MMM vs MTA: Which One a Lead Gen Team Actually Needs

If your question is the narrower one, whether attribution or a lift test should decide a budget, read incrementality vs attribution. If you want the definitions of lift and iROAS, the glossary entry on incrementality has them.

Both try to say how much of your result each channel deserves. They disagree about the unit of evidence: one looks at people, the other at periods.

• Multi-touch attribution (MTA)

  Records the touchpoints of each person who converts (ad clicks, organic visits, referrals) and splits the credit for their conversion across them by a model: first click, last click, linear, time decay, position based or data-driven.

• Marketing mix modeling (MMM)

  A statistical model fitted to aggregated time series: spend or exposure per channel per week, plus trend, seasonality, price and other controls, against a weekly outcome. It estimates each channel's contribution and its diminishing returns.

• Incrementality testing

  An experiment, not a model: one group or region sees the ads, a comparable one does not, and the difference is the effect. Used to check and calibrate both of the above.

The names cause some of the confusion. "Attribution" in MMM vendor copy often means MTA specifically, while the word itself covers any method of assigning credit, including the platforms' own reporting. In this guide MTA means multi-touch attribution on your own first-party click data joined to the CRM, which is a different thing from the totals in Ads Manager. The glossary entry on attribution models covers the rules for splitting credit.

The differences follow from the unit of evidence. Once you know a method looks at weeks rather than people, most of the rows below write themselves.

Read the last row twice. The two methods are not rivals for the same decision. A team that asks MMM which ad set to pause and MTA whether television works will be disappointed by both.

An MMM is a regression with two adjustments that make it fit how advertising behaves. The detail varies by model, but the shape is common to all of them.

• A baselineWhat the business would sell or book with no marketing at all: trend, seasonality, holidays, price changes, a competitor's launch. Everything not explained by the baseline is left for the channels.
• Adstock, or carry-overAn ad seen this week still has some effect next week. The model spreads each week's spend forward with a decay, so a burst of TV in March is allowed to keep working in April.
• Saturation, or diminishing returnsDoubling a channel's spend does not double its result. A response curve flattens as spend rises, which is what lets an MMM say where the next euro is worth least.
• Budget allocationWith a curve per channel, the model can propose a split of a fixed budget that it expects to return more. The proposal is only as good as the curves, and the curves only as good as the data.

The weakness sits in the middle of the method. A regression can only learn from what moved. Meta's Robyn guide gives the example directly: if sales vary week to week but TV spend stayed constant, the model has difficulty determining how TV affected sales. Channels that are always on, at a budget that only ever grows, are the hardest for an MMM to read.

MTA is the opposite trade. It gives you detail and speed, and it gives up anything that does not leave a click behind.

Done properly for lead generation, MTA records every visit with its UTM parameters and click IDs at the moment of the click, joins the anonymous visitor to a person when they fill in a form, call or book, and follows that person's record through the CRM stages to the closed deal. You can then ask which campaigns produced deals rather than form fills, what each keyword's leads were worth, and how the answer changes between a first click and a last click model. None of that is possible with weekly totals.

• It cannot see what was not clicked. A YouTube ad, a podcast read or a billboard that sent someone to search your brand shows up as brand search. MMM can credit it; MTA cannot. The click-through vs view-through guide covers why platform view-throughs do not fill the gap.
• It does not measure cause. A retargeting ad clicked by somebody who was already about to book gets credit for a deal it did not make. MTA tells you who was present at a conversion, not who changed it.
• It loses people who decline consent. A visitor who declines is not measured, and that share differs by country and audience.
• Platform MTA is not your MTA. Each ad platform's own attribution counts only its own touchpoints and claims every conversion it touched, so four dashboards sum to more than the CRM.

Two of the largest ad platforms publish free MMM code, which has made in-house modeling realistic for teams with an analyst. Both were read on their own pages on 7 October 2026.

Robyn describes itself as especially suitable for digital and direct response advertisers with rich data sources. Meridian encourages geo-level data where it exists but can run on national data, and can use reach and frequency. Google launched a partner programme of over 20 measurement firms alongside the general release, which says something about the effort: free code is not a free model. Somebody still has to collect the data, choose the priors and controls, and defend the output.

Both open source projects publish their own data guidance, and it is the most useful thing to read before buying any MMM. The numbers are stricter than most vendor pitches suggest.

• History. Robyn's analyst guide says an MMM needs a minimum of two years of weekly data, and four to five years if the data is monthly. Meridian's guide goes further for a national model: two years of weekly data, 104 points, gives four data points per parameter, which it calls too low to estimate reliably, and it prefers three years.
• Rows per variable. Robyn recommends about ten observations per independent variable. Every channel, control and seasonal term is a variable.
• Variation. Spend has to move, independently per channel, or the model cannot tell the channels apart.
• Geography helps. Meridian multiplies the data points by modeling regions, which is why it encourages geo-level data where it exists.

Now take an ordinary lead gen account. Five paid channels, two years of weekly data, plus trend, season and two controls such as a price change and the sales team's headcount: nine variables against 104 weeks, or 11.6 rows each. That passes Robyn's rule of thumb, barely, and fails Meridian's preference for a national model. Add a sixth channel or a third control and the margin is gone.

The outcome is the bigger problem. An e-commerce brand fits its model to thousands of orders a week. A lead gen business that cares about revenue fits it to won deals, and 18 won deals a week is a healthy B2B pipeline. If deals arrive more or less independently, chance alone moves a week of 18 by about four either way (the square root of 18 is 4.24), so a normal week runs from about 14 to 22 with nothing changed. A LinkedIn budget at 10% of a €40,000 month, about €920 a week, has to show its effect inside that noise.

• The sales cycle outlasts the adstockA deal that closes 90 days after the first click puts the effect of this week's spend three months downstream. Modeling on qualified leads is faster but measures a step that is not revenue.
• Always-on budgets do not varyMany lead gen accounts run every channel all year at a slowly rising budget. That is good management and bad data for a regression.
• Sales capacity is a hidden variableBooked calls depend on how many reps were free to take them. If headcount changed, it belongs in the model as a control, or the model will credit marketing with a hiring decision.

Incrementality is not a third rival. It is a measurement of one thing at one time, and both MMM and MTA get better when they are checked against it.

An incrementality test, a holdout or a geo experiment, answers one question with real causal evidence: would these conversions have happened without this channel, over these weeks, at this spend? It answers nothing else. It does not split a budget across eight channels or say which keyword works. MMM and MTA do those jobs, and both are estimates. The tests are what keep the estimates honest.

Robyn's own guide strongly recommends calibrating an MMM with experimental results it calls the ground truth, and running incrementality studies on a regular basis. Meridian takes experiment results as priors. Both projects treat an MMM without tests as a model you have no way of checking, and the same is true of MTA.

Calibration means using the test's answer to correct the model's. The arithmetic for MTA is simple enough to do on a napkin.

The ratio is specific to one channel, one period and one setup. Retargeting and brand search usually come out well below 1, because they reach people who were already on their way; a prospecting channel can come out near or above it, because it starts journeys that MTA credits elsewhere. Retest when the budget, the audience or the creative changes a lot.

For an MMM the test result goes in as a prior or a constraint: the model is told that this channel's effect over those weeks was about this much, and fits everything else around it. Both open source projects document how. For running the test itself on CRM deals, incrementality testing for lead gen walks through a geo holdout step by step, and the incrementality calculator does the lift arithmetic.

The budget bands below are our rule of thumb, not a published threshold. What matters more than spend is the shape of the account: how many channels, whether any are offline, and how many outcomes a week the model would have to learn from.

"Unified marketing measurement" is the name vendors give to running all three methods and reconciling them. The idea is right; the name is often a product.

Triangulation means letting each method do what it is good at and using the others as a check. MTA steers the daily work. The MMM sets the yearly split. Tests settle the arguments where the two disagree, and their results flow back as calibration into both. When MTA says retargeting is your best channel and the MMM says it adds little, that disagreement is not a failure: it is the cue for a holdout.

• One outcome definition for all three. If MTA reads won deals, the MMM qualified leads and the test form fills, nothing reconciles.
• The same channel names. A channel taxonomy agreed once, so that "paid social" means the same spend everywhere.
• A test calendar. Decide which channel gets tested each quarter, so calibration is a habit rather than a reaction.

LeadJourney sits on the MTA side of this guide, built for lead generation. One script on the site or in Google Tag Manager records every visit server-side on your own domain, first-party, at 95%+ tracking accuracy, and captures the click IDs (gclid, gbraid, wbraid, fbclid, li_fat_id, msclkid) at the click. The anonymous history is joined to the person at the form, the call or the booking, and native CRM integrations (HubSpot, Salesforce, Pipedrive, Zoho and more) write the source, campaign and journey onto the record and read the stage changes and won deals back. Spend from the ad platforms is joined to leads, deals and revenue per channel, campaign and ad, with first click, last click, linear, time decay and position based switchable on the same events. Setup takes about 21 minutes.

What it does not do, plainly: LeadJourney does not build marketing mix models, run incrementality tests or geo experiments, or count view-throughs, and a visitor who declines consent is not measured. Where it helps the other two methods is the outcome. A lift test or an MMM fitted to form fills measures the wrong thing for lead gen; won deals per campaign and channel, joined to the CRM, are the outcome both should read. And because won deals go back to Meta (Conversions API), Google Ads (offline conversion import), LinkedIn (Conversions API) and Microsoft Ads with the deal value, a platform's own lift study can be set up on deals rather than leads where the study type accepts offline events: check that on the platform before you start. For a DIY geo test, export deals with their source and the customer's region from the CRM.

1. Buying an MMM to fix broken tracking. If deals in the CRM carry no source, the MMM has a weekly total to model and you still cannot say which campaign worked.
2. Reading platform dashboards as MTA. Four platforms each claiming their own conversions is not multi-touch attribution; it is four single-touch reports that overlap.
3. Modeling on form fills. A channel that brings cheap leads that never close looks best on leads and worst on revenue. Pick the outcome the business is paid for.
4. Trusting either without a test. MTA over-credits channels close to the conversion, and an MMM can attribute a season to a channel that happened to rise with it. Tests are the check on both.
5. Expecting one number. The methods answer different questions at different speeds. Agree in advance which one decides which kind of budget move.

The rest of this cluster: incrementality for the definitions, incrementality vs attribution for the argument between the two, incrementality testing for lead gen for running a geo holdout on CRM deals, Meta incremental attribution for Meta's setting, and the incrementality calculator. On attribution: attribution models, multi-touch attribution, B2B marketing attribution, cross-channel attribution and click-through vs view-through conversions. For sending deals back to the platforms: offline conversion tracking. Tools that offer modeling next to attribution are compared in LeadJourney vs Rockerbox and LeadJourney vs Northbeam.