How to Put Influencer and Affiliate Spend in a Marketing Mix Model
Affiliate commissions track sales and creator fees miss posting weeks. Which input to use for each in Meridian and Robyn, and the lift test that calibrates it.
On this page
- Why commissions make a bad spend column
- Why influencer fees land in the wrong weeks
- Which input each activity should be
- Discount code sales are attributed orders
- How to measure incremental lift for each channel
- A worked example with invented numbers
- Feeding the results into Robyn and Meridian
- Where we land
- Sources
To include affiliate and influencer marketing in MMM, give the model what each channel put in front of buyers and keep what it billed you as a separate number. An affiliate commission is earned on a sale, so its cost series follows the KPI it is meant to explain. A creator fee is often invoiced on a contract date, weeks away from the posts that do the work. Our position: model creators on views by posting date and affiliates on clicks or placements, attach fees in a form the engine can spread over that exposure, and calibrate both with a holdout, because discount-code sales are attributed orders.
Why commissions make a bad spend column
An affiliate commission only exists after a purchase. The Associates help center from Amazon says a referral commission is not credited until the customer has purchased the item, accepted delivery and paid in full. When the commission is a percentage of order value, the weekly commission line is a scaled copy of affiliate-attributed revenue, and a regression reading it as spend sees cost rise in every strong week, whatever made the week strong.
Google's Meridian data guide on Google for Developers lets you use spend as a proxy for exposure when impressions or clicks are missing, and warns: "If the model assumes spend represents physical exposures, periods of high ad cost might be interpreted as high ad volume." With commissions, high-cost periods are high-sales periods by construction.
Timing adds a second error. On impact.com, the action locking period runs from when an action is tracked until it locks and can no longer be updated, up to 90 days when set in days alone, and a payment scheduling period follows before the partner is paid. Commissions pulled from accounts payable can land weeks after the orders behind them, so key any commission series to the action date.
Why influencer fees land in the wrong weeks
A flat-fee creator deal is billed on a contract schedule while its effect arrives on posting dates. Three posts in a year can leave one invoice spike in a week when nothing ran and zero spend in the three weeks that mattered.
Meridian has a tidy fix. Its InputData reference on Google for Developers accepts media_spend as one total per channel, with no geo or time dimension, and allocates it across geos and times in proportion to media. Feed post views by posting week as media and the contract total as spend, and cost follows exposure. Three limits apply. The total form covers the whole media_spend array, so every channel in it goes in as a total or none does. The same reference requires a spend breakdown by geo and time period if roi_calibration_period is specified, and if views arrive as one national number, a geo model needs them imputed to geos, which the Meridian data guide recommends for national-only channels.
Robyn wants spend in every period, so do the allocation by hand: spread each fee across its posting weeks, weighted by views. Google's data guide also advises combining low-spend channels with other channels "to avoid issues with ROI estimation," so five small creator deals usually belong in one creator channel.
Which input each activity should be
| Activity | Meridian input | Robyn input | Exposure series | Lift test that calibrates it |
|---|---|---|---|---|
| Paid creator posts (flat fee) | Paid media, fees as one aggregate media_spend total | paid_media_vars plus paid_media_spends spread over posting weeks | Views or impressions by posting date | Same creative run as ads in test regions, matched regions held out |
| Gifted or unpaid creator posts | Organic media, contribution prior | organic_vars | Views or reach by posting date | Rarely testable alone; borrow the paid creator result as a wide prior |
| Content affiliates (reviews, guides) | Paid media, commissions by action date as aggregate spend | Clicks in paid_media_vars, commissions in paid_media_spends | Referred clicks or placement impressions | Pause the publisher group for matched weeks or regions |
| Coupon, cashback and loyalty partners | Separate paid channel with a tight prior from a test | Separate spend column, calibrated | Referred clicks | Pause the group and read total sales against a matched control |
| Creator and affiliate discount codes | Non-media treatment: discount depth by week | context_vars: discount % by week | Discount depth, with redemptions left out | Read through the test for the channel the code rides on |
The input types rest on the documentation. The test designs and the affiliate split are our own. Meridian's page on organic and non-media variables at Google for Developers says "Organic media variables are media activities that have no direct cost," gives them Adstock and Hill effects, and states that "results pertaining to ROI, such as response curves and budget optimization, are not available for organic media variables." That fits seeding, where no fee exists to divide by. The same page lists running a promotion as a non-media treatment, with the variable's minimum as the default counterfactual.
Robyn's Analyst's Guide from Meta defines organic variables as "Any marketing activities without a clear marketing spend," lists social media posts among them, and applies similar transformation techniques to them as to paid media. Promotions sit among context variables, which "will not undergo any transformation techniques," and the guide prefers each promotion's discount % as continuous data over a dummy. Our guide to MMM control variables covers how promotion variables protect media credit.
The affiliate split is our call. In our experience coupon and cashback clicks cluster at checkout, after the decision, while content publishers reach people earlier, and one column would average those two jobs into a single curve. Robyn's guide says to avoid clicks because they miss view-through conversions. A text link's click may be the only exposure available, so accept it and keep that channel's ROI provisional until a test confirms it.
Discount code sales are attributed orders
Incorporating influencer marketing and affiliate channels into an MMM for incrementality means retiring code redemptions as proof of lift. A code ties an order to a creator or partner and stops there. A buyer who had already decided can find the code at checkout, a creator's code can spread to coupon sites, and a loyal customer who happens to follow the creator gets counted as the creator's sale. Redemptions are attribution, and useful as attribution: reconcile them with a test through a calibration multiplier. Keep them out of the model as a KPI or a channel variable, since they are an outcome the channel claims for itself. That gap between credited and caused is what incrementality measures.
How to measure incremental lift for each channel
To measure influencer and affiliate incremental lift for an MMM, you need people who could have been exposed and weren't, which these two channels make harder than paid social does. A creator's post reaches their following wherever it lives, so the practical test moves the content to where you control delivery: run the same creative as ads from your own account in test regions and hold out matched regions, following the design of a geo holdout. That measures the creative under paid delivery, a close cousin of the organic post, so carry it into the model as a prior with honest width.
For affiliates, test by partner group. Pause the coupon and cashback group for a run of weeks against matched weeks, or in held-out regions where your program can enforce it, and read total sales across every channel. Then repeat for content publishers. Robyn's demo script on GitHub names the usual experiment types as identity-based (for example Facebook conversion lift) or geo-based (for example Facebook GeoLift).
A worked example with invented numbers
Measuring incremental lift for influencer and affiliate spend in a marketing mix model starts with putting each cost in the right week. The figures below are hypothetical, for illustration only.
A skincare brand pays three creators $8,000 each, invoiced together in week 2. The posts run in weeks 6, 7 and 19 and draw 400,000, 250,000 and 600,000 views. An invoice-based series puts $24,000 in week 2 and nothing anywhere else. Passed to Meridian as one $24,000 total against the views series, the spend lands in proportion to views.
| Week | Views | Allocated spend |
|---|---|---|
| 6 | 400,000 | $7,680 |
| 7 | 250,000 | $4,800 |
| 19 | 600,000 | $11,520 |
| Total | 1,250,000 | $24,000 |
The same brand pays affiliates 10% of attributed sales. In a normal week, partners are credited with $50,000 of sales and earn $5,000. In a TV flight week, coupon partners are credited with $90,000 and earn $9,000. Commission rose 80% while the partners did nothing different, and a model fed that line as spend would hand them credit that mostly belongs to TV.
Feeding the results into Robyn and Meridian
Marketing mix modeling with influencer and affiliate incrementality from causal lift experiments, geo experiments included, can feed Robyn and Meridian, each through its own input.
In Robyn, the demo script says "Calibration channels need to be paid_media_vars or organic_vars names," and its sample calibration input includes an organic newsletter with zero spend, so an unpaid creator channel can be calibrated too. For causal experiments it says to use calibration_scope = "immediate", Robyn currently accepts only a point estimate, and "The point-estimate has to always match the spend in the variable." The script also asks for experiment spend that matches the input data within 10% over the test dates, one more reason to put fees in posting weeks.
In Meridian, the prior-type table on Google for Developers allows ROI and mROI priors for paid media only; organic media and non-media treatments take contribution or coefficient priors, with contribution as the default. A contribution prior is "The proportion of the total outcome that is attributable to a channel," and the default for organic media is a Beta distribution with a mean of 1%. A seeding result therefore enters as a share of total outcome over the modeling window, and a paid creator or affiliate result enters as an ROI prior. Our walkthrough on how to calibrate a marketing mix model with geo lift results covers matching the KPI, channel scope and time window.
Where we land
Count creators by what their audiences saw, count affiliates by the clicks and placements they delivered, treat codes as promotions, and give no channel's ROI weight in a budget decision until a holdout has set its prior. The coupon tier is the one we would test first, since its cost tracks sales most closely. Our data and analytics team is a further resource for a model spec review, and the Attribution Doctor runs a lightweight media mix model in your browser from monthly spend and revenue by channel.
Sources
- Amazon: Associates Central help, Shopping Cart window and referral commission crediting
- impact.com: Set action locking and payment scheduling periods
- Google for Developers: Meridian data collection guide
- Google for Developers: Meridian InputData API reference
- Google for Developers: Organic media and non-media treatment variables (Meridian)
- Google for Developers: Choose and configure treatment prior types (Meridian)
- Meta: An Analyst's Guide to MMM (Robyn)
- GitHub: Robyn demo script, demo.R (facebookexperimental/Robyn)
