Marketero · Entrevistas
Bonus: Marginal ROAS & How Meta’s Algorithm Really Works w/ Constantine Yurevich
26 ideas de marketing sacadas de «Bonus: Marginal ROAS & How Meta’s Algorithm Really Works w/ Constantine Yurevich», de DTC Podcast. Sobre todo de canales…
- 26 ideas despiezadas
- Fuente DTC Podcast
- Dura 48 min
- Publicada en 2025
Ver la entrevista entera en YouTube
De qué habla
Las 26 ideas de esta entrevista se reparten así entre las etapas del marketing. El reparto es lo que distingue una entrevista de otra: dice en qué se detuvo de verdad quien hablaba.
Las 26 ideas, con su cita y su minuto
Cada una está sacada de un punto concreto de la entrevista. El enlace del titular lleva a
la página de esa idea, donde se ve quién más la dice por separado; el del pie salta al
segundo exacto del vídeo para que la compruebes tú.
Marginal ROAS, no Average ROAS: optimiza por la última unidad, no el promedio de toda la inversión
Medición y experimentosConstantine Yurevich: 'the first step is to forget about average ROAS and never optimize your campaigns based on average ROAS. Average ROAS doesn't matter'. La razón: si tu campaign tuvo 5X ROAS en los primeros $10K, eso NO significa que los próximos $10K generen otro 5X. La realidad: 'eventually you start investing more and at this spend point it starts diminishing greatly and you invest additional 5,000…
There are many misconceptions about ROS and there are also many different things you call ROS. Platforms do not show marginal roles because they want you invest as much as possible. A lot of D2C brands they invest so much into activities you should not be investing. When it comes to budget allocation, you should never look at is it better to just give Meta your budget and let them determine where it goes. For most of the D2C brands, for example, this is one of the most important things. I know how meta works and matter also has a model on creative level. Technology will always do whatever you…
Escúchalo en el minuto 0:00
Canales de adquisiciónConstantine: 'for example like retargeting campaigns — yes, last click is very biased towards retargeting for example on Facebook. But once you start investing more and more and more money into retargeting, even though it shows amazing ROAS based on last click attribution, at some point additional money you spend into retargeting won't bring any additional revenue because you just exhaust your audience and you show…
which attribution you use. So for example like retargeting campaigns. Yes, last click like uh of course last click is very biased towards retargeting for example on Facebook. But once you start investing more and more and more and more money into retargeting, even though it shows amazing roles based on last click attribution, at some point additional money you spend into retargeting won't bring any additional revenue because you you just exhaust your audience and you show this creatives and this ads so many times that people just keep clicking but they do not buy anymore. Like adding…
Escúchalo en el minuto 12:54
New events necesitan warm-up: no lances un campaign con un evento nuevo al frío, primero entrena el modelo
Canales de adquisiciónConstantine: 'sometimes you optimize towards purchase but then you see oh we don't have enough purchases let's test some new event. Maybe we create some fancy event, combine this and that and you launch new campaign and it doesn't perform. It doesn't perform not because this event is bad but because you don't have warm warmed up model for this event. And that's why what I usually recommend is first if you have event…
importantly it will reuse event level model so that's why sometimes you optimize towards purchase but then you see oh we don't have enough purchases let's test some new event. Maybe we create some fancy event, combine this and that and you launch new campaign and it doesn't perform. It doesn't perform not because this event is bad but because you don't have warm warmed up model for this event. And that's why what I usually recommend is first if you have event that you've never used before for example many of our clients especially lead generation businesses but also like D2C businesses who…
Escúchalo en el minuto 24:46
El ROAS promedio oculta el ROAS marginal negativo: la decisión de escalar debe basarse en el marginal
Medición y experimentosLos canales publicitarios muestran ROAS promedio (ingreso total dividido gasto total), pero este número puede ser positivo mientras el ROAS marginal —el retorno del último dólar invertido— ya es negativo por rendimientos decrecientes. Esto lleva a asignar más presupuesto a campañas que en realidad destruyen valor. La solución es medir el ROAS marginal mediante cambios controlados en el presupuesto (shifts de ±15%)…
you you invest $1,000 a day in your you get $2,000 a day. So your ROS is 2x, campaign. But this additional 1,000 mean it means like overall I invest average gross is still 1.25. is like in a sense negative. It's 0.5. marketers and most of the ad platforms campaign where you have like 1.1 rows marginal ros is already 0.5 so you're one of the solutions looking at it this because marginal rows can be measured call it uh controlled budget shifts and So we make this like plus minus uh 15% after that you just reallocate your roles is already 0.1. And it makes perfect sense in this example because
Escúchalo en el minuto 5:34
Medición y experimentosCuando el ROAS marginal alcanza 1x (por cada dólar invertido se recupera exactamente uno), cualquier gasto adicional es pérdida neta. Un caso concreto: una campaña que invierte $1.000/día con retorno de $2.000/día tiene ROAS promedio 2x, pero al duplicar el gasto a $2.000/día con solo $500 adicionales de retorno, el ROAS marginal baja a 0.5x (se pierde dinero en cada dólar incremental) aunque el ROAS promedio total…
you you invest $1,000 a day in your you get $2,000 a day. So your ROS is 2x, decide, okay, two is amazing. ROS like I campaign. But this additional 1,000 mean it means like overall I invest average gross is still 1.25. is like in a sense negative. It's 0.5. marketers and most of the ad platforms campaign where you have like 1.1 rows marginal ros is already 0.5 so you're one of the solutions looking at it this
Escúchalo en el minuto 5:34
Brand search campaigns no son gold: separá brand de non-brand en optimization, no confíes en last click
Medición y experimentosConstantine: 'most of the marketers already know this and they already do the same way as they separate prospecting from retargeting. So I believe this is already something we have in the marketing even for any junior marketer know that it should be evaluated separately. Of course like Google made an attempt to mix everything together but marketers know better. They've met a lot of resistance and a lot of critiques…
lot of the platforms with all of their AI automations, they like to include as much brand as possible to make their numbers look good as well. So they're it's almost like the AI developments are sort of incentivizing you to, you know, include as much brand uh spending into their algorithms so that their so that their numbers on platform look good. So that's another thing to be wary of. Yeah, of course. Of course, you should separate brand from non-brand. So I believe most of the marketers already know this and they already do the same way as they separate prospecting from retargeting. So I…
Escúchalo en el minuto 15:04
El ROAS marginal detecta saturación de campañas antes que el ROAS promedio
Medición y experimentosEl ROAS promedio no muestra cuándo una campaña empieza a rendimientos decrecientes: una campaña que rinde 5x ROAS con $200K de presupuesto puede caer a 4.8x o 4.7x al agregar $50K extra, y ese cambio luce inocuo en el promedio. En cambio, el ROAS marginal (el retorno del último dólar invertido) puede caer a 0.5x o 0.9x, indicando que el siguiente dólar gastado destruye valor. Ninguna plataforma de anuncios muestra…
imagine you've had a campaign which performed really great like 5x rows for a long time up up to like up to 200k budget it performed really well but now for some reason like uh the audience was exhausted like or for some reason like this campaign cannot perform and cannot scale anymore and it starts showing this these ads to the same audience and your returns might be diminishing quite rapidly. you already might be having like zero like like 0.90.5 marginal rows. Well, in average ROS you won't be able to detect this because it was so good for the such a huge budget. It was 5x now you invest…
Escúchalo en el minuto 37:30
Optimizar por leads o volumen nunca por valor produce leads baratos e inútiles
Medición y experimentosCuando una herramienta de atribución predictiva recibe como instrucción 'maximizar cantidad de leads' o 'minimizar CPA por lead', cumplirá la instrucción al pie de la letra: encontrará audiencias masivas de bajo costo (India, Arabia Saudita) que generan miles de leads por muy poco, pero que casi nunca convierten en clientes reales. La tecnología hace exactamente lo que le pides, no lo que realmente necesitas. El…
Yeah, I will give you an example. So, we've had a self-s serve client. Uh, and essentially they were in education space. So, actually it's like higher education where you can apply from anywhere in the world. It's a lead generation business and of course it's a long application process and after 3 months finally you pay for like 50k or 100k depends on the course you're picking. we've built an attribution for them and they started uh optimization based on marginal analytics and instead of optimizing towards leads or instead of implementing lead scoring and optimizing towards lead value they…
Escúchalo en el minuto 31:48
Static budget no revela marginal ROAS: haz ±15% budget shifts automáticos para construir la curva de diminishing returns
Medición y experimentosConstantine: 'marginal ROAS can be measured only in dynamics. So if you have static budget, the only thing that you know is average ROAS because you don't have any budget shifts, you don't know any changes, you don't know how additional spending influences additional revenue or how decrease in spending influences revenue'. El método: 'controlled budget shifts — our platform makes this automatically. We make this…
whole many companies do like this for the whole year you just have static daily spend. you just invest like $1,000 per day and you never ever change budget for this campaign. And because of this, there is no way to measure marginal rows because marginal rows can be measured only in dynamics. So if you have static budget, the only thing that you know is average rows because like you don't have any budget shifts, you don't know any changes. You don't know how additional uh spending uh influences additional revenue or how decrease in spending uh influences uh revenue. You don't know this. That's…
Escúchalo en el minuto 9:48
Top-funnel investment scales linearly: cuando tu bottom-funnel satura, expand budget a demand gen
Canales de adquisiciónConstantine: 'so many companies worry about the incrementality of their brand campaigns and of course like we know about the cannibalization and uh that like people might be searching for your brand but they click on your ads. So you keep investing more and more and even though like we know that last click is very biased towards brand but at some point brand returns diminish so greatly that even brand campaigns…
lot of the platforms with all of their AI automations, they like to include as much brand as possible to make their numbers look good as well. So they're it's almost like the AI developments are sort of incentivizing you to, you know, include as much brand uh spending into their algorithms so that their so that their numbers on platform look good. So that's another thing to be wary of. Yeah, of course. Of course, you should separate brand from non-brand. So I believe most of the marketers already know this and they already do the same way as they separate prospecting from retargeting. So I…
Escúchalo en el minuto 15:04
Canales de adquisiciónConstantine: 'meta has many machine learning models. It's not like you launch a new campaign and absolutely new machine learning model launches and starts training based on like 50 signals you send per campaign. So meta has layers. So it has like account level machine learning model that simply understands like learns based on all data in your account. Then it has like pixel level data. So you deploy pixel or…
So it's like yeah what can you say about really understanding the meta algorithm? So how meta works? So meta has many machine learning models. It's not like you launch a new campaign and like absolutely new machine learning uh model launches and it starts training based on like 50 signals you send per campaign. So meta has layers. So it has like account level machine learning model that simply understands like learns based on all data in your account. Then it has like pixel level data. So you deploy pixel or conversions API on your specific website. So it trains on the data of this specific…
Escúchalo en el minuto 23:49
Premature diversification: si tu single channel todavía tiene headroom, escala a $1M ahí antes de diversificar
Canales de adquisiciónConstantine: 'most of the brands they scale too fast. I've seen DTC brands that actually invest like $200K a month and they already diversify across TikTok, Snapchat, out of home, TV. But like what I've observed in the best scaling brands is that with Facebook only you can scale up to $1 million per month easily. The idea is why like all this diversification. Of course I understand the importance of diversification…
have been able to do that essentially unlock budget to bigger and better demand generation operations when they avoid spending too much saturating the bottom part of the funnel. What I see in the market that uh most of the decisions in terms of like whether invest in lowerfunnel, upper funnel, uh super upper funnel like brand awareness, so-called brand awareness at most out of home campaigns or something. Yeah. Most of the brands they scale too fast. So I've seen uh DTC brands that actually invest like 200k a month and they already diversify across Tik Tok, Snapchat, out of home, TV, they…
Escúchalo en el minuto 16:22
Reallocate por marginal ROAS: a veces mover budget de 1.5X Google a 1.1X Meta es lo correcto (cuando Google ya está saturado)
Medición y experimentosConstantine: 'sometimes we had really crazy recommendations. So some of our clients initially when we release this feature they ask like why are you recommending me to allocate budget from our like Google campaign which has like 1.5 ROAS to this Facebook campaign which has like 1.1 ROAS. So this definitely something is is something is wrong here. And then we show them actually your Google campaign already has…
after that you just reallocate your budget based on marginal rows instead of average rows. And sometimes we had really crazy recommendations. So some of our clients initially when we release this feature they ask like why are you recommending me to allocate budget from our like Google campaign which has like 1.5 rows to this Facebook campaign which has like 1.1 rows. So this definitely something is something is wrong here. Uh and then we show them actually your Google campaign already has diminishing returns and your like real like marginal roles is already 0.1. And it makes perfect sense in…
Escúchalo en el minuto 11:01
In Meta, broad targeting es optimal: intereses y lookalikes son innecesarios si tienes buen creative + eventos entrenados
Canales de adquisiciónConstantine: 'right now Meta probably is the only platform where this is absolutely not necessary. So in meta you can go like 100% broad and if you properly train algorithms of meta with proper signals meta will be able to find anyone in this world'. Caso personal: 'I have a dog and I was looking for new food for dog and I was chatting with my friend in WhatsApp and like in 30 minutes I've seen like hundreds of ads…
prospecting campaigns just to focus on reaching uh more new customers. Um but definitely u I I see a lot of inexperienced marketers who don't understand how Meta's algorithms work. So a lot of people still um create this audiences, they target based on interests, they build lookalike audiences etc. While right now Meta probably is the only platform where this is absolutely not necessary. So in meta you can go like 100% broad and if you properly train algorithms of meta with proper signals meta will be able to find anyone in this world and uh I I was even surprised I I have a dog and I was…
Escúchalo en el minuto 22:01
Brand awareness campaigns sin recognition no funcionan: no hagas TV/display hasta tener omnipresence
Canales de adquisiciónConstantine: 'a lot of D2C brands especially below like 1-2 million per month they invest so much into activities you should not be investing. It's kind of like marketing textbook thinking like I should be everywhere or I should be omni channel. DTC podcast told me to be omni channel so I should be everywhere rather than letting the data tell the picture. So a lot of what I see a lot of D2C brands especially below…
forget about it. So a lot of what I see a lot of D2C brands especially below like uh 1 2 million per month they invest so much into activities you should not be investing. It's kind of like marketing textbook thinking like I should be everywhere. or I should be omni channel. DTOC podcast told me to be omni channel so I should be everywhere rather than letting the data tell the picture. Yeah, it's very for example for our company so we are a software as a service company. So like we know exactly who are our target accounts. We can list like I don't know 100,000 companies like in D2C space in…
Escúchalo en el minuto 19:37
Con presupuestos estáticos es imposible medir ROAS marginal: hay que desestabilizar el sistema
Medición y experimentosSi el presupuesto diario de una campaña nunca cambia, solo se puede conocer el ROAS promedio, porque no hay variación en el gasto contra la cual medir el impacto incremental del último dólar. Para obtener el ROAS marginal hay que introducir cambios deliberados de presupuesto (shifts controlados de ±15% sobre el gasto actual) que permitan observar cómo responde el ingreso a cada nivel de inversión y así construir la…
campaign and for the whole year for the because marginal rows can be measured call it uh controlled budget shifts and So we make this like plus minus uh 15% And now once we have this diminishing after that you just reallocate your
Escúchalo en el minuto 9:45
Testea nuevos creatvos añadiéndolos a campañas existentes con presupuesto alto, no en campañas de test propias
Canales de adquisiciónMeta tiene modelos de aprendizaje automático a nivel de creativo que requieren datos para entrenarse. Cuando lanzas un nuevo creativo en una 'campaign de test' con presupuesto limitado, no hay suficientes señales para construir ese modelo y el test tarda semanas o no produce resultados útiles. La táctica correcta es añadir el nuevo creativo a una campaña que ya funciona bien con mucho presupuesto: ya existe un…
frankly speaking, yeah, we're exactly on the data side. So, we are not involved in launching any campaigns for our clients design. So, this is uh all of this is done inhouse or by digital agencies for sure. But of course I know how meta works and meta also has a model on creative level. So again all your creatives like uh meta has specific models that are linked to your creatives and of course like if you have really well performing creatives combined with event that you've been using in another campaign. When you launch an absolutely new campaign with new strategy of course you should be…
Escúchalo en el minuto 42:19
Cómo activar un nuevo evento de conversión en Meta sin gastar presupuesto en campañas de test
Canales de adquisiciónCuando creas un nuevo evento de conversión en Meta (por ejemplo para una campaña de lead generation o para optimizar LTV en D2C), no primero correr la campaña en modo 'warm up': enrégala solo a tu audiencia de retargeting o combínala con prospecting sin separar ambos. Una vez que el modelo de aprendizaje automático del evento se calibra (típicamente tras unas semanas), apagas esa campaña temporal y lanzas la nueva…
what I usually recommend is first if you have event that you've never used before for example many of our clients especially lead generation businesses but also like D2C businesses who optimize towards LTV for example or only towards new customers versus existing customers they create a new event and they just want to use this new event for optimization and they just run a new campaign it doesn't perform first you need to warm up your event level model and for this purpose you can create new event and you can run it only for your retarget targeting audience or in this particular case you…
Escúchalo en el minuto 25:08
No diversifiques canales ni campañas hasta tener 500+ conversiones en una sola combinación
Canales de adquisiciónCon presupuestos pequeños ($50K en adelante), dividir entre múltiples plataformas o múltiples campañas por plataforma es contraproducente. Facebook necesita enviar suficientes señales para construir modelos de aprendizaje automático robustos; si solo generas 500 conversiones totales entre 4 plataformas, cada una recibe datos insuficientes para optimizar bien. También es un error mirar el rendimiento de una campaña…
never ever diversify and never ever have many campaigns because like for smaller budget in many cases Facebook will perform uh worse compared to larger budget. So if you invest like 50k on Facebook, 50k on Tik Tok, 50k on Snapchat, 50k on YouTube, this is how machine learning works like to build a robust machine learning model and for Facebook to build this proper segment, proper lookalike audience based on the signals you send, you need to send as many signals as possible. So if you just invest 50K and you have just like I don't know 500 conversions, it might be not enough for Facebook to…
Escúchalo en el minuto 35:09
Canales de adquisiciónMeta tiene acceso a datos dentro de WhatsApp, Instagram y la red social que permiten identificar audiencias mejor que cualquier interés o lookalike creado manualmente. Lanzar campañas con audiencias de interés, lookalikes o audiencias personalizadas en Meta es innecesario si se entrena correctamente el algoritmo con eventos de conversión adecuados. La práctica recomendada es 100% broad targeting con señales de…
definitely u I I see a lot of probably is the only platform where this was chatting with my friend in WhatsApp your so perceived like audience you know conversions and they believe that this
Escúchalo en el minuto 22:09
Presupuestos pequeños en muchos canales impiden toda medición útil y se desperdician
Canales de adquisiciónDistribuir $10K mensuales entre TikTok, Snapchat, DV3, búsqueda pagada, Performance Max y display (6 canales) deja cada canal con volumen insuficiente para test geo holdout, pruebas incrementales o visitor scoring. Sin señal estadística, no se puede confirmar si un canal genera ventas o no. Constantine documentó el caso de una marca que no registraba conversiones last-click ni efecto medible por visitor scoring en…
Most of the brands they scale too fast. holdout tests you won't be able to invests like uh 10K a month on Tik Tok, no way. And when we connected our connected our like visitor scoring these campaigns are useless and they say believe that people going to see our
Escúchalo en el minuto 16:48
El modelo de atribución es irrelevante cuando los rendimientos decrecientes dominan: el ROAS marginal los hace desaparecer
Medición y experimentosEn campañas con rendimientos decrecientes intensos (retargeting saturado, brand search saturado), el ROAS marginal ya es negativo independientemente del modelo de atribución usado. Esto significa que debates como last-click vs multi-touch vsMM pierden relevancia: cuando la campaña está agotada, ninguna atribución muestra rentabilidad incremental. La consecuencia práctica es que la señal para cortar o reducir…
understand that you should be investing unleashed that actually attribution which attribution you use. So for additional money you spend into click but your marginal ROS can be they click on your ads but this exactly very very bad and in this case like look good even based on last click
Escúchalo en el minuto 12:20
Las plataformas de anuncios no quieren que uses ROAS marginal porque mostraría que tu siguiente dólar pierde dinero
Medición y experimentosNinguna plataforma publicitaria (Meta, Google, TikTok) incluirá la métrica de ROAS marginal en su panel, porque la recomendación universal de sus account managers ('aumenta el presupuesto') es exactamente lo opuesto a lo que esa métrica revelaría: que existe un punto donde cada dólar adicional invertido genera menos retorno que el anterior, y eventualmente destruye valor. Por eso el ROAS marginal es un diferenciador…
brands need marginal rows and there are no ad platforms that can calculate this because for ad plat platforms this will be a disaster. it would like like the best recommendation of most of account managers in ad platforms is increase budget and if they going to introd if they if they going to introduce marginal roles into Google or Facebook most of the recommendations will be decrease budget that's why I believe this feature will never be introduced in this platform so uh we see like that for most of the D2C brands for example this is one of the most important things because like profit…
Escúchalo en el minuto 28:43
Marcas DTC sin reconocimiento de marca no pueden replicar el modelo omni-canal de Coca-Cola
Canales de adquisiciónLa idea de estar "en todas partes" funciona para Coca-Cola porque tiene presencia en cada punto de venta, reconocimiento masivo acumulado y saturación de marca. Cuando una marca DTC que nadie conoce lanza una campaña en TV, display o YouTube sin esa base, los usuarios ven el anuncio, no lo recuerdan al día siguiente y no compran. El trigger de "vi el anuncio y ahora lo veo en el estante" no existe para quien no…
like when you're a well-known brand when skip in in 5 10 minutes you're going to invest so much into activities you software as a service company. So like campaign, you might be just like uh may not be in every retail store that the store versus someone who's spending,
Escúchalo en el minuto 18:56
Los incentivos de plataformas, agencias y CMOs están alineados en aumentar gasto, no en maximizar beneficios incrementales del cliente
Medición y experimentosCada actor del ecosistema publicitario tiene un objetivo propio que no coincide necesariamente con el del anunciante: Meta y Google quieren cobrar premium por cada dólar; las agencias que cobran comisión sobre el presupuesto gastado ganan más al escalar inversión; los CMOs que protegieron su posición con campañas de TV necesitan una tecnología de medición que justifique esa inversión ante el CFO. La única persona…
You can just add this creative to your existing campaign and see whether it flies or dies. Yeah. Or crashes. Um this I think this is good news. Like I think a lot of times people are talking about okay in the next 2 to 3 years media buyers will be you know uh replaced by AI but with what you're saying these platforms are not incentivized to be thinking about uh you know be thinking about marginal rorowass in the way that you're talking about it. Do you have any predictions for the media buying profession and the media buyers in the next two to three years? Yeah. So I I believe the biggest…
Escúchalo en el minuto 44:24
Meta es el único canal que puede absorber $1M mensuales con audiencias 100% broad sin perder rendimiento
Canales de adquisiciónSegún Constantine, con Facebook/Meta solo es posible escalar hasta $1M mensuales manteniendo retornos positivos antes de necesitar diversificación. Esta cifra funciona como punto de referencia: antes de añadir TikTok, Snapchat, OOH o TV, una marca DTC debería agotar primero la capacidad de Meta con audiencias broad y señales adecuadas. Esta directriz cuantifica el umbral de escala típico para una marca DTC antes de…
So I've seen uh DTC brands that actually different things. But like what I've with Facebook only you can scale up to 1 because you won't be able to run geio platform like Meta, do you see
Escúchalo en el minuto 16:52
Seguir leyendo
Esta es una de las 1872 entrevistas despiezadas. Puedes
ver todas, leer las ideas por tema o
empezar por la ruta guiada si no sabes por dónde.