Media Mix Modeling: A Strategic Guide for Growth Leaders

Media mix modeling (MMM) is a statistical method that uses aggregate historical data to estimate how much each marketing channel and tactic contributes to a business outcome, typically revenue or conversions. Its core use is strategic budget allocation: you run the model, see which channels are working and which are burning capital, and reallocate accordingly. According to eMarketer, a significant portion of US brand and agency marketers plan to increase investment in MMM, driven largely by signal loss from privacy changes that have made user-level attribution unreliable.

Use MMM when you need cross-channel visibility that spans both offline and digital, when privacy constraints limit user-level tracking, or when you need board-ready scenario planning with confidence intervals rather than last-click guesses.

The three-point verdict:

• Ideal use cases: Annual and quarterly budget planning, channel mix rebalancing, and measuring offline media (TV, radio, out-of-home) alongside digital.
• Expected output: Channel contribution estimates, marginal ROAS by channel, saturation curves, and simulated allocation scenarios with uncertainty ranges.
• Who should own it: A senior analytics lead or growth executive, with finance as a co-owner on the budget implications and a data engineer handling input preparation.

Key Takeaways

Media mix modeling is the most reliable method for cross-channel, privacy-safe budget allocation, but its value depends entirely on data quality, validation discipline, and organizational commitment to acting on the outputs.

Table of Contents

• What media mix modeling actually measures, and how it differs from "marketing mix modeling"
• How the core mechanics of MMM actually work
• What data you need before building a model
• Step-by-step phases from data intake to scenario planning
• What an MMM actually produces and how to use the outputs
• How MMM compares with MTA and incrementality experiments
• Known limitations and how to mitigate them
• How to operationalize MMM outputs into planning cycles
• A practical 90-day plan to get your first MMM running
• Connecting MMM outputs to broader marketing strategy and budget planning
• Real-world applications where MMM has driven measurable outcomes
• Tools and platforms commonly used for MMM
• Where MMM is heading: AI, machine learning, and faster cadence
• Handling digital and offline channels together in an omnichannel model
• What I actually focus on when running an MMM engagement
• What Ashafrazier delivers for teams ready to act on MMM
• Sources

What media mix modeling actually measures, and how it differs from "marketing mix modeling"

The terms get used interchangeably, but the distinction matters when you are scoping a project. Wikipedia's entry on marketing mix modeling draws the line clearly: media mix modeling focuses on paid media channels (TV, paid search, social, display, audio, out-of-home), while marketing mix modeling is the broader discipline that can also include pricing, distribution, promotions, and non-media factors like product availability.

In practice, most modern MMM projects sit somewhere between the two. A paid-media-only model is faster to build and easier to explain to a CMO. A full marketing mix model that includes pricing elasticity and distribution coverage is more complete but requires richer data and longer timelines.

The privacy-safe, aggregate-data advantage is what makes MMM relevant again in 2026. The model never touches individual user records. It works on time-series aggregates, which means it functions in a cookieless environment where MTA tools are losing signal. Common outputs include:

• ROAS and marginal ROAS by channel: What each additional dollar in a channel returns at current spend levels.
• Contribution estimates: What percentage of total sales each channel drove over the measurement period.
• Saturation curves: Visual representations of diminishing returns, showing where a channel is over-invested relative to its response.

How the core mechanics of MMM actually work

Three building blocks drive every MMM output: adstock (carryover), saturation (diminishing returns), and baseline/seasonality decomposition. Understanding these is what separates a team that can interrogate a model from one that treats it as a black box.

Adstock models the fact that advertising exposure does not produce an immediate and complete response. Some effect carries forward into future periods. The Think with Google MMM Guidebook describes adstock as a decay transformation applied to spend or impression data before it enters the regression. Higher decay rates suit brand channels; lower rates suit direct-response formats where the effect is immediate and short-lived.

Saturation functions model diminishing marginal returns. As spend in a channel increases, each additional dollar produces less incremental outcome. The Hill function and the Michaelis-Menten curve are the two most common shapes used in practice. The output is a response curve: flat at low spend, steep in the middle, and flattening again at high spend. Reading where your current spend sits on that curve tells you whether you are in the efficient zone or past the point of productive investment.

Baseline and seasonality capture everything that drives sales independent of media: organic demand, seasonal patterns, price changes, competitor activity, and macro conditions. Separating baseline from media-driven contribution is what makes the channel estimates credible. Without it, a model will attribute holiday demand spikes to whatever media happened to be running at the time.

Bayesian MMM methods, as described in Google Research's paper on carryover and shape effects, estimate full posterior distributions over parameters rather than single point estimates. This gives you credible intervals on every channel's contribution, which is far more useful for board-level scenario planning than a single ROAS number with no uncertainty range. The trade-off: Bayesian models are sensitive to prior choices when sample sizes are small, so prior selection requires deliberate justification.

Model families break into two camps. Additive models assume channels contribute independently and their effects sum linearly. Multiplicative models allow interaction effects, where the impact of one channel scales with the level of another. HBR's practitioner refresher notes that multiplicative specifications often fit consumer goods data better because media effects tend to amplify baseline demand rather than add to it flatly. Classical OLS regression is still widely used for its interpretability. Bayesian estimation, via tools like Stan or Pyro, is increasingly preferred when teams need uncertainty quantification.

Pro Tip: Apply the response-curve (saturation) transformation before the adstock transformation, not after. The Think with Google Guidebook is explicit on this order: saturation first, then carryover. Reversing the order produces systematically biased coefficient estimates.

What data you need before building a model

The quality of an MMM output is a direct function of input quality. No amount of modeling sophistication fixes bad data.

Required inputs checklist:

• Weekly or daily time-series spend by channel and placement (TV, paid search, paid social, display, audio, OOH, email, affiliate).
• Impressions or GRPs where available, especially for brand channels where spend alone is a poor proxy for exposure.
• Primary outcome metric at the same temporal granularity: revenue, transactions, or qualified leads.
• Pricing and promotional flags: price changes, discount events, and promotional periods.
• Distribution and availability signals: out-of-stock events, new market entries, or distribution expansions.
• External controls: holiday calendars, weather indices for weather-sensitive categories, and macro indicators like consumer confidence for high-consideration purchases.

The Think with Google Guidebook recommends impression-level variables wherever possible, because spend can be held constant while reach and frequency shift significantly, producing different response patterns that spend alone cannot capture.

Recommended granularity and minimum window:

A minimum of multiple years of weekly data is recommended to enable models to separate seasonal effects from media effects, which improves the reliability of channel estimates.

Common pitfalls that break model validity:

• Multicollinearity: Channels that always move together (TV and digital brand spend both ramping for every product launch) make it impossible to isolate individual effects. The fix is variance in spend patterns, either historically or through designed experiments.
• Missing channels: Omitting a channel that is correlated with included channels pushes its effect onto the nearest correlated variable, biasing all coefficients.
• Inconsistent time windows: Mixing fiscal-week and calendar-week data, or aggregating some channels monthly and others weekly, introduces systematic error.
• Data leakage: Including future information in the training window (for example, using a full-year promotional calendar as a feature when the model is being validated on a holdout period) inflates apparent accuracy.
• Non-stationary inputs: Revenue or spend series with strong trends need detrending or differencing before modeling, or the model will fit the trend rather than the media signal.

Step-by-step phases from data intake to scenario planning

A well-run MMM project moves through six phases. The timeline and cost vary significantly depending on whether the team is doing it in-house, using a hybrid approach with external tooling, or commissioning a full consultant engagement.

Phase summary: Discovery and data audit → feature engineering → model selection and estimation → validation and backtest → scenario planning and optimization → operationalization.

Phase checklist and decision gates:

1. Discovery: Confirm outcome metric, identify all spend sources, audit data completeness, and flag gaps before any modeling begins.
2. Feature engineering: Apply adstock and saturation transformations, create control variables, and document all transformation choices.
3. Model selection: Choose additive vs. multiplicative specification, classical vs. Bayesian estimation, and set priors if Bayesian.
4. Validation: Run a holdout backtest on the final 8–13 weeks of data. The model should predict holdout-period outcomes within an acceptable error range before any outputs are used for decisions.
5. Scenario planning: Use the validated model to simulate reallocation scenarios and generate marginal ROAS curves for each channel.
6. Operationalization: Schedule model refresh cadence, assign data steward, and integrate outputs into the quarterly planning cycle.

Cost ranges depend primarily on data engineering complexity, modeling cadence, and required customization. A DIY approach using open-source tools (Python with PyMC, or R) has near-zero software cost but significant internal labor. Hybrid approaches using commercial MMM platforms typically run in the range of tens of thousands of dollars annually. Full consultant-led engagements for mid-market companies generally run from low six figures for a one-time project to higher retainers for ongoing quarterly refreshes. Gartner notes that modern platforms have made MMM faster and more accessible than the legacy perception suggests, which has meaningfully compressed both timelines and costs.

What an MMM actually produces and how to use the outputs

The canonical outputs of a completed MMM are five things: channel contribution estimates, marginal ROAS by channel, saturation curves, simulated allocation scenarios, and uncertainty intervals around each estimate.

Channel contribution tells you what percentage of total measured sales each channel drove over the period. A typical waterfall chart shows baseline (organic demand), then stacked media contributions.

Marginal ROAS is the more actionable number. It answers: at current spend levels, what does one additional dollar in this channel return? A channel with a high average ROAS but a low marginal ROAS is saturated. You are getting good returns on existing spend, but the next dollar is inefficient. The marginal ROAS curve is what drives reallocation recommendations.

Saturation curves visualize the full response function for each channel. Reading them correctly requires noting where current spend sits on the curve. Spend in the flat upper portion means diminishing returns have set in. Spend in the steep middle portion means the channel still has room to scale.

Simulated allocation scenarios are where MMM becomes a planning tool rather than a reporting tool. You hold total budget constant and shift spend across channels to find the allocation that maximizes predicted outcome. Most modern platforms generate these scenarios interactively.

Statistic to anchor decisions: eMarketer reports that over a quarter of US brand and agency marketers identify MMM as a reliable measurement methodology, a figure that reflects both the method's maturity and the growing unreliability of user-level attribution in a privacy-constrained environment.

Uncertainty intervals matter more than most teams acknowledge. The first requires more data or an experiment before you act on it. The second is a reliable signal.

How MMM compares with MTA and incrementality experiments

These three methods answer different questions. Treating them as interchangeable is one of the most expensive mistakes in measurement strategy.

eMarketer's MMM FAQ frames the architecture clearly: MMM provides long-term strategic allocation guidance, MTA handles near-real-time digital touchpoint optimization, and incrementality experiments deliver the highest causal precision for specific channel or creative questions.

Recommended measurement architecture:

• Use MMM for annual and quarterly budget allocation across all channels, including offline.
• Use MTA (where signal quality permits) for in-flight optimization of digital campaigns within the budget envelope MMM sets.
• Use geo or holdout experiments to validate MMM channel coefficients for your two or three highest-spend channels annually. Experiment results should feed back into prior calibration for Bayesian models.

The triangulation principle: no single method is sufficient. MMM tells you where to allocate; experiments confirm whether the allocation assumption is causal; MTA tells you which creatives and placements to optimize within the channel.

Known limitations and how to mitigate them

MMM is observational, not experimental. That distinction carries real risk if teams treat model outputs as ground truth rather than informed estimates.

Critical limitations:

• Aggregation bias: Weekly or monthly aggregates can mask within-period variation. A channel that performs well on Tuesdays and poorly on weekends looks average in weekly data.
• Multicollinearity: Correlated spend patterns across channels make coefficient estimates unstable. The model may assign credit to the wrong channel simply because they moved together historically.
• Omitted-variable bias: Any channel or factor that influences sales but is not in the model will have its effect absorbed by correlated variables that are included.
• Small-sample sensitivity: As Google Research's Bayesian MMM paper documents, when datasets are small, posterior distributions become dominated by prior choices. A poorly chosen prior can produce confident-looking but misleading results.
• Model mis-specification: Choosing the wrong functional form (additive when the true relationship is multiplicative, or vice versa) produces systematically biased contribution estimates.

How each limitation misleads decisions:

1. Multicollinearity inflates the coefficient of whichever correlated channel the model happens to favor, leading to over-investment in that channel.
2. Omitted-variable bias can make a channel look effective when it is simply correlated with an unmeasured driver like distribution expansion.
3. Small-sample sensitivity in Bayesian models can produce narrow credible intervals that look precise but reflect prior assumptions more than data.

Mitigation checklist:

• Run a holdout backtest on the final 8–13 weeks of data before using any outputs for decisions. HBR's practitioner guide identifies failure to backtest as the primary cause of overconfident allocation recommendations.
• Conduct sensitivity analysis: vary key parameters (decay rates, saturation shape) and observe how much outputs change. High sensitivity to parameter choice signals model fragility.
• Cross-validate channel coefficients against geo experiments for your largest spend channels.
• Document all prior choices in Bayesian models and make them available for stakeholder review.
• Stage budget reallocations: implement 20–30% of the recommended shift in the first quarter, measure actual results, and compare against model predictions before committing to the full reallocation.

How to operationalize MMM outputs into planning cycles

A model that produces outputs once and sits in a folder is a sunk cost. The goal is a repeatable process where MMM informs budget decisions on a defined cadence.

Define owners before you define cadence. Three roles are non-negotiable: a model steward (owns the technical model and refresh process), a data steward (owns input data quality and pipeline), and a decision owner (the CMO or VP of marketing who commits budget based on outputs). Without named owners, model outputs drift into advisory status and get ignored when they conflict with intuition.

Governance checklist:

• Version control for model code and configuration files.
• Documentation of all data sources, transformation choices, and prior selections.
• A data provenance log that tracks when inputs were last updated and by whom.
• Decision thresholds: define in advance what marginal ROAS level triggers a reallocation recommendation and what confidence interval width requires an experiment before acting.
• A scorecard that compares model predictions against actual outcomes each quarter.

Integrating outputs into the growth marketing KPI dashboard is the step most teams skip. MMM outputs should appear alongside CAC, LTV, and payback period in the same planning view, not in a separate analytics report that finance never reads.

A practical 90-day plan to get your first MMM running

The 90-day outcome is either a validated model with scenario outputs ready for a board presentation, or a clear handoff plan if the data gaps discovered in week two require remediation before modeling can begin.

Week-by-week milestones:

1. Weeks 1–2: Stakeholder alignment. Define the outcome metric, confirm the measurement period, identify all spend data sources, and assign the three owner roles.
2. Weeks 3–4: Data audit. Pull all spend and outcome data. Document gaps, inconsistencies, and missing channels. Decide whether gaps are fatal to the model or manageable with proxies.
3. Weeks 5–6: Data remediation and feature engineering. Fix identified gaps, apply adstock and saturation transformations, build control variables.
4. Weeks 7–9: First model run. Estimate the model, review coefficient signs and magnitudes for face validity, and flag any results that contradict known business logic.
5. Weeks 10–11: Holdout validation and backtest. Withhold the final 8–13 weeks of data, run the model on the training period, and evaluate prediction accuracy on the holdout.
6. Week 12: Scenario planning and board presentation. Generate three to five allocation scenarios, attach uncertainty intervals, and present marginal ROAS curves by channel.

When to DIY vs. when to hire:

DIY works when your team has a data scientist with regression and Bayesian modeling experience, clean spend data going back at least two years, and a decision-maker who will wait 12–16 weeks for outputs. Hire a consultant when speed matters, when the data requires significant engineering, or when the model needs to be credible to a board or investor audience that will scrutinize the methodology.

Questions to ask any consultant: How do you handle multicollinearity in correlated spend patterns? What is your holdout validation protocol? How do you communicate uncertainty to non-technical stakeholders? What does the handoff look like so the internal team can refresh the model independently?

Pro Tip: The most common reason a 90-day MMM project fails is not the modeling. It is discovering in week three that spend data for two major channels is inconsistent or missing for 18 months. Run the data audit before you scope the modeling timeline, not after.

Connecting MMM outputs to broader marketing strategy and budget planning

MMM outputs are most valuable when they feed directly into the annual planning cycle rather than arriving as a post-hoc analysis. The channel priority matrix framework is a natural complement: MMM tells you the measured contribution and marginal efficiency of each channel; the priority matrix tells you where to test new channels that are not yet in the model.

The practical integration looks like this. In Q3, run the annual model refresh. Use saturation curves to identify channels that are over-invested relative to their marginal ROAS. Shift budget from saturated channels into under-invested ones or into new channel tests. In Q4 planning, present the MMM-backed allocation alongside CAC and LTV projections. Finance signs off on a budget that is grounded in measured returns rather than historical precedent or gut feel.

One discipline that separates teams that get value from MMM and those that do not: they treat the model's uncertainty intervals as budget guardrails. If a channel's marginal ROAS credible interval spans a wide range, they do not reallocate heavily based on that estimate alone. They run a geo experiment to tighten the estimate, then reallocate. This staged approach, described in the era of efficient growth framework, is what prevents MMM from becoming a sophisticated way to light capital on fire with more confidence.

Real-world applications where MMM has driven measurable outcomes

The most instructive MMM case studies share a common structure: a business with significant offline spend, a measurement gap that last-click attribution could not fill, and a reallocation decision that improved total ROAS without increasing total budget.

A consumer packaged goods brand running TV, digital video, paid search, and in-store promotions is the classic MMM use case. The model typically reveals that TV has a higher contribution to baseline demand than digital attribution tools suggest (because TV drives search behavior that gets credited to paid search in last-click models), and that paid search is partially capturing TV-driven demand rather than generating it independently. The reallocation: reduce paid search budget slightly, maintain TV, and reinvest the savings into digital video, which the MMM shows is under-invested relative to its marginal ROAS. Total budget stays flat; measured revenue increases.

A direct-to-consumer brand scaling from $20M to $100M in revenue faces a different version of the same problem. As paid social spend scales, marginal ROAS declines, but the team cannot tell whether the decline is saturation, creative fatigue, or audience exhaustion. MMM separates these effects by modeling the saturation curve independently of creative performance. The output: the channel is not saturated at current spend levels, but the adstock decay rate has shortened, suggesting creative fatigue. The fix is creative refresh, not budget reduction. This kind of diagnostic is impossible with MTA alone.

For a multi-location services business, MMM can incorporate local market variables (population density, competitive presence, local promotions) alongside national media spend, producing market-level contribution estimates that inform where to concentrate media investment for the next expansion phase.

Tools and platforms commonly used for MMM

The MMM tooling landscape spans open-source frameworks, commercial platforms, and full-service measurement providers.

Open-source frameworks are the starting point for teams with strong data science capability. PyMC (Python) and Stan (available via R and Python interfaces) are the most widely used Bayesian modeling libraries. Meta's open-source Robyn package, built in R, provides a full MMM workflow including adstock and saturation transformations, budget optimizer, and visualization outputs. Google's Meridian (released in 2024) is a Python-based Bayesian MMM framework designed for scalability and incorporates the carryover and shape-effect methods from Google Research's published work.

Commercial platforms add workflow automation, data connectors, and scenario planning interfaces on top of the modeling core. Platforms in this category vary widely in their modeling transparency, so evaluating how well a vendor explains its adstock and saturation specifications is a reasonable proxy for overall rigor.

Full-service measurement providers combine platform access with analyst support. These are appropriate when the internal team lacks modeling expertise or when the engagement requires custom specification for complex channel mixes.

The choice between open-source and commercial depends on three factors: internal data science capacity, required refresh cadence, and the need for non-technical stakeholder interfaces. A team that can maintain a PyMC or Robyn workflow gets maximum transparency and control. A team that needs a CMO to run scenarios without analyst support needs a commercial interface.

Where MMM is heading: AI, machine learning, and faster cadence

The most significant shift in MMM over the past three years is speed. Legacy MMM projects took six months and cost hundreds of thousands of dollars. Gartner's current guidance explicitly tells CMOs to reassess that assumption: modern tooling has compressed timelines and made near-real-time scenario planning achievable.

Machine learning is entering MMM in two specific ways. First, automated feature selection and hyperparameter tuning reduce the manual effort in model specification. Second, gradient-boosted models and neural networks are being tested as alternatives to regression-based MMM for datasets large enough to support them, though interpretability remains a challenge. The Bayesian approach remains dominant for most business applications because it produces uncertainty intervals that regression and ML alternatives do not naturally provide.

The other major trend is the integration of MMM with incrementality testing infrastructure. Rather than running experiments and MMM as separate workstreams, leading measurement teams are using experiment results to calibrate MMM priors, producing models that combine the causal validity of experiments with the cross-channel coverage of MMM. This is the direction the Google Research Bayesian MMM framework points toward: priors informed by experimental lift estimates rather than set arbitrarily.

Privacy regulation will continue to accelerate MMM adoption. As third-party cookies complete their deprecation and mobile identifier access tightens, the aggregate, privacy-safe architecture of MMM becomes not just a preference but a practical necessity for measuring offline and cross-channel impact.

Handling digital and offline channels together in an omnichannel model

The omnichannel challenge in MMM is not technical. It is a data alignment problem. Digital channels produce daily or even hourly spend and impression data. Offline channels like TV and radio produce weekly GRP data at best, and out-of-home often produces only monthly spend figures. Aligning these at a consistent temporal granularity without losing signal is the core engineering challenge.

The practical approach: model at weekly granularity and aggregate digital data up to weekly. This loses some within-week variation in digital performance but preserves the ability to include TV, radio, and OOH in the same model. For businesses where within-week digital variation is strategically important, a two-stage model can work: a weekly MMM for cross-channel allocation, and a separate daily digital model for in-channel optimization.

Offline channels require proxy variables when direct exposure data is unavailable. TV GRPs by market are the standard proxy for TV exposure. For OOH, spend by market is often the only available input. The key discipline is consistency: use the same proxy definition across the entire measurement window. Changing how you measure a channel mid-series introduces a structural break that the model will misinterpret as a genuine change in channel effectiveness.

One underappreciated benefit of including offline channels in MMM: it often reveals that digital attribution tools have been over-crediting digital channels for demand that offline media created. TV drives branded search. Radio drives direct traffic. OOH drives foot traffic that converts in-store. None of these pathways appear in a digital attribution model. MMM captures them all, which is why the channel contribution estimates from MMM and from MTA rarely agree, and why MMM is the more complete picture for strategic allocation.

What I actually focus on when running an MMM engagement

I think that ratio is backwards.

The technical work matters, but the model is not the hard part. The hard part is getting clean, consistent spend data across every channel for two-plus years, convincing the media team to share data they consider proprietary, and then presenting uncertainty intervals to a CMO who wants a single number to put in the board deck. Those are organizational and communication challenges, not statistical ones.

In practice, I spend the first two weeks of any engagement on the data audit, not the model. If the spend data has gaps, inconsistencies, or missing channels, no amount of modeling sophistication fixes it. I would rather delay the model start by three weeks to fix the data than run a fast model on bad inputs and spend the next quarter defending outputs that do not hold up to scrutiny.

On modeling choices: I prefer Bayesian specifications for most engagements because the credible intervals are genuinely useful for staged reallocation decisions. But I am explicit with clients that prior choices matter and that I will document and justify every prior. A Bayesian model with undisclosed priors is not more rigorous than OLS. It is just less transparent.

The trade-off I accept most often is cadence over precision. A quarterly model refresh with good-enough data beats an annual model with perfect data. The business environment changes. Spend patterns shift. A model that is 18 months old is not a measurement tool; it is a historical artifact. I would rather have a slightly noisier model that is current than a precise model that is stale.

What Ashafrazier delivers for teams ready to act on MMM

Running a media mix model is one thing. Building the organizational capability to act on it, quarter after quarter, is another. Ashafrazier's growth consulting engagements cover the full arc: data audit and spend alignment, model specification and estimation, holdout validation, scenario planning, and a structured handoff so your internal team can refresh the model independently.

Engagement models are scoped to where you are. A fixed-scope project delivers a validated model and board-ready scenario outputs in 5–7 weeks. A retainer adds quarterly refreshes and ongoing scenario planning as the budget cycle demands. For companies that need ongoing strategic leadership to operationalize outputs, a fractional CMO engagement integrates MMM into the full planning cycle alongside CAC, LTV, and channel prioritization decisions. Every engagement includes documented methodology, so the outputs are defensible to finance and investors, not just the marketing team.

If you want to pressure-test your current measurement approach before committing to a full engagement, the Growth Score Calculator is a fast way to identify where your unit economics and attribution gaps are costing you the most.

Sources

The sources below are the primary references behind the technical claims and practitioner guidance in this article.

• FAQ on media mix modeling: How to build a modern measurement strategy - eMarketer
• Marketing mix modeling - Gartner
• Marketing Mix Modeling Guidebook - Think with Google
• Bayesian methods for media mix modeling with carryover and shape effects - Google Research
• Marketing mix modeling - Wikipedia

Recommended

• The Channel Priority Matrix: Where Early-Stage Companies Should Actually Spend — Asha Frazier
• Things I've Learned — Asha Frazier
• The 60-Day Turnaround: Taking a Cash-Burning DTC Brand to a $100M Exit — Asha Frazier
• Marketplace Growth: When to Push Supply vs. Demand (And What Actually Compounds) — Asha Frazier