Google's Ads Decoded podcast published a 48-minute episode on November 18 in which experiments product lead Sagar Shah and search experiments specialist Al Nevarez explained how Google Ads splits traffic in its campaign tests, how it decides whether a result is statistically significant, and how a newer test of budget increases runs each arm on half of its budget. Host Ginny Marvin closed with a rule of thumb that Google's Ads & Commerce Blog repeated today: an 80% confidence interval for directional changes, and 95% for major structural ones.

In Short

Google used an episode of its Ads Decoded podcast to explain how the experiments built into Google Ads work, from the statistics underneath them to newer tests for budgets, bid targets and AI Max. It matters to anyone who raises a budget or switches on Google's automation, because these tests are Google's answer to whether a change actually worked before it is rolled out across an account. Google's guidance is to start with its pre-built templates, give a test four to six weeks and accept 80% confidence for small changes. Google's own help pages do not fully agree with the episode on test length or confidence levels, and support for shared budgets in experiments is still a pilot.

The episode

The video, titled "A/B testing in Google Ads: How to prove what works," appeared on the Google Ads YouTube channel on November 18, 2026. The channel shows 930,000 subscribers. By the evening of November 19 the video had 1,097 views, 23 likes and four comments, and it sat in the "Ads Decoded by Google Ads" podcast playlist, which lists 27 episodes.

The description opens with a question. "How do you know a campaign change works before you scale it? Experiments in Google Ads let you validate everything from new features like AI Max for Search to daily budget shifts and changes in your Smart Bidding strategies," according to the page. It promises a guide to "when to choose templated or custom test setups, and how to match confidence thresholds to your goals," and links to Google's help article on the statistical methodology behind experiments.

Google's Ads & Commerce Blog followed today with a short post titled "Build strategic A/B tests to grow your ad performance." It calls Shah and Nevarez "product leads," frames experimentation as "a strategic compass" in "an era of rapid AI innovation," and condenses the episode into three rules. The player embedded in that post gives the running time as 47:55.

The quotations in this article come from the episode's YouTube transcript, reproduced as written. That transcript does not label speakers, so attributions follow Marvin's hand-offs and the guests' references to each other. It also contains recognition errors that matter for meaning: "Jenny" for Ginny, "jackknife-free sampling" for jackknife resampling, "AMX research" for AI Max for Search, "roast target" for ROAS target, "controller treatment arm" for control or treatment arm, "multi-campus" for multi-campaign, "funnel URLs" for final URLs and "form pill" for form fill. Where a quoted passage contains one of these errors, the article says so.

The series began when Google launched Ads Decoded on August 29, 2025, with an episode on AI Max and Performance Max channel reporting. In October, product manager Tal Akabas used it to say that target or budget changes cause no learning period, and Google Search data science vice president Shivani Mohan told advertisers that AI search queries now run three times longer.

"Trust us" and "prove it"

Marvin framed the episode with a line from Nevarez. "And for me, the overarching theme of this conversation really comes down to something that Al said, to use experiments to move from trust us to prove it," she said. Experimentation, in her words, "isn't just a nice to have, it's really both a safety check and a strategic compass."

Nevarez gave the original version when asked why experiments are worth running. "At the end of the day, I guess it's, it basically comes down to going from trust us to prove it," he said. The platform provides "that scientific rigor in the experiment, um, so you don't have to bring the lab coat to get kind of the lab grade truth from your test."

Shah was more direct about the commercial subtext. "At the end of the day, we wanna make sure that advertisers have the tools they need in order to be able to sort of validate that the features that we build, um, and the sort of the platform is able to do a good job for them," he said.

The arrangement is worth stating plainly. The features under test are Google's, the auction they run in is Google's, and the arithmetic that decides whether they worked is Google's too.

The statistics under the hood

Every system for A/B testing has to answer the same question: whether the gap between two groups is larger than chance alone would produce. Marvin said Google does not "often talk a lot about the methodology" and described the model as "jackknife resampling to calculate confidence inter- intervals and statistical significance in our AB experiments."

The null hypothesis

Using conversions as the example metric, Nevarez said that "an AB experiment platform calculates the probability that the observed uplift in conversions would have occurred if there was actually no difference between the control and the treatment." Put more briefly: "It's the probability that the data matches the null hypothesis." Because advertisers hope a change causes a real increase, "we'd like that conditional probability to be very low, really low."

He described three steps: "Number one, calculate the variance of the data. Number two, calculate the standard error, and number three, calculate the confidence intervals." The jackknife is Google's method for the first step, he said. (The transcript renders the method here as "jackknife-free sampling," a recognition error.) "The jackknife is a versatile method that effectively detects outliers and helps reduce bias," Nevarez said. "It also includes bucketing of the traffic, um, to further increase the accuracy of those confidence intervals we calculate in step three."

Buckets and outliers

According to Nevarez, the split randomisation "together with the jackknife ensures that any given user will be consistently assigned to the same controller treatment arm, and they will also be consistently assigned to the same bucket during that trial." ("Controller treatment" is the transcript's rendering of "control or treatment.") The purpose is protection against unusual users. "Imagine if there's a user who converts many times on your campaign or he has an exceptionally high conversion value," he said. "Here, the jackknife process will protect us against that outlier, uh, thereby helping us avoid false positive results later on."

The resampling is what statisticians call leave-one-out. It means "recalculating the experiment results many, many times, each time leaving out a bucket or some of the users," Nevarez said. "So if the variance changes a lot, when I remove a couple people, that's an indication it's probably an outlier and the system will detect that."

Shah added the comparison with the usual alternative. Without an A/B test, advertisers typically run "a pre-post of some sort and see how things moved," which "gets you into challenges of seasonality or other reasons that may be sort of, you know, not controlled for." A test that misses significance, he said, is "just basically saying, 'Hey, you're seeing a difference between control and treatment, but we don't know that that difference is anything other than noise, right?'"

What the help page adds

The help article linked from the episode, titled "The statistical methodology behind experiments," supplies figures the podcast left out. It states that "Jackknife resampling is applied to bucketed data to calculate the sample variance of the percent change of a metric," and that "Twenty buckets are used in the control arm and twenty buckets are used in the treatment arm." It calls the jackknife "the standard at Google."

Two points in that documentation sit awkwardly next to the episode. First, the page says that "Two-tailed significance testing is then run using the 95% confidence interval," a single fixed level, while the episode and a separate help page describe a choice of levels with 80% as the default. Second, the page distinguishes two kinds of split. In a search-based split, users are placed in either arm at random each time a search occurs; in a cookie-based split, a user stays on one version however many times they search. Nevarez's description of users being "consistently assigned" fits only the second.

The page also says advertisers cannot rebuild the buckets or run the jackknife on aggregated experiment data themselves, and that Google has "no internal tools to do so on behalf of our clients." None of the three Google help pages cited in this article displayed a publication or update date as of November 19.

Templated or custom

When advertisers open the experiments page, Marvin said, "they are presented with options for pre-built goal-based experiments or the option to build custom experiments from scratch." Nevarez said custom experiments let an advertiser "adjust anything you like in that treatment arm," with a caveat: "if you adjust too many levers in this custom experiment, you may arrive at the end of your experiment and you're not really sure which lever caused the change."

Templates test "one thing at a time," and they work differently underneath. "Instead of creating, for example, a new campaign copy or a draft as we do with the custom experiments, we're actually splitting the traffic fifty fifty within the single campaign," Nevarez said. "This helps sometimes to minimize errors, and this approach may also lead to a reduced learning period because you're acting on that same campaign that you started with."

Shah separated "big decisions," such as changes to account structure or measurement setup, from "day-to-day optimizations, which is, 'Hey, how should I change my creative or what roast target should I set? Or what happens if I tweak my budget a little bit?'" ("Roast" is the transcript's rendering of ROAS.) The templates are recent, he said: "We've only had custom options for a long time, right?"

Sync and what the system cannot know

Custom experiments carry their own maintenance problem. Marvin said the sync feature "is on by default for custom search and display experiments" and is "supported by the APIs as well as in the UI." Shah said Google introduced it "four or five years ago now" because of "experiment arms that were just mismatched." In "a six to eight week experiment," he said, "sometimes longer you forget that this thing is running and so you, you change something and then that invariably ends up distorting the results of the experiment."

Sync copies changes made to the control campaign over to the treatment campaign, and it can only be switched on at the start. "We, we can only set it up at the very beginning of running an experiment because that's when both the, the arms are matched and we know what should be synced and what shouldn't be synced," Shah said. The constraint reflects a wider one. With a custom experiment, "our systems don't fundamentally have a deep understanding of the exact set of things that you're trying to test," he said. Google can see that a bidding strategy differs between arms, "But did you intend to actually test that or was that something that just happened because the arms are mismatched? It's not something we, we'll always know."

AI Max for Search and the multi-campaign pilot

Nevarez described the AI Max test as "an AB template under experiments campaign features," named "AI Max for search," and said it "has become the most popular way to test AI Max." He gave no figure for that popularity. The template exposes "the campaign level settings, which includes text customization and final URL expansion," along with brand controls. Final URL expansion lets Google send a click to a landing page other than the one the advertiser specified. Search term matching "is more of an ad group level thing," Nevarez said, edited in the original campaign rather than in the experiment.

The more consequential detail concerned which campaigns can enter a test. Nevarez described "the new, uh, very exciting campaign selection feature," which puts one campaign or several into a single experiment. "And we just launched the pilot this month to actually allow you to select a shared budget or a bidding portfolio as well," he said. Each selected campaign "will be split fifty fifty, just like as if it was one campaign," and results are assessed in aggregate. With more campaigns, "you'll have more statistical power and you can possibly get to a conclusion sooner, um, due to the higher volume." Marvin said shared budgets had "sort of been a obstacle to be able to use experiments."

That obstacle was documented. When Google announced on August 20 that advertisers could test budgets and ROI targets across several Search campaigns in one A/B test, PPC Land noted that AI Max experiments could not be created for campaigns on a shared budget or a portfolio bid strategy, among six blocking conditions. A portfolio bid strategy applies one automated bidding goal across several campaigns, which is why such campaigns are hard to split one at a time. The episode is the first public account of a pilot that relaxes the restriction. It does not say which advertisers or markets are included, and because no recording date is given, "this month" can only be placed on or before November 18.

Budget and bid target tests

Marvin set the new tests against Google's forecasting tool. "Performance Planner often offers simulated forecasts for changing budgets, smart bidding targets, and now in search, the new budget and bidding targets experiments will allow advertisers to AB test their budgets and, uh, target changes live," she said.

Nevarez tied the feature back to his opening line. "This really kind of touches on the prove it thing I said earlier on, right? Don't just trust, you know, um, that more budget may help. You can now demonstrate scientifically." Whether the setup screen offers a budget or a target test depends, he said, on "what bidding strategy your campaign is using and whether or not it's a budget constrained or not."

His budget example is the clearest description of the mechanics Google has given. "If your current budget was about $100 per day and you wanna test $150 per day, during the trial, the c, the control arm will get a $50 budget and the treatment will get a $75 budget," Nevarez said. If the advertiser applies the result, "the main campaign will move up to the $150 that you set at the onset of the campaign." The halving is deliberate: "we do that half splitting to kinda fairly compare during the trial itself." On PPC Land's arithmetic, the campaign's combined daily budget during that trial is $125, a quarter above the original $100 and a sixth below the $150 under test.

For a target change, the treatment arm simply takes the new value. "Let's say that it was originally $20 and you'd like to test $30. So during the trial, the treatment arm will simply be set to the $30," he said. Target CPA advertisers "typically look for conversion uplift," and Target ROAS advertisers "typically look at conversion value," though the key metrics are configurable and the report covers others regardless. Nevarez added: "we don't impose any spend thresholds or limitations or minimums on these experiments. You can kinda set the budget or the target to anything you like."

Test described in the episode Setting before the test Value tested Control arm during trial Treatment arm during trial Setting if applied
Budget increase $100 per day $150 per day $50 per day $75 per day $150 per day
Target CPA change $20 $30 Not stated $30 $30
Minimum spend None, according to Nevarez - - - -

Google's August 20 post said the multi-campaign budget and ROI target test would roll out in September. The episode gave no current rollout status.

Manual CPC against Smart Bidding

Marvin raised a frequent setup error: testing manual bidding against Smart Bidding with a poorly chosen starting target. Shah said the approach mirrors a first move to automated bidding. "What we typically recommend is that you, you look at your historical performance that your campaign on manual CPC is achieving," he said. "So if you're achieving a $15 CPA or 100% ROAS, then that's the target that I would recommend you set in your experiment."

His caveat concerned timing. "Make sure you're accounting for conversion delay, right?" Shah said. Recent days with conversions still to arrive should be left out, "because then you might end up setting a much higher CPA than you, than you actually need to." A target set too tightly creates a different problem: the treatment "takes longer to ramp up," and "it becomes difficult to parse out what the effect of the experiment actually was."

How long to run

Shah called four to six weeks "our standard recommendation" and also "a heuristic." The aim is enough conversions that "your confidence intervals can become narrow enough and you can clearly distinguish between control and treatment." His rough schedule was to "let it ramp up for about a week or so, then, like, measure results for, you know, four to five weeks," then end the experiment and wait for delayed conversions before comparing the arms. Counted that way, the full cycle runs past six weeks, though Shah gave no length for the waiting period.

Google's guidance on premature evaluation has long stressed excluding days with incomplete attribution, and conversion lag can run from seven to 21 days, a range PPC Land reported in March alongside an earlier Ads Decoded episode on Smart Bidding.

Ad variations, assets and Performance Max

Ad variations test one change across many campaigns. Shah said the tool is "intentionally, um, simple and straightforward," meant for swapping a headline, description or final URL "across, like, multiple campaigns or across your account." Testing combinations across "creatives or, like, targeting or your landing pages" calls for "a smaller custom experiment."

Marvin listed the asset tests: "asset uplift experiments, as well as experiments that allow you to test up to, I think, 20 groups of video assets," and "a PMAX, final URL expansion on experiment." Her figure of 20 video asset groups was hedged on air. Google's help page on the experiments page describes its separate video experiments as having two to four arms, and the episode did not reconcile the two.

For retailers comparing a Performance Max campaign built only on a product feed with one carrying added assets, Shah said a test is possible provided "everything else is the same across the board, right? Your budgets, your targets, all of those kinds of settings." The video test exists "because we know that videos are expensive to make," and because advertisers "have concerns or questions about engaged view conversions." He added an unprompted point: "YouTube is not the same as videos, right?" Performance Max spend on YouTube also covers "image-based formats on YouTube text-based formats as well," he said.

Google expanded asset A/B testing from retail campaigns to all Performance Max campaign types in January, as a beta that locks the tested asset group for the duration. Two months before that, in November 2025, Demand Gen gained asset uplift A/B experiments with 70%, 80% and 95% confidence thresholds.

Uplift experiments and "comparable campaigns"

Marvin turned to the "PMAX uplift experiment and the new demand gen uplift experiment," which "automatically identifies and assigns what we call comparable campaigns to both the control and treatment arms." This, she said, "sometimes confuses practitioners if they see that their existing search or display spend shifts, or they wonder why their PMAX traffic is getting throttled."

Shah said the uplift tests measure "the effect of adding a Performance Max or Demand Gen campaign to your account on top of your existing account structure." Comparable campaigns exist because a new Performance Max campaign can take traffic from existing Search campaigns. "So we look at shared variables, things like are the domains the same? Are you targeting the same conversion actions? Is the location targeting sort of similar and so on," he said. The control is the account as it stands; the treatment is the account plus the new campaign. "We don't wanna like artificially take credit for conversions that were just shifted from another campaign type," Shah said.

That is a claim about incrementality: the conversions that would not have happened without the new campaign, rather than those it merely absorbed. Google first offered Performance Max experiments in January 2023 and added a dedicated uplift test in October 2023. The episode gave no launch date for the Demand Gen version and did not say how many comparable campaigns a typical test pulls in.

When a test shows no improvement

Nevarez gave the episode's most structured answer when Marvin asked what to do after a flat result. He began with the report's secondary metrics, "as something may actually have gone well in an unexpected way," and then listed checks. In order, according to Nevarez:

  1. Confirm that conversion tracking works. "That's paramount to, uh, a lot of the, uh, primary metrics."
  2. Check that the experiment ran "for the recommended four to six weeks." With "a very low conversion count, it might need to run longer."
  3. For AI Max tests run with text customisation and final URL expansion switched off, "consider running the experiment again, but with those enabled."
  4. Review the budget. "If during the trial that treatment arm becomes budget constraints, you're gonna probably limit the effect that the treatment has in comparison to the control."
  5. Look at the search terms report "and consider if you need to add some negative keywords if the clicks that are coming through might not be relevant."
  6. If the test lacks conversions, add more campaigns to it through the multi-campaign option, for "more statistical power."

The third item is the only one that answers a weak automation result with more automation. It sits alongside wider evidence on AI Max: when Google scheduled the conversion of broad match campaigns to AI Max from September 1, it cited an average of 7% more conversions at a similar CPA or ROAS, from internal data that excluded retail, while an analysis of more than 250 retail campaigns by Smarter Ecommerce found AI Max conversions coming in at about 35% lower ROAS than other match types.

Measuring the right goal

Asked what setup matters most, Shah said the issue "is not experiment specific." When advertisers say a feature "didn't work for me," he said, "it's often because the goal that they are setting, right, or have given the platform or like have pointed the automation tool, um, isn't the actual like goal that they care about." His example came from lead generation: "Well, they have a shallow conversion, like a form pill on a website, but the thing they actually care about is a qualified lead, right?" ("Form pill" is a recognition error for form fill.) He called it "probably like the single most important thing in general."

When CPCs rise

Marvin raised a pattern she said "comes up a lot": advertisers who see cost per click rise in a test and conclude that it failed, even when the CPA or ROAS target was met. Shah conceded that a single-metric answer "ignores the lived reality of, you know, explaining to your, like, finance department, like, why things worked a certain way," then explained why CPCs can rise under Smart Bidding. "It might be that, turns out you may have to bid more to get users that are, like, more likely to convert, right?" he said, adding that bidding in "more expensive auctions" can lift CPCs while CPA stays "on target." Impression share on some keywords may also fall, "because, hey, it turns out we're not entering a bunch of auctions, um, that were not as efficient before."

Three rapid answers

Uneven spend. Asked why a 50/50 split can produce unequal spend, Nevarez said an A/B test is "fairly splitting the traffic and the functionality," and a better-performing treatment can win more auctions. "It will always confined to the budget that you set for the campaign," he said. "So it shouldn't be a lot more." Shah then offered "an extreme example": at a $10 CPA, a control arm with 10 conversions spends $100, and a treatment that finds 20 conversions at the same CPA spends $200, "even if I, like, am splitting traffic evenly." He did not say what budget that example assumed. "The goal is not to even spend out," Shah said.

Confidence levels. "Okay, we now offer 80%, 85%, and 95% confidence intervals," Marvin said, before asking when the lower level is enough. Nevarez answered: "Yeah, I think we, we feel that the 80% is typically good enough when, especially when combined with the advertiser's judgment, and the goal is, you know, to look for a directional improvement from what you're setting in the treatment." For features such as AI Max, he said, "these things are just gonna keep getting better over time." If an advertiser intends to apply the treatment to campaigns that were not part of the test, "maybe they wanna be a little stricter and set it to the 95 and just be really, really sure," he said. The level can be changed in reporting at any time.

What 80% means is worth spelling out. An 80% confidence interval is built so that, across repeated experiments, it misses the true effect one time in five; at 95%, one time in twenty. That is a standard property of confidence intervals, not a figure from the episode.

The post-test slump. Asked why performance can dip right after a winning treatment is applied, Nevarez said "these modern, uh, ad products need some time to learn," particularly when a custom experiment is applied back to the base campaign. "It's just a matter of being patient," he said, while the original campaign returns to "a mature level again, just like you had in the treatment arm." The episode did not say how long such a dip lasts. The answer also sits uneasily with Akabas's statement in October that target or budget changes on an established campaign trigger no learning period, though a custom experiment can apply a wider set of changes than a single target edit.

Where Google's own documents differ

The episode, the blog post and Google's help pages give different answers on basic questions.

Question Ads Decoded episode (Nov 18) Ads & Commerce Blog (Nov 19) "Monitor your experiments" help page "The statistical methodology behind experiments" help page "About the Experiments page" help page
Confidence levels 80%, 85% and 95% offered 80% for "directional tweaks," 95% for "major structural overhauls" "You can pick your own confidence intervals (80% is the default confidence interval)" "Two-tailed significance testing is then run using the 95% confidence interval" Not stated
How long to run Four to six weeks, plus about a week of ramp-up and time for delayed conversions "Four to six weeks to account for conversion lag" "We recommend allowing your experiment to run for 2 to 3 weeks to gather data" Not stated "At least 4 to 6 weeks" when results are undecided
Applying a winner Chosen at setup: automatic if the treatment wins, or manual Not stated Auto-apply "is enabled by default" for some experiment types Not stated Not stated

The 85% option Marvin mentioned does not appear on the "Monitor your experiments" page, which names 80% as the default and uses 95% as an example. The monitoring page's two-to-three-week recommendation falls well short of the four to six weeks that the episode and the blog post advise.

How the templated tests are set up

According to Nevarez, the AI Max for Search template follows five steps:

  1. Open the Experiments page, under Campaigns, and choose the AI Max for Search template from the campaign features templates.
  2. Select one campaign or several. Under the pilot described in the episode, a shared budget or a bidding portfolio can also be selected.
  3. Confirm the AI Max features to switch on in the treatment arm: text customisation and final URL expansion at campaign level, with brand controls available. Search term matching is set at ad group level and edited in the original campaign, not in the experiment.
  4. Review the default key metrics on which statistical significance will be assessed, and keep or change them.
  5. Choose the outcome at the end: apply automatically if the treatment arm wins, or apply manually.

Each selected campaign is then split 50/50 within itself rather than copied.

The budget increase test, according to Nevarez, works as follows:

  1. Select one or more campaigns. The screen shows each campaign's current daily budget.
  2. Enter the desired budget for the treatment arm in a text box.
  3. During the trial, the control arm runs on half of the original budget and the treatment arm on half of the new budget.
  4. If the advertiser applies the result, the campaign moves to the new budget.

Google's three rules

The blog post's three rules, as published:

  • "Start with templated setups: Pre-built templates isolate single variables, split traffic 50/50, and minimize setup errors for cleaner data."
  • "Focus on business outcomes: Measure qualified conversions rather than shallow proxies, and give tests four to six weeks to account for conversion lag."
  • "Match confidence to decision scope: An 80% confidence interval works well for directional tweaks. Reserve 95% certainty for major structural overhauls."

Marvin's spoken version of the first rule named "AMX research," the transcript's rendering of AI Max for Search. Nevarez's closing remarks summarised Google's pitch in one sentence: "You basically have the power of a data science team just a couple of clicks away."

Why it matters

Google has spent 2026 moving more decisions inside a Google Ads account into automated systems. Marvin said in August that targets become the efficiency lever in budget-capped campaigns after Google's August 17 bidding change, and Google scheduled the conversion of broad match campaigns to AI Max from September 1. Each change raises the same question for an advertiser: did it work, and compared with what? The experiments tools are Google's answer, and this episode is the most detailed public account so far of how they reach a verdict.

The halved budgets, the in-campaign split, the comparable campaigns and the multi-campaign pilot all change how a result should be read.

The limits are equally clear. Advertisers cannot reproduce Google's calculation from their own data, according to Google's help page. The default confidence level of 80% accepts a higher chance of a false winner than the 95% level Google's methodology page still describes, and the monitoring page lists automatic application of favourable results as on by default for some experiment types. None of this makes the tests unreliable. It does mean that "prove it" is proof as Google computes it.

Google offers other routes to the same question. It cut the minimum budget for its incrementality studies to $5,000 in 2025, using Bayesian methods, a different statistical approach from the jackknife and confidence intervals used in campaign experiments. In August, Google said Performance Planner's suggested changes could be applied to live campaigns in one click, which lets an advertiser act on a simulation without running a test at all. This episode placed the live experiment above the forecast.

Several things went unsaid. The episode gives no recording date, no list of markets or advertisers in the shared budget and portfolio pilot, no results from any experiment, no launch date for the Demand Gen uplift test and no confirmation of the 20 video asset groups Marvin mentioned. Nor does it reconcile its four-to-six-week advice and its 85% option with the help pages that say otherwise.

Key facts

  • Episode: "A/B testing in Google Ads: How to prove what works," Ads Decoded by Google Ads, published on the Google Ads YouTube channel on November 18, 2026, https://www.youtube.com/watch?v=1mTKp8qdEpA
  • Running time: 47:55, according to the player in Google's blog post
  • Host: Ginny Marvin, Ads Product Liaison, Google
  • Guests: Sagar Shah, experiments product lead; Al Nevarez, search experiments specialist (both called "product leads" in the blog post and "product experts" on YouTube)
  • Companion post: "Build strategic A/B tests to grow your ad performance," Google Ads & Commerce Blog, November 19, 2026
  • Methodology documentation: "The statistical methodology behind experiments," Google Ads Help, answer 9232676 (no date shown): jackknife resampling on bucketed data, 20 buckets per arm, two-tailed testing at 95%
  • Other help pages cited: "Monitor your experiments" (answer 6318747: 80% default confidence, 2 to 3 weeks recommended, auto-apply on by default for some types); "About the 'Experiments' page (formerly drafts and experiments)" (answer 10682377: at least 4 to 6 weeks when undecided)
  • Location in Google Ads: Experiments tab under Campaigns
  • Experiment types discussed: custom experiments (Search and Display, sync on by default); AI Max for Search template; budget and bidding target experiments (Search); multi-campaign selection, with a pilot for shared budgets and bidding portfolios launched "this month"; ad variations; asset uplift experiments; video asset experiments; Performance Max final URL expansion experiment; Performance Max uplift experiment; Demand Gen uplift experiment
  • Confidence levels named in the episode: 80%, 85%, 95%
  • Budget test example: $100 per day tested against $150 per day runs as $50 against $75 during the trial
  • Target CPA example: $20 tested against $30
  • Spend minimums for budget and target tests: none, according to Nevarez

Timeline

Summary

Who: Ginny Marvin, Google's Ads Product Liaison, with Sagar Shah, experiments product lead, and Al Nevarez, a search experiments specialist, on Google's Ads Decoded podcast. The subject affects Google Ads advertisers who test bidding, budgets, AI Max, assets, Performance Max and Demand Gen.

What: A detailed account of how Google Ads experiments work: jackknife resampling on bucketed traffic to estimate variance and confidence intervals, templates that split a single campaign 50/50, budget tests that run each arm on half its budget, a pilot that admits shared budgets and bidding portfolios to multi-campaign tests, comparable campaigns in uplift tests, and advice that 80% confidence suits directional changes while 95% suits major ones. Google's help pages differ from the episode on test length, available confidence levels and the significance level used.

When: The episode was published on November 18, 2026, and Google's Ads & Commerce Blog summarised it on November 19, 2026. The shared budget and portfolio pilot launched "this month," according to Nevarez.

Where: On the Google Ads YouTube channel and the Google Ads & Commerce Blog, with the features located in the Experiments tab of Google Ads accounts.

Why: Google has moved more account decisions into automated bidding and AI Max during 2026, and it presents its experiments as the way advertisers can check those changes before scaling them. The episode explains how Google reaches its verdicts, though advertisers cannot reproduce the calculation independently.