In short
- Mix modelling needs about two years of weekly data, several channels and creator activity that varies by week. Without that, use tracking, a survey and a holdout.
- Creator spend is a small, bursty line that is hard to separate from festival demand. In our simulation, festival-only spend widened the estimate's range by about two thirds.
- Calibrate the model with a holdout result and read intervals, not point estimates. An interval that spans break-even is not yet a decision.
- Move budget on a model's advice in steps, our working rule being 10 to 20%, then verify the move with a test.
Marketing mix modelling, or MMM, estimates how much each channel contributed to sales by fitting a statistical model to your weekly spend and sales history, with seasonality, price and promotions as controls. For influencer marketing it works only when you have long enough data, several channels and creator activity that varies from week to week. Otherwise it returns intervals too wide to act on.
Our guide to influencer marketing attribution lists mix modelling as one method among eight and rates its effort "very high". This post goes deeper on one question: can a mix model read your creator programme, and if so, how do you feed it and read it? It builds on influencer marketing incrementality testing, which provides the experiments that keep a model honest, and on the definitions in how to measure influencer marketing ROI. The wider plan is in the complete guide to influencer marketing in India. The scale question matters: in a 2024 industry report covered by Social Samosa, 56% of brands spent more than 2% of budget on influencers, which can still be a small line beside paid social and search.
What a mix model can and cannot tell you
| Question | Links and codes | Holdout test | Mix model |
|---|---|---|---|
| Which creator drove this order? | Partly | No | No |
| How many sales did creators cause in total? | Floor only | Yes, for the test period | An estimate with an interval |
| Does the effect carry into later weeks? | Hard to see | Only if the test has a tail | Yes, through carry-over terms |
| Where does the next ₹1 lakh do most good? | No | One channel only | Yes, through response curves, with caution |
| Did creators lift other channels? | No | Yes | Partly |
A model never replaces a creator-level view. It tells you about the programme, not about creator number 7.
Readiness checklist
These are our working rules, not published standards, and we found no India-specific threshold for any of them.
| Condition | Why it matters | Working rule |
|---|---|---|
| History | The model must see seasons more than once | About two years of weekly data |
| Variation | A line that never moves cannot be measured | Creator activity changes week to week, including some weeks outside festivals |
| Channels | Several channels with their own histories let the model separate them | Three or more |
| Stakes | The model costs analyst time | A budget move large enough that a 10% error matters |
| Clean data | Missing promotions pollute estimates | Weekly sales, a price and promotion log, stock-out days |
| An analyst | Someone must explain it to finance | Named, with time set aside |
If two or more are missing, use tracking, a survey and a holdout, and revisit in a year.
Feeding creator activity into the model
| Variable | How to build it | Pitfall |
|---|---|---|
| Creator fees | By the week each post went live, not the invoice date | Advance payments and barter value hide timing |
| Creator-ad media | A separate paid channel, by delivery week | Merging it with fees double counts the activity |
| Creator views | Optional second measure when fees differ a lot from reach | Views come from creator screenshots and arrive late |
| Carry-over | The share of an effect still working each following week | YouTube tends to carry longer than Stories; fit it, do not guess |
| Saturation | A curve where each extra rupee earns less | Needs weeks of high spend to estimate |
| Controls | Festivals, sale events, price, discount depth, stock-outs, distribution changes | An unlogged promotion is credited to a channel |
| Other channels | Paid social, search, video, out-of-home, email, affiliates | Missing a channel moves its sales into the others |
A weekly layout looks like this. The rows are invented.
| Week | Net sales (₹ lakh) | Paid social | Search | Creator fees | Creator-ad media | Sale event | Average discount | Stock-out days |
|---|---|---|---|---|---|---|---|---|
| Week 40 | 182 | 10.4 | 6.1 | 0.0 | 0.0 | No | 8% | 0 |
| Week 41 | 204 | 11.0 | 6.3 | 1.8 | 0.6 | No | 8% | 0 |
| Week 42 | 251 | 13.2 | 7.0 | 3.1 | 1.4 | Yes | 18% | 1 |
Why a small, bursty creator line is hard to read
Two things go wrong. A small line has a small signal against weekly noise. And creator work is often timed to festivals, so the model has to separate a creator effect from a festival effect that moves at the same time.
We ran a toy simulation to show the pattern. It had 104 weeks of sales, a true creator return of ₹2 of sales per ₹1 spent, landing in the week of spend, fixed true returns and carry-over for two other channels, weekly noise of about 8% and a festival effect that was larger in one of the two years. The model used a single festival control. We ran each case 600 times and recorded how far the estimated creator return strayed from the true 2.0. It is a pattern, not India data.
| Creator spend as a share of media | Always-on, varying week to week: 5th to 95th percentile (median) | All spend in festival weeks: 5th to 95th percentile (median) |
|---|---|---|
| 3% | −8.7 to 13.2 (2.0) | −17.1 to 19.0 (2.0) |
| 6% | −3.7 to 7.5 (2.2) | −7.4 to 12.0 (1.8) |
| 11% | −0.8 to 4.8 (2.0) | −2.3 to 6.4 (2.1) |
Read it in two steps. First, the estimate is right on average and very uncertain: at a 3% share it could land anywhere from a large loss to six times the truth, and even at 11% the range runs from below break-even to more than double. Second, putting all creator spend into festival weeks made it worse, with ranges about two thirds wider, because creator spend and festival demand move together and the model has to guess which is which. In a variant where creator effects carried over into later weeks, the festival-timed estimates were also pulled well above the truth: the model gave the creators credit for the festival.
The practical lesson is to design creator activity so a model can read it later. Include some waves outside festival weeks, vary timing and intensity between regions where you can, and log everything else that changes. Our guide to always-on versus campaign influencer marketing covers the rhythm trade-offs.
Plan creator activity so a model can read it later
You can act on this before you ever build a model. These choices cost little and make a future model far more useful.
| Choice | Why it helps |
|---|---|
| Keep a low level of weekly creator activity all year | Gives the model many small movements to learn from, not three large ones |
| Stagger launches by region, with some regions starting a month later | Creates natural comparison groups for both the model and tests |
| Plan one pause of three to four weeks outside festival season | Shows what sales do without creators when nothing else is moving |
| Keep a dated log of every price change, sale and stock-out | The model can only control for what you recorded |
| Keep post dates, fees and creator-ad media in one sheet by week | Saves weeks of data cleaning later |
Read the output with intervals
A hypothetical output for a brand whose break-even return is 1.5, which a contribution margin of about two thirds produces:
| Channel | Share of media spend | Return per ₹1 (median) | 90% interval | Reading |
|---|---|---|---|---|
| Paid social | 55% | 1.9 | 1.5 to 2.4 | Just above break-even |
| Search | 30% | 2.8 | 2.0 to 3.7 | Clearly above |
| Marketplace ads | 9% | 1.4 | 0.9 to 1.9 | Straddles break-even |
| Creators | 6% | 2.1 | 0.4 to 3.9 | Straddles break-even, very wide |
| Creators, after calibration | 6% | 1.8 | 1.3 to 2.4 | Narrower, above break-even |
The creator line alone says very little: an interval from 0.4 to 3.9 includes a heavy loss and a large gain. The last row shows calibration. A holdout test measured creators at a return of about 1.7 in the test period, and the model was told so. Open-source MMM libraries, including Meta's Robyn and Google's Meridian, accept experiment results for this purpose. The calibrated line is narrower and agrees with the test. Our guide to incrementality testing shows how to design the test that gives the number.
Check the model before you trust it
| Check | How | Red flag |
|---|---|---|
| Out-of-sample forecast | Hold back the last 8 to 12 weeks, refit and predict them | Error far above normal week-to-week noise |
| Stability | Drop different blocks of weeks and refit | The creator return jumps between very different values |
| Sign and size | Look at each channel's estimate | Negative creator effect, or an implausibly large return |
| Sensitivity | Change the carry-over assumptions | The conclusion flips |
| Agreement with tests | Compare with holdout results | The model sits outside the test's interval |
Use the result in steps
Move budget on a model's advice in small steps, our working rule being 10 to 20% of the line, then run a test on the move. Refresh the model each quarter and keep a log of what changed, so a shift in return can be traced to a change in the data and not only to the market. Do not cut creators because a single run puts the low end of the interval below break-even. Check it against the calibration and the creator-level scorecard first, and our post on forecasting influencer campaign results shows how to feed the result into the next plan.
Build, buy or skip
| Option | Fits | Watch |
|---|---|---|
| Skip for now | Short history, or creators well under about 5% of media | Use tracking, a survey and one holdout |
| A regression of weekly sales on creator views, with controls | A directional read on one outcome | Do not call it a mix model |
| Open-source library with an analyst | A brand with a data team | Analyst time and governance |
| Vendor or consultancy | A brand without a team | Ask the questions below |
Questions for any vendor: Will you share the code and the data pipeline? How are priors set, and can you calibrate with our tests? Do you report intervals? What is the out-of-sample error? How often is the model refreshed? How do you separate festival bursts from creator effects? We found no India price data, so compare the cost with the budget it could move.
Mistakes to avoid
- Modelling with six months of data.
- Reading the creator line as a precise number.
- Timing every creator wave to the same festival and then trusting the model to separate them.
- Merging creator fees and creator-ad media.
- Leaving promotions and stock-outs out of the controls.
- Cutting creator budgets on one uncalibrated run.
- Treating the model as a creator ranking.
If you want to plan creator activity so that a future model can read it, send us a brief. Our influencer marketing strategy service covers the audit and the measurement plan, and the rates guide covers the fees that sit in the spend line. For the cheaper signals that can sit beside a model, see influencer impact on branded search.
Sources
- The State of Influencer Marketing in India: share of brands spending more than 2% of budget on influencers (2024 report; older data), Social Samosa (EY, Collective Artists Network and Big Bang Social report) (2 Apr 2024)
- Coverage of a 2026 India creator-economy report: brand survey on revenue targets and campaign-by-campaign judging, Campaign India (21 May 2026)
Frequently asked questions
What is marketing mix modelling for influencer marketing?
It is a statistical model that splits weekly sales across channels, baseline demand, price, promotions and seasonality, using your spend and sales history. For influencers, it estimates the contribution of creator fees and creator ads over time. It gives an estimate with an interval, not an order-by-order count.
How much data do I need for mix modelling?
Common practitioner guidance is about two years of weekly data, roughly 104 weeks, but that is a rule of thumb, not a standard. Variation matters as much as length: if creator spend only appears in festival bursts, the model cannot separate it from festival demand however many weeks you have.
Can mix modelling measure influencer marketing?
Yes, when creator activity is large enough and varies across weeks and regions. When it is a small, bursty share of media, the interval around its return is usually too wide to act on. Pair it with a holdout test, as our incrementality testing guide explains.
What is the difference between mix modelling and incrementality testing?
A test withholds creators from a group and measures one cause directly. A mix model works from history across all channels and cannot be certain about any one. Tests are sharper but narrower. Use test results to calibrate the model, then use the model to see the whole portfolio.
How much does a marketing mix model cost?
We found no published India price data. The cost is analyst time if you build in-house, or a vendor or consultancy fee. Compare it with the size of the budget the model could reallocate, and skip it if the likely reallocation is small.
Should a small brand build a mix model?
Usually not yet. Without long history, several channels and enough creator spend to detect, the model returns wide intervals. Start with links, codes, a survey and one holdout, as in our attribution guide, and revisit when spend and history have grown.