vendor has not disclosed its head-to-head count.

The AI tools described in Part Three change quickly. Several are in beta or only announced, and agent benchmarks date within months, so each figure carries its date. We found no rigorous study of how much time AI tools save on routine account management specifically.

Part Four leans on research into automation and AI at work in other fields, none of it in paid media. The task split and the team model are our reading of it.

The calculator presets come from Meta experiments run between November 2019 and March 2020, before app tracking restrictions and before today’s campaign types.

Sources

Reading list

All 41 sources, with what each supports
1Meta Platforms, third quarter 2025 earnings call, October 29, 2025, and second quarter 2026 earnings call, July 29, 2026, with advertising revenue from the accompanying earnings releases.Platform disclosureSource of the $60 billion and $75 billion run-rate statements and of quarterly advertising revenue of $50.1 billion and $59.4 billion. The 30.0% and 31.6% shares are our calculation: run-rate divided by four times quarterly advertising revenue.
2Alphabet, first quarter 2026 earnings call, April 29, 2026, and second quarter 2026 earnings call, July 22, 2026.Platform disclosureSource of the statement that more than 30% of customers’ Search spend uses AI Max or Performance Max, and of AI Max adoption by half a million advertisers.
3Tinuiti, Digital Ads Benchmark Report, quarterly editions from the fourth quarter of 2024 to the fourth quarter of 2025.Agency dataDrawn from the retail accounts Tinuiti manages. Source of the Performance Max share of Google shopping investment among retailers running both campaign types, and the Advantage+ shopping share of retail Meta spend. The base for the Performance Max share changed during 2025. Tinuiti sells management of these campaign types.
4Haus, “The Meta Report: Lessons from 640 Haus Incrementality Experiments,” July 2025, and “Is Meta Incremental?”, August 2025.VendorSource of the 39% Advantage+ share of typical portfolios and the head-to-head comparison with manual campaigns. The number of head-to-head tests is not disclosed. Haus sells incrementality testing.
5Madison and Wall, “How AI-Powered Advertising Totals $142 Billion By 2030,” February 2026, commissioned by Adobe, with the update reported by Digiday, September 10, 2026.Analyst estimateSource of the 2%, 8%, 12% and 27% shares of US advertising and the definition quoted in 1.1. Method described only at a high level.
6Optmyzr, analysis of match-type spend across 30,000 Google Ads accounts, May 2026, using February 2026 data.VendorSource of the finding that broad match is the largest match type by budget and that exact match’s share fell by nearly 10 points from 2022 to 2026.
7Aggarwal and colleagues, “Auto-bidding and Auctions in Online Advertising: A Survey,” arXiv 2408.07685, August 2024.Platform researchWritten by 26 Google researchers. Source of the statement that platform revenue is capped by the value advertisers can pay within their constraints, with equality when the constraints bind.
8Dalessandro, Hook, Perlich & Provost, “Evaluating and Optimizing Online Advertising: Forget the Click, but There Are Good Proxies.” Big Data 3(2), 2015, 90–102.Peer reviewedSource of the finding that clicks are poor proxies for purchases and that brand site visits are a good one.
9Yang, Eckles, Dhillon & Aral, “Targeting for Long-Term Outcomes.” Management Science 70(6), 2024, 3841–3855.Peer reviewedSource of the surrogate targeting finding and the $4 million to $5 million estimate at the Boston Globe.
10Google Ads, value-based bidding guidance, citing Google internal data from March 2021.Platform claimSource of the median 14% increase in conversion value from moving from target CPA to target ROAS. No method published.
11Gordon, Moakler & Zettelmeyer, “Predicted Incrementality by Experimentation (PIE) for Ad Measurement.” NBER Working Paper 35044, April 2026.Working paper2,226 experiment and conversion pairs from Meta, November 2019 to March 2020. Source of the calculator presets: travel, the full sample and e-commerce, three of the verticals the paper breaks out. Co-authored with a Meta researcher on Meta data.
12Optmyzr, “Is PMax Cannibalizing Search?”, July 2025.VendorFebruary 2025 data. Source of the overlap and conversion-rate comparisons. Optmyzr treated differences under 10% as insignificant, without a statistical test, and sells Performance Max tooling.
13Google Ads Help, Performance Max serving priority, and the October 2024 change to how Performance Max and Standard Shopping campaigns compete.Platform documentationSource of the exact-match priority rule and the Ad Rank rule.
14Blake, Nosko & Tadelis, “Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment.” Econometrica 83(1), 2015, 155–174.Peer reviewedSource of the eBay brand-keyword finding.
15Simonov, Nosko & Rao, “Competition and Crowd-Out for Brand Keywords in Sponsored Search.” Marketing Science 37(2), 2018, 200–215.Peer reviewedSource of the 1% to 4% and 18% to 42% figures.
16Gordon, Moakler & Zettelmeyer, “Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement.” Marketing Science 42(4), 2023, 768–793.Peer reviewed663 experiments. Source of the measured and estimated lift figures. Co-authored with a Meta researcher on Meta data.
17Braun & Schwartz, “Where A/B Testing Goes Wrong: How Divergent Delivery Affects What Online Experiments Cannot (and Can) Tell You About How Customers Respond to Advertising.” Journal of Marketing 89(2), 2025, 71–95.Peer reviewedSource of the explanation of divergent delivery in platform A/B tests.
18Burtch, Moakler, Gordon, Zhang & Hill, “Characterizing and Minimizing Divergent Delivery in Meta Advertising Experiments.” MSI Working Paper 25-140, November 2025.Working paper3,204 lift tests and 181,890 A/B tests on Meta. Source of the balance comparison. Four of the five authors are employed by or contract with Meta.
19Meta’s description of incremental attribution, as reported by Social Media Today, September 2025.Platform claimSource of the description of the feature. No validation method published.
20Google: “Open Source Google Ads API MCP Server,” Google Ads Developer Blog, October 7, 2025, and its developer documentation; “Google’s AI advisors: agentic tools to drive impact and insights,” November 12, 2025, and “3 new ways Ads Advisor is making Google Ads safer and faster,” April 21, 2026; and the Ask Advisor announcement at Google Marketing Live, May 2026.Platform documentationSource of the read-only status of the official connector, the capabilities and rollout of Ads Advisor, and the description of Ask Advisor. Ask Advisor’s availability was described inconsistently across Google pages at the time of writing.
21Meta, first quarter 2026 earnings call, April 29, 2026, and the Meta Ads AI Connectors and Ads MCP server developer documentation.Platform documentationSource of the open beta, the create-campaign, ad set and ad capabilities, the paused-by-default rule and the confirmation step before activation.
22Amazon Ads, MCP server open beta announcement, February 2, 2026, and Ads Agent and Creative Agent announcements, November 11, 2025.Platform documentationSource of the create, update and delete capabilities and the review-and-approve requirement before launch.
23Meta Platforms earnings calls: October 30, 2024; January 29, 2025; April 30, 2025; and April 29, 2026.Platform claimSource of the generative AI tool adoption figures, the image generation and video generation performance claims, and Zuckerberg’s description of the goal.
24Alphabet, fourth quarter 2025 earnings call, February 4, 2026, as transcribed by a third party.Platform claimSource of the nearly 70 million text assets figure.
25Ju & Aral, “Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance.” arXiv 2503.18238, revised February 2026.Working paper2,234 participants, 11,024 ads, about 5 million impressions on X. Source of the output, text quality, image quality and homogeneity findings.
26Noy & Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science 381, 2023, 187–192.Peer reviewed453 professionals on writing tasks. Source of the 40% time and 18% quality figures.
27Reisenbichler, Reutterer, Schweidel & Dan, “Frontiers: Supporting Content Marketing with Natural Language Generation.” Marketing Science 41(3), 2022, 441–452.Peer reviewedField experiments in two industries. Source of the ranking and cost findings and the quote on the human editor.
28Becker, Rush, Barnes & Rein, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” METR, arXiv 2507.09089, July 2025.Working paperRandomized trial with 16 developers and 246 tasks. Source of the 19% slowdown and the 20% perceived speedup.
29Xu and colleagues, “TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks.” arXiv 2412.14161, version of September 2025.PreprintSource of the 30% full-completion rate for the best agent tested.
30Yao, Shinn, Razavi & Narasimhan, “τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains.” arXiv 2406.12045, June 2024.PreprintSource of the under-50% success rate and the repeated-attempt reliability figure on retail tasks.
31METR, “Time Horizon 1.1” and “Clarifying limitations of time horizon,” January 2026, building on Kwa and colleagues, “Measuring AI Ability to Complete Long Tasks,” arXiv 2503.14499, 2025.Research organizationSource of the roughly five-hour time horizon for the leading model in January 2026, the doubling rate, and the note on reliability-critical tasks.
32Beck, Eckman, Kern & Kreuter, “Bias in the Loop: How Humans Evaluate AI-Generated Suggestions.” Harvard Data Science Review 8(2), 2026; arXiv 2509.08514.Peer reviewed2,784 annotators reviewing AI suggestions. Source of the 82% and 31% error-catching rates.
33Acemoglu & Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33(2), 2019, 3–30.Peer reviewedSource of the displacement and reinstatement framework.
34Autor, Chin, Salomons & Seegmiller, “New Frontiers: The Origins and Content of New Work, 1940–2018.” Quarterly Journal of Economics 139(3), 2024.Peer reviewedSource of the finding that most current US employment is in job specialties introduced after 1940.
35Dietvorst, Simmons & Massey, “Overcoming Algorithm Aversion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify Them.” Management Science 64(3), 2018, 1155–1170.Peer reviewedSource of the adoption and accuracy findings.
36Hoffman, Kahn & Li, “Discretion in Hiring.” Quarterly Journal of Economics 133(2), 2018, 765–800.Peer reviewedSource of the finding on managers who hired against test recommendations, across 15 firms.
37Google Ads announcements, 2025: the Keyword blog, January 23 and April 30; the AI Max for Search launch, May 6; Google Marketing Live, May 21; and the Google Ads Help summary of 2025 Performance Max updates.Platform documentationSource of the control timeline in 3.2.
38Brynjolfsson, Li & Raymond, “Generative AI at Work.” Quarterly Journal of Economics 140(2), 2025, 889–942.Peer reviewed5,172 customer support agents. Source of the productivity and tenure findings.
39Dell’Acqua and colleagues, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Harvard Business School Working Paper 24-013, 2023.Working paper758 consultants. Source of the inside- and outside-the-frontier results.
40Dratsch and colleagues, “Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance.” Radiology 307(4), 2023, e222176.Peer reviewed27 radiologists in three experience groups reading 50 mammograms with AI suggestions, 12 of the 40 test cases deliberately wrong. Source of the 79.7% to 19.8% and 82.3% to 45.5% accuracy figures.
41Dell’Acqua, Ayoubi, Lifshitz, Sadun, Mollick, Mollick, Han, Goldman, Nair, Taub & Lakhani, “The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork.” Organization Science, 2026, in press. DOI 10.1287/orsc.2025.20702.Peer reviewedPreregistered field experiment with 791 Procter & Gamble professionals. An earlier version is NBER Working Paper 33641, April 2025, reporting 776. Source of the individual-with-AI, functional-silo and top-tier solution findings; the last as summarized by co-author Ethan Mollick.