If your attribution tool and your gut disagree, the answer is usually not marketing mix modeling, at least not yet. Multi-touch attribution (MTA) follows individual journeys and credits the clicks it can see; marketing mix modeling (MMM) uses aggregate spend and sales data to estimate what every channel contributed, including the ones nobody clicks. Which one you need depends on your spend, your data history and the decision you’re making, and most growth-stage companies get further by combining lighter methods first.
How multi-touch attribution works and where it breaks
MTA stitches together the touchpoints a user had before converting and splits credit among them. Rules-based models use fixed logic: first touch, last touch, linear, U-shaped. Data-driven models compare converting and non-converting paths and assign credit statistically. GA4’s data-driven attribution, CRM attribution reports in tools like HubSpot and standalone attribution software all work this way.
Its strengths: it’s close to real time, reaches down to ad and keyword level, and matches how teams optimize day to day.
It breaks in predictable places:
- It only sees what it tracks. Consent refusals, ad blockers, cookie limits and device switches leave holes, and walled gardens like Meta and Google don’t hand outside tools user-level impression data.
- It measures presence, not cause. A touchpoint that shows up before a sale gets credit whether or not it changed anything. Retargeting and brand search sit right before purchase, so they collect credit for demand other channels created.
- Dark and offline channels are invisible. Podcasts, word of mouth, events, communities, and the LinkedIn post or ChatGPT answer someone read without clicking.
- B2B buying committees fragment the path. The person who fills in the demo form is often not the one who first found you, so contact-level journeys miss the start of the story.
That’s usually why the tool and your gut disagree: the tool is accurate about clicks and silent about everything else. If the disagreement is between tools rather than between a tool and your gut, see why GA4 and ad platforms don’t match.
How marketing mix modeling works
MMM ignores individual users. It takes a time series, usually weekly, of your outcome (revenue, orders, new customers or pipeline) alongside each channel’s spend or impressions, and fits a statistical model that estimates each channel’s contribution. It also controls for what moves sales without marketing: seasonality, holidays, pricing, promotions, launches and the baseline demand you’d get anyway.
Two transformations make it more than a simple regression:
- Adstock (carryover). An ad that runs this week keeps working for a while. The model estimates how fast each channel’s effect decays.
- Saturation (diminishing returns). The last thousand dollars in a channel does less than the first. The model fits a response curve for each channel.
The outputs are channel contribution and ROI, marginal ROI at current spend, and response curves you can plan budgets against. Open-source options include Meta’s Robyn, Google’s Meridian and PyMC-Marketing. Bayesian models can take your experiment results as priors, which anchors the model to causal evidence instead of correlation alone.
MMM has its own failure modes:
- It needs variation. If every channel’s spend rises and falls together, the model can’t tell them apart. Flat, steady spend gives it almost nothing to learn from.
- It’s slow and coarse. You get channel-level answers monthly or quarterly. It won’t tell you which creative or ad set to scale.
- It’s sensitive to setup. Different reasonable modeling choices can produce different answers. An uncalibrated MMM can be confidently wrong.
- Small channels disappear. A channel carrying a small slice of spend usually sits below the noise.
Data and spend requirements
| Multi-touch attribution | Marketing mix modeling | |
|---|---|---|
| Input data | User-level touchpoints and conversions | Weekly spend, outcomes and control variables |
| History needed | Works as soon as tracking is clean | Typically two or more years of weekly data |
| Dependence on tracking | High: consent, cookies, identity | Low: no user tracking needed |
| Offline and view-based media | Mostly invisible | Included if you record spend or impressions |
| Granularity | Campaign, ad, keyword | Channel or major tactic |
| Refresh speed | Daily | Monthly or quarterly |
| Best decision | Optimizing within a channel | Allocating budget across channels |
| Typical failure | Over-credits touches close to purchase | Can’t separate channels that move together |
MTA’s entry cost is mostly hygiene: consistent UTMs, source data captured on the CRM or order record, and server-side tracking where consent allows.
MMM has no official minimum. My working rule, not an industry standard: a model starts earning its cost when paid media runs in the mid six figures a year or more, across at least three channels, with about two years of weekly data and enough deliberate variation in spend (seasonal pushes, pauses, launches) for the model to read. B2B companies need more than that. If you create a handful of qualified opportunities a week and deals take months to close, the signal is thin and the lag between spend and outcome is long.
Which fits your stage
| Stage | Typical profile (working ranges) | Use | Skip for now |
|---|---|---|---|
| Early traction | One or two paid channels, under roughly $20k a month | Platform reporting, GA4, store or CRM totals, blended CAC, self-reported attribution | MMM, paid attribution software |
| Growth stage | Three or more channels, roughly $20k to $100k a month | All of the above, plus clean MTA in GA4 or the CRM, incrementality tests on the biggest channels, MMM-ready data collection | Full vendor MMM, unless offline media is a large share |
| Scale | $100k+ a month, often with TV, podcasts, out-of-home or retail | MMM calibrated with experiments, MTA for tactical optimization, self-reported attribution | Trusting any single source |
Two adjustments. Ecommerce brands with daily orders qualify for MMM earlier than B2B companies, because their weekly sales are less noisy and respond to spend in days, not months. And any company putting real money into channels MTA can’t see, such as podcasts, connected TV or sponsorships, needs MMM or geo experiments sooner than spend alone suggests.
Self-reported attribution as a third lens
The cheapest measurement method is asking: “How did you hear about us?” It captures what tracking misses, such as the podcast, the colleague’s recommendation, the LinkedIn post or the AI assistant answer, and it starts producing data the day you add it.
How to set it up so the data is usable:
- B2B: an open-text field on demo and signup forms. Open text surfaces sources you’d never think to list.
- Ecommerce: a post-purchase survey on the order confirmation page, with a short list of options plus “Other (please specify).” Randomize the order so the first option doesn’t win by position.
- Categorize monthly into a fixed taxonomy. An LLM can do the first pass on open-text answers; spot-check it.
- Store the answer on the contact or order record so you can join it to revenue and pipeline, not just count responses.
Know its limits. Memory favors whatever was recent or memorable, and “Google” often means “heard about you somewhere, then searched.” It can’t give you ROI on its own, so use it for direction, not bids.
The real value comes from comparing it with MTA. A hypothetical example with made-up numbers: a B2B SaaS company’s CRM credits 70% of pipeline to direct traffic and brand search. On the same deals, 40% of self-reported answers name the founder’s podcast appearances or LinkedIn posts. Neither source is wrong. The podcast created the demand and search captured it, and only one of those shows up in the click path.
Triangulation: a practical measurement stack
No method answers every question, so give each lens the decision it’s good at and a fixed cadence.
| Layer | Source | Cadence | Decision it drives |
|---|---|---|---|
| Business results | Store or CRM revenue, MER, blended CAC | Weekly | Is marketing working overall; total budget |
| Tactical attribution | Platform reporting, GA4 or CRM attribution | Daily to weekly | Which campaigns, ads and keywords to scale |
| Self-reported | “How did you hear about us?” answers | Monthly | Which channels create demand, including dark ones |
| Experiments | Incrementality tests | Quarterly | How far to trust each channel’s reported numbers |
| MMM (once you qualify) | Calibrated model | Quarterly or twice a year | Cross-channel allocation and diminishing returns |
When the lenses disagree, use rules set in advance:
- MTA high, self-reported low, channel sits close to purchase (retargeting, brand search): suspect over-crediting and test it before scaling.
- MTA low, self-reported high (podcasts, organic social, community): likely demand creation MTA can’t see. Protect the budget and test with a geo holdout or on/off test.
- All lenses agree: move budget with confidence.
- MMM and an experiment disagree: trust a well-run experiment and recalibrate the model.
The incrementality testing guide covers how to design those tests so they settle arguments instead of starting new ones.
Getting started
You don’t need a model to start building toward one.
- Pick one outcome metric from the store or CRM: net revenue, new customers or qualified pipeline
- Build a weekly table with spend by channel, the outcome metric and a notes column for promotions, launches, price changes and tracking outages; this is exactly the input a future MMM needs
- Fix UTMs and source capture so MTA data is clean enough to use
- Add “How did you hear about us?” and store the answer on the record
- Review the three lenses side by side once a month
- Run a first incrementality test on your largest or most doubted channel
- Vary spend on purpose, with planned pushes and pauses, so a future model has signal
- Revisit MMM once you have about two years of weekly data and three or more meaningful channels
When I set up measurement for growth-stage companies, the weekly table and the self-reported field come first, because they’re cheap and change budget conversations quickly. Wiring this stack together, from tracking to dashboards to calibration, is what my marketing attribution work covers.
Get it built
If your attribution reports and your instincts point in different directions, I can audit your measurement, set up the triangulated stack and tell you honestly whether an MMM is worth it yet. See pricing or get in touch.