Meta的Andromeda算法：来自53,000条广告的启示

最引人注目的事情是，这个时期投入的广告制作量巨大。

The biggest things that popped out immediately is the the just volume of ad production that went into this time of year.

从去年的约13,000条广告增加到今年的超过50,000条。

So from approximately 13,000 ads that were launched last year for this dataset to over 50,000 this year.

接下来我将向你们展示一些具体的例子，说明广告投放量是如何分布的，以及这种量级策略为何奏效，为什么‘快速试错’和‘创意速度’在未来至关重要。

So I'm gonna show you we're gonna show you some some specific examples of how that breaks down and and how the volume strategy paid off and why this idea of failing fast and creative velocity is is paramount moving forward.

但在进入创意部分之前，我想先简单谈谈出价策略，以及它在我们这个周末的数据集中是如何体现的。

But before we go into the creative piece, I wanna talk just a little bit about the ideas of of bidding and how that actually played out in in our data set for for the weekend.

本集《电商实战手册》由Tax Cloud赞助播出。

This episode of the ecommerce playbook is brought to you by Tax Cloud.

现在是电商的旺季。

It's peak season for ecommerce.

你的团队正在快速运转，订单纷至沓来，每一天都像一场冲刺终点的赛跑。

Your team's moving fast, orders are flying in, and every day feels like a sprint to the finish line.

但随着销售额攀升，你的销售税义务也可能变得越来越复杂。

But as sales climb, your sales tax obligations are probably getting more complicated too.

这就是Tax Cloud的用武之地。

That's where Tax Cloud comes in.

他们是Shopify品牌信赖的销售税申报合作伙伴，直接处理与各州的注册、申报和合规事宜，让您无需操心。

They're the trusted sales tax filing partner for Shopify brands handling registrations, filings, and compliance directly with the state so you don't have to.

在这个季节，您要处理的事情已经够多了，销售税不该是其中之一。

With everything else you're juggling this season, sales tax shouldn't be one of them.

Tax Cloud为您保驾护航。

Tax Cloud has you covered.

请访问 taxcloud.com/thread 开始使用。

Start at taxcloud.com/thread.

大家好。

Hey, folks.

欢迎收听《电商实战手册》播客。

Welcome to the Ecommerce Playbook podcast.

我是主持人理查德·加芬，就职于Common Thread Collective，担任数字产品战略总监。

I'm your host, Richard Gaffin, director of digital product strategy here at Common Thread Collective.

我身边的是我们的付费媒体副总裁托尼·切普切普，他也是我主持DDC热线的搭档。

And I'm joined by our VP of paid media, Tony the Chopper Chop, also my cohost on the DDC hotline.

但他今天加入我们是为了聊聊安德罗美达。

But he's joining us today to talk a little Andromeda.

托尼，最近怎么样，老兄？

Tony, what's going on, man?

理查德，一切都好。

Richard, all is well.

你呢？你自己怎么样？

How about how about yourself?

嗯。

Yep.

如果你一切都好，托尼，那我也一切都好。

If all is well with you, Tony, all is well with me.

我觉得你这件衬衫真不错。

I'm gonna say I really like your shirt.

这种中式立领设计应该回归潮流。

Like that the Mandarin collar needs to come back.

这看起来很棒，我很受启发。

It's a good look, and I'm inspired.

但没错。

But Yeah.

酷。

Cool.

让我们直接切入正题吧。

Let's let's let's jump straight into here.

所以，我们今天讨论的核心内容是Andromeda。

So, like, the purpose behind kinda what we're talking about today, we're talking about Andromeda.

特别是，我们讨论的是我们在BFCM期间收集的数据，如何体现了当前市场环境与当时相比的巨大变化。

And particularly, we're talking about the ways in which the data that we gathered over BFCM encapsulates the vast difference in the current in the landscape between then and now.

当我们刚刚查看Andromeda的时间线时，发现它是在BFCM之后推出的，这意味着我们可以清晰地对比去年和今年黑色星期五的情况，看看算法如何改变了我们对平台的策略。

So as we were we were just kind of looking into the timeline here on Andromeda, and it rolled out right after BFCM, which means we kind of have a clean comparison between what Black Friday looked like last year and what it looked like this year in terms of how the algorithm has changed our approach to the platform.

所以，我们这边的托尼一直在收集线索。

So what we got Tony here, he's been gathering he's been gathering the clues.

他一直在分析这些线索。

He's been analyzing them.

今天我们主要关注的是，Andromeda算法如何奖励数量、格式和策略，特别是它对创意产生的影响。

And what we're kind of looking at today is is how the Andromeda algorithm rewards volume format and strategy, particularly how it affects creative.

所以，托尼，你来深入讲讲吧？

So, Tony, why don't you dive into it?

给我们简单介绍一下这个研究项目的背景，以及我们希望通过它获得什么。

Give us, like, a a little bit of background behind this sort of research project and and kinda what we were hoping to get out of it.

是的。

Yeah.

是的。

Yeah.

正如你所说，这很有趣。

It was interesting to to your point.

在我们开始讨论之前，我记得Andromeda的对话是从十二月开始的。

Just before we got to talking, we I I remember, like, the Andromeda conversation starting in December.

是的。

Yeah.

我们去查了一下，发现他们实际上是在12月2日，也就是网络星期一那天，Meta在他们的工程博客上发布了关于Andromeda的内容。

And we went and looked it up, and it was actually December 2 that they they posted about it Meta posted about it on their engineering blog, which happened to be Cyber Monday.

所以，这个想法正是源于那次发布。我们今年已经多次谈到Andromeda，它本质上是一种与传统媒体平台截然不同的机制，允许我们采用不同的出价策略，围绕着获取客户的成本、你的出价和预算之间的关系等等。

So it so the idea for this sort of kind of came out of that where, you know, we've talked about Andromeda a lot this year and and how, really, it's a a fundamentally different mechanism underneath the hood from traditional media platform that allows us to bid, you know, using different bidding methodologies and is around this relationship between the cost to acquire a customer and how much you bid and how much your budget is, etcetera, etcetera.

我们用来描述Andromeda的传统说法是，它是一个由AI驱动的内容分发系统。

Really traditional stuff to the the language that we've heard around Andromeda that we use is it's it's an AI driven content delivery machine.

所以，它的核心理念是将最合适的广告内容与最有可能对这些内容做出回应的用户精准匹配。

So so it's all about this idea of pairing the exact right piece of content with the user that's most likely to respond to that content.

今年我们一直有一些理论和假设，关于这会对广告的形式和数量产生怎样的影响。

And and we've had some some some theories and some theses throughout the year about what that means for ads, sort of the type of ad and the amount of ads that you need.

我们想利用这个机会，结合BF和BFCM数据集，对比一下年度变化，看看能得出什么结论。

And we wanted to use this opportunity to compare, to really look at some year over year things on on the BF BFCM dataset to to see what we could tease out.

嗯。

So Mhmm.

这里的分析涵盖了超过53,000条广告，涉及约147个客户，这些客户在黑色星期五期间产生了1.27亿美元的收入和1450万美元的媒体投放投入。

So the setup here is the analysis is on just over 53,000 ads across to approximately a 147 clients, which represent a $127,000,000 in revenue and 14 and a half million dollars in media investment across Cyber five.

从感恩节开始，一直到网络星期一。

So starting on on Thanksgiving through Cyber Monday.

好的。

Alright.

这就是背景。

That's the setup.

明白。

Okay.

不错。

Cool.

那我们就开始吧，试着弄清楚这里究竟发生了哪些变化。

Well, then let's let's let's roll into it and and kinda try to understand exactly what changes have occurred here.

嗯。

Yeah.

嗯。

Yeah.

所以有几件事特别突出。

So a couple things really jumped out.

最明显的一点是，这个时期投入的广告制作量巨大。

So the one of the biggest things that popped out immediately is the the just volume of ad production that went into this time of year.

从去年的约13,000条广告增加到今年的50,000多条。

So from approximately 13,000 ads that were launched last year for this dataset to over 50,000 this year.

接下来我会向你们展示一些具体例子，说明这种细分情况，以及为什么这种快速试错和创意速度的策略至关重要。

So I'm gonna show you we're gonna show you some some specific examples of how that breaks down and and how the volume strategy paid off and why this idea of failing fast and creative velocity is is paramount moving forward.

但在进入创意部分之前，我想先简单谈谈出价策略，以及它在本周末我们的数据集中是如何体现的。

But before we go into the creative piece, I wanna talk just a little bit about the ideas of of bidding and how that actually played out in in our dataset for for the weekend.

所以，K。

So K.

有几件事先做个铺垫。

A couple things just just to set the table.

我们分析了在Meta平台上可用的不同出价策略下，1450万美元的媒体投放支出。

We analyzed that 14 and a half million dollar media spend across the different bid strategies that that are available for us to use in Meta.

要获得这个数据集非常困难。

This is a remarkably difficult dataset to get to.

原因在于，Meta并不容易让我们提取这些信息，而且出价设置可以在广告系列或广告组层级进行，这使得数据获取更加复杂。

And the the reason why is because, well, Meta doesn't make it easy to extract this information, and the the the bidding can be set either at the campaign or at the ad set level, which which makes it even even funkier to get to.

无论如何，我们与数据科学团队合作，成功提取了所有数据。

Either way, we were able to work with our data science team and and pop this all out.

这项分析中有一些显著的发现。

And a couple of things really stood out from this analysis.

第一，我们有三个主要的出价目标，占据了绝大部分支出。

Number one, we had three main bidding objectives that represented the vast majority of the spend.

分别是最高成本、最低成本、最高投放量、最低成本和成本上限。

So highest cost, lowest cost, highest volume, minerals, and cost cap.

这三个目标占了总投资的99%。

Those three accounted for 99% of the investment.

最突出的是，最低成本策略无疑是推动流量的主要引擎。

The thing that that stood out is lowest cost was for sure for sure the volume engine.

在该时期内，它占据了最大的支出比例。

It represented the largest portion of the spend during that period.

然而，真正有趣的地方在于平均订单价值（AOV）方面。

However, where it really gets interesting is on the AOV side.

因此，最低成本与MINROAS或成本上限之间的平均订单价值差异接近三倍。

So the difference between the AOV on lowest cost versus MINROAS or cost cap is nearly three x.

最低成本策略的平均订单价值约为15.50美元。

The average AOV on the lowest cost is approximately $15.50 dollars.

矿物策略和成本上限的平均订单价值约为150美元。

The average AOV on minerals and cost cap was approximately a $150.

因此，当我们思考如何应用出价原则及其与安德罗美达的关系时，这是一个核心概念。

And so this is such a this is such a core idea when we when we think about how we apply the bidding principles and and the relationship to to Andromeda.

在这种情况下，出价不仅仅改变了你的Meta广告活动的ROAS或CPA。

In this case, bidding doesn't just change the the ROAS or the CPA from from the outcome of your meta ads campaign.

它从根本上改变了你所获取的客户类型。

It fundamentally changes the type of customer that you acquire.

好的。

Okay.

嗯。

Yeah.

不。

No.

我正在查看数据。

I'm I'm just kinda looking at the data right now.

这太有趣了。

It's fascinating.

所以，我们在这里看到的是，在MINROAS和成本上限这两种出价策略下，CPA至少是最高流量活动的两倍，AOV也大约是三倍。

So what we're looking at here is on MINROS and cost cap, the both of those bid strategies had at least double the CPA of the highest volume campaign, and then also had triple the AOV, roughly.

所以，我的意思是，这表明作为成本上限的出价策略能够找到所谓的优质客户。

So, I mean, I guess what that's pointing to is the Mineral as a cost cap bidding strategies are able to find, quote, unquote, better customers.

它们面向的是那些愿意花更多钱的人。

They're put in front of folks who wanna spend more.

但你觉得为什么会发生这种情况呢？

But what do you think is the why is that happening?

也就是说，是什么算法机制让这种出价方式能够找到更好的客户？

Like, what's the what's the sort of algorithmic trick that causes that bid style to find a better customer?

我认为，我们对安德罗米达的信念在于，出价中所整合的信号包括视频观看时长以及个人行为，比如谁更常加购、谁更常结账，而这些都属于基于价值的出价策略。

Well, I think the part of what we believe to be true about Andromeda is the the signals that are being incorporated in the the bidding are things like watch time on videos and behavior for the individual around, like, who adds cart more often, who checks out more often, and both of these value based bidding strategies.

我们也会把‘最高价值’策略纳入其中，因为我觉得这个策略更加有趣。

And and we'll throw we'll throw highest value in there as well because I think this one is is even more interesting.

所有这些基于价值的出价策略都体现了Meta真正识别出高价值客户的能力。

All of these value based bidding strategies represent Meta's ability to actually tease out higher value customers.

我认为真正有趣的地方在于，我不认为存在所谓对或错的出价策略。

I think where where it becomes really interesting is I don't I don't think there's a I don't believe that there's a right or wrong bidding strategy.

嗯。

Mhmm.

但很明显，它们是不同的工具。

But it is very clear that they are different tools.

所以，例如，如果你的品牌产品之间的平均订单价值差异很大，比如有些产品卖20美元，有些则卖200美元，你就必须非常谨慎地设计你的广告账户结构，特别是针对出价策略。

And so so, for example, like, if you're a brand that has a really big difference in the AOV of your products, let's say you sell some things for $20 and you have other things for, you know, $200, you you have to be really conscientious about how you structure your advertising account specifically to bidding.

因为如果你只是把所有产品都放进一个以最低成本为目标的广告系列中，使用最低成本出价，你肯定会推动最低价值的销售。

Because if you just lump all of if you just lump that all of your products into one campaign around lowest cost, using lowest cost, you are for sure going to drive the lowest value sales.

嗯。

Mhmm.

另一方面，

On the flip side of the

如果你的产品数量较少，平均订单价值和利润率的范围也比较集中，那么最低成本、最高销量的出价方式确实能带来大量投放效果，正如黑色星期五和网络星期一期间所展现的那样——实际上大部分投资都通过这种最低成本、最高销量的设置完成。

coin, if you have a smaller group of products and sort of tighter ranges for your AOVs and even for your margin, for sure, lowest cost and highest volume is a is a bidding methodology that can get a bunch of delivery, as illustrated by what happened in BFCM where actually the majority of the investment went through went through this lowest cost, highest volume set.

所以，我认为这并不是对与错的问题。

So I don't think it's a it's a it's not a question of right or wrong.

这是一个问题，关键在于从根本上理解这些工具的用途，其实并不复杂，因为答案就在名字里。

A question it's a matter of fundamentally understanding what these tools are for, and it's not that complicated because it's in it's in the name.

是的。

Yeah.

当然。

Sure.

我认为特别有趣的是，通过这个庞大的数据集——超过1450万美元的广告支出——验证了名字所体现的内容，本质上就是你所能获得的结果。

And it it's I think it's just really cool to see it born out over over this significant dataset, over 14 and a half million dollars of media spend that that what what is in the name is is essentially what what you're gonna get.

我有点好奇想进一步探讨的是关于最高价值的出价策略，从表格中你可以看到，这种策略的平均订单价值最高，接近200美元，但转化量却是最低的。

The the thing that the thing that I'm sort of, like, curious to tease out and understand a little bit more is bidding around highest value, which you'll notice from the table had had the highest AOV, almost a $200 AOV, but the lowest measure for bras.

也就是最高的每次转化成本。

So, like, in in the highest CPA.

是的。

So Yeah.

这看起来像是出价策略中仍处于待完善阶段的部分。

This this feels like the part of the bidding stack that's, like, still a work in development potentially.

是的。

Yeah.

但如果你有一个非常高客单价的产品，那么这种最高价值出价策略似乎是一个值得探索的方向，因为你将找到那些真正认同这种产品类型的客户。

But if you have like a really high if you have like a super high AOV product, this feels like a a good area of exploration into this highest value bidding because you're you're gonna go find those those customers that that resonate with that that type of product.

问题是这样的。

Here's the thing.

每花一分钟担心销售税，就少了一分钟可以用来关注你的客户、产品、品牌和业务增长。

Every minute spent worrying about sales tax is a minute that could have been spent on your customers, your product, your brand, and growing your business.

这就是为什么这么多Shopify品牌选择与Tax Cloud合作。

That's why so many Shopify brands work with Tax Cloud.

他们通过直接处理各州的注册、申报和合规事务，简化了销售税的复杂性。

They take the complexity out of sales tax by handling registrations, filings, and compliance directly with the states.

无需复杂设置，无需等待数日才能获得支持，只需一个值得信赖的合作伙伴，他们能准确处理并为你腾出时间，专注于真正重要的事情。

No complex setup, no waiting days for support, just a trusted partner who gets it right and gives you time back to focus on what really matters.

把时间还给自己，专心经营你的业务。

Get back to running your business.

访问 taxcloud.com/threat。

Visit taxcloud.com/threat.

这是 taxcloud.com/threat。

That's taxcloud.com/threat.

嗯。

Yeah.

这很有趣。

That's interesting.

嗯。

Yeah.

所以，我的意思是，正如你之前提到的，Monroe广告活动会找到一个最低的ROAS。

So so, I mean, I guess, like, to to your point earlier, you know, a Monroe campaign is going to find a minimum ROAS.

因此，在最基本的层面上，这确实是最高效的支出方式。

And so to some in some basic level, so it makes sense that that would be the most efficient spend.

对吧？

Right?

结果是2.6。

Which has turned out to be 2.6.

呃，实际上，抱歉。

Well, actually, sorry.

它并没有成为最高效的支出方式。

It didn't turn out to be the most efficient spend.

它的ROAS更高，但成本上限实际上也更高。

It had a higher ROAS, but cost cap was actually even higher.

成本上限背后的思路是，你实际上是在说：CPA或最高CPA为这个数值，两侧会有一些浮动空间。

The idea behind the cost cap is you're you're essentially saying, hey, CPA or a maximum CPA of this, and there's a little bit of wiggle room on either side.

但你觉得，为什么它能拉高AOV，并且比其他出价方式表现得更高效呢？

But is there what's the reason do you think that that was able to draw out a higher AOV and actually perform more efficiently than the rest of the bid styles?

我认为，从定义上讲，这是最高效的出价方式，因为它约束最多。

Well, I think it's it's, by definition, the most efficient bidding, so it has the most constraint.

所以，我不惊讶成本上限对我们来说代表了最高的数值，因为成本上限和最低ROAS出价之间的区别在于：成本上限会严格限制在某个CPA，不会超过这个值；而如果你看一个最低ROAS活动的结果散点图，有些CPA会更高，有些会更低，最终会平衡到那个ROAS水平。

So I'm I'm not surprised that cost cap has represents the largest the the highest for us because the difference between the cost cap and mineralized bidding is cost cap's gonna sort of cap at a specific CPA and not go for anything above that, where MinRoAS if you looked at a scatter plot of the outcomes of a of a minimum ROAS campaign, some of the CPAs are gonna be higher, some of the CPAs are gonna be lower, and it's gonna balance out to that ROAS.

但另一方面，这两者之间确实存在差异，虽然差异并不大。

So but but the flip side of that coin is and the the difference isn't isn't huge, but there is a difference.

通过MinRoAS投放的媒体预算更多。

There's more media spend delivered through MINROAS.

你打算怎么称呼它？

What what do wanna call it?

大概多10%吧。

Maybe 10% more.

这额外的10%媒体预算通过MinRoAS投放，实际上就是那些高于成本上限阈值的CPA。

So that that 10% additional media spend that's going through MINROAS is effectively the CPA the CPAs that were above the cost cap threshold.

对。

Right.

有意思。

Interesting.

这就是为什么你是专家，托尼，而我不是。

Well, this is this is why you're the expert, Tony, and I'm not.

但让我们继续进入下一个部分，谈谈

But let's let's roll on to our next kind of section here talking about

所以

So

增长、广告量和扩展。

growth, ad volume, scaling.

但没错，你来说吧。

But, yeah, jump into it.

嗯。

Yeah.

所以当我们观察每年支出增长与每年广告量之间的不同关系时，我认为安德罗美达效应变得无可否认。

So the this is when we look at these the different different sort of relationships between the spend growth year over year and the ad volume year over year, and this is where I believe the Andromeda effect becomes undeniable.

首先，看看广告量。

So first things first, look at look at the volume.

创意制作的数量同比激增。

So creative production literally exploded year over year.

我们将静态广告数量增加了158%，视频广告数量增加了141%。

So we we increased static ad volume by a 158% and video ad volume by a hundred forty hundred forty one percent.

两种格式的增长率都非常相似，约为44%的年增长率。

So both formats grew at a very similar rate, approximately 44% year over year.

明白吗？

K?

从2024年推出的约15,000条广告增加到2025年的45,000条。

From about 15,000 total ads that were launched in 2024 to 45,000 in 2025.

是的。

Yeah.

两种广告类型的数量都增长了，我们生产的量都增长了三倍。

So both ad types grew, the amount that we produced, and they both grew, like, three x.

我认为这里真正有趣的是，随着静态广告的支出增加，我们观察到了回报递减。

What I think is really, really interesting here is as as spend scaled through the static ads, we observed a a diminishing return.

广告回报率（ROAS）实际上同比下降了8%。

The ROAS actually dropped 8% year over year.

视频广告表现出略有不同的行为，它们吸收了近100万美元的额外支出，而视频广告的ROAS不仅保持稳定，甚至略有上升。

Video ads showed a slightly different behavior where they absorbed nearly $1,000,000 in extra spend, and the video ROAS actually held and even slightly increased year over year.

因此，这证实了我们的观点：在Entromeda时代，视频广告能产生更密集的数据信号，如观看时长和互动率，从而在不破坏效率的情况下吸收规模增长。

And so this this is a confirmation of our thesis that in the Entromeda era, video create creates denser data signals, like watch time and engagement, and allowing it to absorb scale without without breaking efficiency.

因此，我今年初就有一个假设。

And so I had a I had a thesis coming into this year.

让我先退一步说。

Well, let me back up a step.

每年，在黑色星期五和网络星期一期间，静态广告传统上都占据了大部分投放规模和媒体预算。

So every year, we for BFCM, static ads have traditionally represent represented the majority of the scale, the majority of where the media investment goes.

我的理论是，今年这种关系可能会发生变化。

And my theory was that maybe this year that relationship would shift.

视频广告的占比可能会超过静态广告，创下历年新高。

It would be more more video than static than it has been in in years past.

但这一预测在本数据集中并未得到验证。

So that actually didn't bear out in this dataset.

但实际验证的是，视频广告能够更高效地吸收更多的支出。

But what did bear out is that video was able to absorb more of that spend more efficiently.

所以下一年。

So next year.

是的。

Yeah.

是的。

Yeah.

所以跟我聊聊这个观点吧，我只是想让你展开一下：视频之所以能吸收更多支出，是因为数据点更多。

So talk to me a little about I just wanted you to tease out this idea that, like, the video video was able to absorb more spend because of the greater number of data points.

这是不是因为算法拥有了更多信息，从而能更精准地投放给目标人群，进而提升了效率？

So is that a question of the algorithm having more information with which to serve it to the right people or something like that, and that creates additional efficiency?

还是说那里发生了别的什么情况？

Or what's kinda what's going on there?

是的。

Yeah.

没错。

Exactly.

完全正确。

That's exactly right.

所以是的。

So yeah.

所以，我们以观看时长为例吧。

So the well, let's use watch time as a for example.

好吗？

K?

嗯。

Mhmm.

视频会向安德罗米达引擎发送一个信号，比如你被推送了视频。

So video is going to send a signal back to to the Andromeda engine around let's say you get you get served videos.

对于芬德吉他视频，理查德，你从头到尾看完了。

And for Fender guitar videos, Richard, you watch that whole thing end to end.

当然。

Absolutely.

对吧？

Right?

没错。

That's right.

你从来不会点开离开。

You never click off it.

但对于吉布森，

But for Gibson,

吉他，我不该说这个

guitar I shouldn't I shouldn't be

这么说。

saying this.

吉布森吉他很棒。

Gibson guitars are great.

它们都很棒。

They're all great.

是的。

Yeah.

但让我们为某个东西用一个不同的类别。

But let's use a different category for for whatever.

水泡。

Water bobbles.

你只是滑动，直接划过去。

You just scroll you just scroll right past.

好吗？

Okay?

是的。

Yeah.

根本不关心。

Couldn't care less.

所以一个正在发送非常强烈、强烈的信号，而另一个则肯定在发送较弱的信号。

So one is sending a really strong, intense signal, and and the other is sending a lesser intense signal for sure.

嗯。

Mhmm.

明白了。

Gotcha.

好的。

Okay.

不错。

Cool.

这个幻灯片上还有其他你想补充的吗？

Any anything else on this slide you wanna add?

嗯，我认为最后一点是缩放关系。

Well, I think the the last thing is the the scaling relationship.

对。

So Right.

这里明显存在体积与规模之间的关系。

This there's a obvious relationship between volume and scale.

因此，生成更多广告的账户通常管理着更高的支出水平。

So accounts that produce more ads generally manage higher spend levels.

但在这个特定的数据集中，图表中并没有低体积、高支出的象限。

But there is in this particular dataset, there was no low volume, high spend quadrant in the chart.

明白吗？

Okay?

所以，零零情况是指我们有一些账户支出很高，但广告数量却很少。

So zero zero situations where we had an account that spent a ton but only had a few ads.

是的。

Yep.

我认为这反映了过去几年的情况，那时我们或许可以依靠一个高规模的赢家，让我们能够真正有力地推动不仅在黑色星期五和网络星期一，而且贯穿全年，但今年我们的数据集中并没有这种情况。

And I think this represents in in years past where we would maybe get away with having one one winner that's that scale and allowed us to to really meaningfully push into not only BFCM, but but throughout a year, and that didn't exist this year for our dataset.

是的。

Yeah.

是的。

Yeah.

对。

Right.

所以，过去可能存在一些情况，你有几个表现优异的广告，能够吸引大量投放预算之类的。

So so the idea being that, like, in in the past, there's potentially scenarios where you had a few, like, winning ads that were able to maybe transcend, pull a bunch of spend in, whatever.

此时，广告数量与投放预算之间的关系已经非常明确。

At this point, the relationship between volume and spend is is sort of undeniably clear.

就像广告越多，预算越高，这就是全部了，没什么复杂的。

Like, it's like more ads means more spend, and there's that's sort of the long and the short of it, and there's not a lot of nuance to it.

好的。

Okay.

不错。

Cool.

我们继续吧。

Let's let's let's roll on here.

所以我认为，这可能是从中浮现出来的最令人兴奋的一些内容。

So I got I think this is where this is probably some of the most exciting stuff that that popped out of it.

这个想法有两个部分。

So two parts to this idea.

进一步深入探讨一下量的概念。

So just more teasing into the volume idea.

我们推出了39,000条广告，但都没能花掉15美元。

We launched 39,000 ads that failed to spend $15.

经典。

Classic.

大约有1,100条广告带来了56%的收入。

About 1,100 that drove 56% of the revenue.

因此，这种必须愿意快速失败——绝大多数时候都是失败的——正是我们所处的世界。

And so this idea of having to be willing to fail fast 98 most most of the time is the this is the world that that we live in.

是的。

Yeah.

但如果我们需要成千上万次尝试，没错。

But if we need thousands of shots on goal Yeah.

这是一个我们一直讨论的想法，我认为让很多人感到非常不满意。

Which is a which is an idea that we've been talking about forever, and I think feels really dissatisfying to to a lot of people.

我们应该瞄准哪里呢？

Where where should we where should we aim?

什么是，没错。

What is the Yeah.

什么能给我们最大的概率，让某样东西进入赢家类别？

What gives us the best probability of making it to make having something make it to the winner category?

嗯。

Mhmm.

用户生成内容广告的全垒打率，绝对是实现规模的最佳选择。

The home run rate for UGC ads is absolutely your best bet for scale.

有意思。

Interesting.

一个

An

一个个体的用户生成内容广告，根据这个数据集，成为高消费赢家的可能性几乎是标准广告的两倍。

an individual so an individual UGC ad, based on this dataset was nearly twice as likely to become a high spending winner than a standard.

有趣。

Interesting.

是的。

Yeah.

如果你想找到下一个能够处理一万美元或更多广告预算的概念，你绝对必须测试用户生成内容视频广告。

If you wanna find the next concept that can handle 10,000 more 10,000 or more dollars in spend, you absolutely have to be testing UGC video ads.

所以这并不是我想在这里强调的一个重要区别。

So this isn't a this is an important sort of I wanna draw an important distinction here.

这并不意味着你可以免除创意制作的必要性，Andromeda系统仍将继续需要它。

This is not a a hall pass to not go into the creative production necessity that the Andromeda system is going to continue to require.

但它确实邀请我们关注我们从BFCM期间的用户生成内容创作者广告中开始看到的那种精准性。

But it is an invitation into the the type of precision that we're beginning to see coming out of BFCM with UGC content creator ad specifically.

另一个从UGC内容创作者中浮现出来的点，与高价值客户这一概念相关，即在这个数据集中，UGC广告的平均订单价值（AOV）为140美元，而标准创意广告仅为84美元。

Another thing that popped out of the UGC content creator thing, which relates back to this idea of higher value customers, the the average order value, the AOV for UGC ads in this dataset was a $140 versus $84 on standard creative.

所以在Meta的术语中，我们一直谈论广告的命中率，而这是我看过的最有力的证据。

So in meta language, we've talked about the hit rate of an ad forever, and this is like the the most smoking gun that I've ever seen.

要获得这种类型的数据非常困难，因为数据量庞大且杂乱无章。

It's it's really hard to get to this this data this this type of data because it's so big and so messy.

例如，为了提取这些数据，我们必须从我们拥有的55,000条广告数据中筛选，通过广告名称、广告组名称或广告系列名称来识别UGC创作者内容的白名单。

So for example, we in order to actually extract this out, we had to take the 55,000 rows of data that we had around ads and look for to deduce UGC creator content whitelisting out of either the ad name or the ad set name or the campaign name.

要获得这个数据集真的非常棘手。

It was really tricky to get to this this data set.

但一旦我们做到了，信息就跃然纸上：如果你想对一个成功广告做出良好预测，这并非保证。

But once we did, the information just jumped off the page that if you if you wanna have a good bet at a winner, it's not a guarantee.

但如果你想提高获胜几率，这类内容对你的媒体组合至关重要。

But if you wanna have a good bat at a winner, this type of content is really important for your media mix.

有意思。

Interesting.

嗯，这个击球的类比其实很贴切，因为像0.353和0.195这样的数据，本质上就是顶尖棒球运动员和普通球员之间的差距，而这正是我们在这里所要探讨的差异。

Well, it's it's such a actually, the at bat analogy is pretty good because, you know, point three five three or whatever and point one nine five are, like, the difference between a great great great baseball player, one of the best of all times, and and that's kind of essentially the difference that we're looking for here.

我想请你稍微展开一下，关于这张幻灯片上提到的‘标准广告是更安全、更高效的选择’，你具体指的是什么？

I'm interested to for you to tease out a little bit like what what you mean on this particular slide when you say that standard ads are quote safer efficiency plays.

我们所观察到的是，UGC要么比标准广告失败得更惨，要么成功得更惊人，还是说背后还有别的原因？

If what we're looking at for is it that UGC is like feast or famine, like a UGC will either fail harder than a standard or succeed more wildly, or is there something else at play here?

是的。

Yeah.

完全正确。

That's exactly right.

标准广告组的平均回报率实际上比UGC更高。

So so the average return on the standard the standard ad set, the average ROAS is actually higher than the UGC.

所以我觉得‘饥荒或盛宴’这个说法非常贴切。

So I think feast or fam is a is a really good is a really good way of framing it.

对吧？

Right?

所以，你绝对不能只用这些信息去投资UGC风格的内容。

So so what you shouldn't do with this information is only invest in UGC style content.

那会非常危险。

That would be really risk.

对吧？

Right?

而且这会适得其反，完全违背了Meta关于创意多样性的指导建议。

And and counterproductive and and very much not in line with the the guidance that we're getting from Meta around creative diversity, etcetera, etcetera.

但你绝对需要投资这种内容风格，因为到2026年，我可以向你保证，明年你花费最高的广告之一，一定会是一个UGC广告。

But what you absolutely need to do is invest in this style of content because over the course of 2026, what I can guarantee you is that one of your top spending ads next year will be a UGC ad for sure.

好的。

Alright.

让我们来谈谈这张幻灯片。

Let's let's let's talk a little bit here about this slide.

用户制作的促销广告，价格仅为一半。

Promo ads by customers for half the price.

首先，这里有一些定义，促销广告和非促销广告有什么区别？

So first off, with some definitions here, what's promo promo ad versus non promo?

我们这里说的是什么？

What are we talking about here?

任何与黑色星期五、网络星期一促销相关的广告。

So anything that was specific to, like, a Black Friday, Cyber Monday offer.

好的。

K.

有意思。

Interesting.

这一直是长期以来被反复提及的一种说法，比如我甚至听过泰勒说过：‘宣传颜色，卖黑色。’

This has been another sort of trope that's got thrown around for for some time, which is, like, I've heard I've even heard Taylor say, like, advertise the color, sell the sell the black.

你懂的？

You know?

当然。

Sure.

是的。

Yeah.

这些年来，我们一直有一些轶事性的观察，比如我们为黑色星期五和网络星期一制作了大量促销广告，但全年最能带来高消费的常青广告，反而会在周末期间持续放量。

And we've seen we've had these anecdotal things that have popped up over the years around, like, we make all these prom promotional ads for Black Friday and Cyber Monday, but your your evergreen top spending ad from the rest of the year will be the one that will scale through through the weekend.

对吧？

Right?

嗯。

So Mhmm.

我们想稍微梳理一下，看看数据集中实际出现了什么情况，结果有几件事引起了注意。

We wanted to tease this out a little bit and see see what actually manifested in the in the dataset, and then a couple things stood out.

首先，支出的占比是均分的。

So number one, the share of the spend was was split evenly.

是的。

So Yeah.

当然，我们确实能够通过黑色星期五和网络星期一的促销广告完成媒体投放。

For sure, we were able to deliver media spend through promotional ads on Black Friday, Cyber Monday.

对我来说，这就像回到了Andromeda之前的时代。

And and to me, this is like going back to this, like, pre post Andromeda.

对吧？

Right?

在Andromeda之前，你有一个全年都在投放的常规广告，并且积累了大量的数据历史。

So pre Andromeda, you have a BAU ad that's been running all year long and has all of this data history with it.

所有这些数据历史都是对Andromeda之前的投放系统的一种信号。

And all of that data history is a signal to the to the pre Andromeda bidding machine.

这是一个好广告。

This is a good ad.

投放它。

Serve it.

对吧？

Right?

嗯。

Mhmm.

有时候，我们在推广广告的投放上会遇到一些挑战。

And sometimes we would have we would have some challenges getting our promotional ads to deliver.

明白吗？

Okay?

今年，后Andromeda时代，Andromeda知道这款BFCM促销广告在当前时刻极具相关性。

This this year, post Andromeda, Andromeda understands this this sale ad for BFCM is extremely topical for this moment in time.

投放它。

Deliver it.

我们在支出关系中也看到了这一点：我们的促销广告在该期间的支出占比大约达到一半。

And we see that in in the spend relationship where our our promo ads, like, accounted for approximately half of the spend during the period.

另一个值得注意的是，这多少有点回溯到传统的BAU广告或在网络星期一表现良好的广告。

The the other things that that popped out is and this is sort of like a little bit hearkening back to the the BAU ads or the ones that performed during Cyber Monday.

非促销广告的广告投资回报率实际上高出约十分之一点。

The the ROAS for nonpromotional ads actually was higher by, you know, a tenth of a point.

非促销广告的ROAS为2.55，而促销广告为2.4。

So point 2.55 for nonpromo ads versus 2.4 for promo ads.

所以这里的要点是，常规广告仍然非常重要。

So the the takeaway here is it it is very much still BAU ads.

在大型促销活动之前就存在于账户中的广告，在促销期间仍将继续发挥重要作用，但这并不意味着促销广告的投放效果会变差或表现明显逊色，也就是说，Andromeda 能够在正确的时间将正确的内容推送给正确的人。

Ads that have existed in the account historically leading up to leading up to a big sale are going to continue to be important during a sale moment, but it's not to say that the promo ads are going to deliver less or perform materially less in in any way, aka Andromeda delivering the right piece of content to the right person at the right time.

是的。

Yeah.

这次分析中最突出的发现是，抱歉。

The real standout from this analysis was sorry.

你先说，理查德。

Go ahead, Richard.

不。

No.

我本来想早点提一下这个观点。

I was gonna, like, kind of make the point earlier.

你刚才谈到的是 Andromeda 前后出价方式的真正差异。

You were you were talking about kind of like the this being a real of the difference between bidding styles pre Andromeda and post Andromeda.

因为正如我们在开始录制前稍微提到的，我们现在正从基于竞价的投放机制转向一种由AI驱动的算法

Because as we were talking about a little bit before we hit record here, the idea is that we're switching from an auction based, like, an auction based bidding kind of setup or whatever to an AI driven algorithmic

交付机器。

thing that I delivery machine.

内容投放系统能够理解创意内容，在某种程度上判断哪些内容会更相关。

Content delivery that reads that essentially is, like, able to read the creative and understand to some level that this is going to be more relevant.

因此，过去我们在这个播客中一直推崇的常规观点——即你的日常广告会表现最佳——

And so now the sort of, like, the conventional wisdom that we've certainly touted on this podcast that, like, your BAU ads are going to be the high performers.

现在平台发生了根本性变化，这一观点不再必然成立。

There's a fundamental shift in the platform now that makes that not necessarily the case.

显然，这些广告仍然表现得非常好，但作为获客工具来看，这种变化特别有意思。

Obviously, like, they're still they still perform really well, but it's just interesting to see this particularly as, like, an acquisition tool.

所以，我们讨论的是新客户获取，还是存在新客户与老客户的区分？或者只是潜在客户？

So if we're talking about we're talking about so are we talking about new customer acquisition here or just like, is there some sort of new versus returning split that's that we're looking at or is it just near leads?

在这个层面不是。

Not on this layer.