Media-Mix Modelling for Earned Links and Digital PR

TL;DR. Media-mix modelling (MMM) is the one measurement method that can finally put earned links and digital PR on the same budget chart as paid media — a top-down, privacy-safe model that estimates each input’s contribution to a business outcome, with no cookies required. But run it naively and it will tell you PR is worthless. The three things MMM handles worst — slow compounding effects, effects that work through mediators like brand search and organic, and collinearity with the “free” organic baseline — are precisely the three things that define earned media. Its credit leaks into base sales and into whatever paid channel it happens to correlate with. The fix is a three-part adaptation (the right input variable, a long carryover, and a nested structure that models PR as a driver of the mediators) plus one non-negotiable step: calibrate the model with a geo holdout, because MMM alone cannot untangle links, PR and organic. Run the Readiness Gate first; without two to three years of varied weekly data and a calibration experiment planned, do not build.

The seat at the budget table earned media never had

Every finance team has seen the same slide. Google Analytics says paid search drove 40% of last quarter’s revenue, Meta’s dashboard claims social drove 35%, and the brand team is waving a lift study worth another 30% — which adds up to 105% of revenue before anyone has mentioned organic, email, or the two-week promotion in March. When every platform grades its own homework, the numbers never reconcile. Digital PR and earned links fare worst of all in that world: they have no dashboard of their own, so their contribution is quietly absorbed by whichever channel was running when the traffic arrived.

Media-mix modelling takes the opposite, top-down view. Instead of tracking individuals, it regresses a business outcome — revenue, conversions, or organic traffic — against every marketing input across two to three years of history, plus controls for seasonality, price and promotions, and reads each input’s incremental contribution. It splits your results into a base (what you would have earned with no marketing at all — loyalty, brand equity, standing organic demand) and the incremental portion each channel actually caused. Because it needs no user-level tracking, it survived the privacy era intact, which is why interest has surged: EMARKETER and TransUnion put the share of marketers planning to increase MMM investment at roughly 47% heading into 2026.

For link builders the appeal is specific. A geo holdout answers “did this one regional campaign cause a lift?” and a page-level model answers “what will this single link do?” — the kind of question we tackle in predicting ranking impact from a single backlink. MMM answers the portfolio question the other two cannot: across everything you do, how much of the business outcome did earned links and digital PR contribute, and what should that be worth next quarter? Get it right and PR finally competes for budget on the same evidence as paid. Get it wrong — and naive MMM gets earned media wrong by default — and you hand the case for defunding your best long-term channel to your own model.

One number MMM surfaces is worth the whole exercise on its own: the gap between what a platform claims and what it actually caused. Platform-reported returns are almost universally higher than the incremental returns a model finds, because platforms take credit for conversions that would have happened anyway. Earned media is the mirror image — it has no platform inflating its numbers, so it is chronically under-credited while paid is over-credited. That asymmetry is precisely why digital PR loses budget arguments it should win: the channels with a dashboard shout, and the channel without one stays silent. MMM is the only method that can let earned media speak in the same units.

Start here: the MMM Readiness Gate

MMM has hard prerequisites, and earned media strains every one of them. Before you commission a model or open Robyn, run your situation through these five checks. If you fail one, the honest move is to fix the data or reach for an experiment instead — not to build a confident model on foundations that cannot support the earned-media question you actually care about.

Check five is the one teams skip and the one that decides whether your earned-media read is trustworthy. The next section explains why the model cannot be trusted on PR without it.

Check two deserves a word of its own, because it is the quiet killer of earned-media MMM. A model learns a channel’s effect from how the outcome responds when the channel changes. A digital PR programme that runs at a steady hum every week gives the model almost nothing to learn from — there is no “off” period to compare against — so the coefficient becomes unidentifiable and defaults to whatever the priors or the correlations suggest. Paradoxically, the more consistently you run PR, the harder it is to measure. The fixes are deliberate: introduce real variation by concentrating activity into distinct waves rather than a flat drip, exploit the natural spikes that digital PR and newsjacking already produce around news cycles, or supply the missing information from outside the data with an experiment. A flat line cannot be modelled, only assumed.

Why naive MMM under-credits links and PR

This is the section to slow down for, because it is the reason so many teams conclude “the model says PR does nothing” and quietly cut the budget. Three structural features of MMM collide with three defining features of earned media, and every collision pushes the estimate the same way: down.

1. Earned media is slow; MMM rewards the fast

A backlink keeps passing authority for months or years, and press coverage compounds long after the news cycle moves on. That is the carryover (or adstock) effect, and it is longer for earned links than for anything else in the mix. A model that ignores carryover counts only the conversions that land in the same week as the activity and attributes the later ones to whatever channel was running when they arrived — usually paid search. The literature is blunt about the size of this error: accounting for a four-week carryover can roughly double a channel’s measured ROI. For earned media, whose tail runs for months, an unmodelled or paid-length carryover does not shave the estimate — it guts it.

2. Earned media works through mediators; additive MMM collapses them

PR and links rarely convert anyone directly. They work through mediators: coverage lifts brand awareness, which raises branded search; links raise authority, which raises organic rankings; both then feed conversions. A standard additive model has no place for that chain — it treats every input as a direct, independent driver of the outcome. So the credit that should flow “PR → brand search → conversions” gets assigned to brand search, and the credit for “links → organic → conversions” gets assigned to organic. The mediator takes the bow; the earned media that created the mediator goes uncredited. This is the same trap MMM practitioners describe when demand drives paid search, which is why sophisticated models add an indirect or “model-within-a-model” layer — the fix we build in Section 5.

A concrete version makes the loss obvious. A brand lands a national data story; over the next fortnight, searches for the brand name climb, and those branded searches convert at four or five times the rate of cold traffic. A naive model sees branded search spike and conversions follow, so it credits branded search — a channel that costs almost nothing and looks gloriously efficient — while the digital PR campaign that manufactured the entire spike is recorded as having done little. The brand-search team gets a bonus; the PR team gets a budget cut. Nothing in the data was wrong; the model simply attributed the effect to the last mediator in the chain instead of the earned media that set it moving.

3. Earned media is collinear with the baseline; the model calls it “free”

Here is the deepest problem. PR, links and organic demand all tend to move together — they rise when the brand is healthy and fall when it is not. When two inputs are highly correlated (above roughly 0.8), the model cannot separate their individual effects, and adding more history does not help; it reinforces the ambiguity. Regularisation — the ridge penalty in Robyn, priors in a Bayesian model — does not resolve this. It just picks one plausible attribution from many equally plausible ones and reports it with a straight face. The related failure, baseline leakage, is where the model cannot cleanly separate what marketing drove from what was already happening — so slow, always-on earned media gets swept into the “base” and labelled free demand.

The net effect, stated plainly. Run a naive MMM and earned media’s value is split three ways: some is counted early and lost to a too-short carryover, some is handed to the mediators (brand search, organic) it actually created, and the rest is absorbed into the baseline as “free” demand. What is left — the number the model reports for PR — is a fraction of the truth. The model is not lying; it is confidently answering a question it cannot identify. A PR contribution near zero is almost never evidence that PR failed. It is evidence the model was not built to see it.

The scale of what hides in the baseline is easy to underestimate. In a representative decomposition, a DTC brand found that 55% of revenue was base sales — loyalty, organic search and word of mouth — with only 45% ruled marketing-driven. Earned media’s fingerprints are all over that 55%, and a model that treats the base as untouchable demand will never give links or PR a penny of it. The rest of this guide is about clawing that credit back honestly.

There is a quick diagnostic for whether your base is hiding earned media. A healthy base should be roughly flat or slowly trending — it represents demand that exists regardless of this quarter’s marketing. If your modelled base instead rises in step with your PR and link activity, that co-movement is the tell: real incremental effect has been misfiled as baseline demand. Plot the base against your earned-activity input and look for correlation. When the two track together, you have not discovered unusually loyal customers; you have found the credit your model is refusing to assign to the campaigns that earned it. That single chart is often the most persuasive thing you can put in front of a finance team, because it shows the leak rather than merely asserting it.

Feeding the model: input variables and carryover

MMM was built for channels measured in spend, and you can dial spend up and down. Earned links and digital PR are not a spend line you control linearly, so the first design decision is what input variable represents them. Pick the highest rung of this ladder your data supports — higher rungs capture more of what actually varies week to week.

Whichever rung you choose, the input is a rate of earned activity over time — which is exactly link velocity expressed as a model variable. That framing also flags a hazard: if your earned activity is a near-flat line, none of these variables can be identified, no matter how good the rest of the model is. Variation is the fuel; a channel that never changes teaches the model nothing.

Carryover: give earned media the longest half-life in the model

Carryover is expressed as a half-life — the number of periods it takes for an effect to fall to half its initial value. Paid channels decay fast; earned media decays slowest of all. Setting this parameter too short is the single most common way an MMM buries link and PR value.

The arithmetic shows the stakes. Suppose a PR spike drives 1,000 conversions in its launch week and carries over at a one-week half-life: 500 the next week, then 250, then 125. A naive model books 1,000; the true four-week total is 1,875 — the measured return nearly doubles once carryover is respected, and that is with a short half-life. Give earned media the months-long half-life it deserves and the gap is far larger. This is the modelling counterpart of a fact link builders already know from the compounding nature of authority: the returns arrive slowly and keep arriving, which is exactly what a short carryover throws away. Where the data cannot pin the decay rate — common when earned activity has little variation — set it from a long-carryover prior rather than letting the model default to a paid-style value.

One more transformation matters: saturation, the diminishing return as activity scales. The tenth tier-one placement in a month moves the needle less than the first, and a model that assumes a straight line will either over-promise the next unit of PR or cap it too early. For earned media the saturation curve is usually gentler than paid — authority compounds rather than fatiguing as fast as an ad audience — but it is still there, and it is what turns the model from a rear-view report into a forward-looking planning tool that can answer “what happens if we double the PR programme?” Carryover and saturation together are what separate a real MMM from a naive regression, and both need to be set with earned media’s slow, compounding character in mind rather than borrowed from paid defaults.

The nested structure: model PR as a driver of the mediators

The mediation problem from Section 3 has a specific fix, and it is the difference between an MMM that sees earned media and one that does not. Instead of forcing PR and links to explain conversions directly — a path they mostly do not travel — you model them as drivers of the intermediate variables they genuinely move, and let those variables drive the outcome. This is the “model within a model” that mature MMM practice uses whenever an upstream activity works through a downstream one.

Concretely, earned media reaches conversions along three paths, and a good model keeps them distinct:

1. The direct-referral path. Coverage and links send referral traffic straight to the site. This is the fast, visible sliver — and the only path a naive model tends to catch.
2. The brand-search path. PR raises awareness, which raises branded search, which converts at a high rate. Model PR as a driver of a brand-search index, and that index as a driver of conversions — so the credit flows back to the coverage that started it.
3. The organic-authority path. Links raise domain authority, which raises organic rankings and non-branded clicks. Model referring domains as a driver of organic performance, not of conversions directly — the same reason a page-level link’s job is to move rankings first.

Structuring the model this way does two things at once. It stops the mediators from stealing earned media’s credit, and it makes the collinearity more tractable, because PR is no longer competing head-to-head with organic to explain the same conversions — it sits one level upstream, explaining the organic and brand signals themselves. It also reframes what you are measuring, which connects to the wider problem of measuring the authority that actually decides outcomes when the classic backlink metrics no longer tell the whole story. This structure is also the honest embodiment of the “multiplier effect” communications teams talk about: earned media does not just add a channel, it lifts the productivity of the others, and a nested model is where that shows up as brand and organic strength rather than a single direct coefficient.

That multiplier is not marketing folklore; it is the finding communications researchers have been pressing for years — that earned, owned and creator activity directly shapes awareness, perception and trust, all of which correlate with future sales, and that most brands fail to capture it because their measurement models are built around paid media. Advanced comms-driven models exist specifically to quantify how earned coverage augments the impact of paid rather than sitting beside it. The practical upshot for a link-and-PR programme is that a nested MMM will often reveal earned media doing two jobs: a modest direct contribution and a larger, easily-missed role making every other channel work harder. Cut it on the direct number alone and you also lose the multiplier you never measured.

The non-negotiable step: calibrate with an experiment

Everything so far improves the model. None of it identifies the earned-media effect, because the PR–links–organic collinearity is a property of the data, not of the model, and no amount of clever structure fully resolves it. The model will still choose one plausible attribution from many. The only way to pin it to the truth is to bring in information from outside the historical data — a controlled experiment.

This is where the two articles in this cluster meet. A geo holdout — a matched treatment-versus-control regional test — produces a causal estimate of a campaign’s incremental lift that owes nothing to the observational tangle MMM is stuck in. You feed that estimate back into the model as a calibration target: a Bayesian prior in Meridian or PyMC-Marketing, or a calibration objective in Robyn. The experiment says “in this test, this activity produced this much lift,” and the model is nudged to agree, dragging the earned-media coefficient off its arbitrary perch and toward the causal ground truth. Practitioners describe this as moving the estimate up the incrementality spectrum, and it is the difference between a number you can defend to finance and one you cannot.

The honest position. An uncalibrated MMM can produce internally consistent results that are systematically wrong, and it is wrong most often exactly where earned media lives: high collinearity and low historical variation. So the rule is simple and firm — do not report an earned-media contribution from an uncalibrated model as if it were a finding. Run at least one geo holdout, feed it in, and let MMM and the experiment triangulate. The experiment supplies causation at a point; the model spreads that truth across the whole mix and keeps it current between tests.

This is also why honest measurement culture matters more than model sophistication. A calibrated range you can stand behind beats a precise-looking point estimate that will collapse the first time it is stress-tested — the same argument we make about communicating uncertainty in link building reporting generally. The teams that get compounding value from MMM treat it as ongoing infrastructure calibrated by a steady cadence of experiments, not a one-off report.

Cadence matters as much as the calibration itself. A single experiment pins the earned-media effect at one moment, but channel effectiveness drifts as algorithms, competitors and your own tactics change. A workable rhythm for most programmes is to refresh the model monthly and run a fresh calibration experiment each quarter on whichever earned activity carries the most budget — rotating through regional PR, then a link-heavy push, then a brand campaign — so the model is never relying on a stale read. Each experiment costs a few weeks of design discipline and returns something no volume of historical data can: an anchor to causation that keeps the whole model honest between tests.

Tooling and a worked read

Three open-source frameworks dominate in 2026, and any of them can run the structure above; the choice is about your team’s skills and stack, not model quality. All three assume a comfortable analyst — budget for that person, or use a guided commercial wrapper.

The frameworks and their prerequisites sit alongside the wider measurement stack in our best link building tools guide. All three need the same fuel: two to three years of weekly data, genuine variation in the inputs, and a monthly refresh.

Put it together with an illustrative read. A brand runs a decomposition and, in the first uncalibrated pass, the model credits digital PR with about 2% of incremental conversions — a rounding error that would justify cutting the programme. The team then runs a geo holdout on a regional PR push, measures a clean lift, and feeds it back as a calibration target. They also lengthen the PR carryover to a months-long half-life and route PR through a brand-search mediator. On the recalibrated pass, PR’s share rises to roughly 9%, and a chunk of what the first model had parked in “base” and “paid search” moves to where it belonged.

The numbers are illustrative, but the direction is not: calibration, longer carryover and a nested structure all move credit the same way — toward earned media and away from the baseline and the paid channels that were quietly banking it. That is the entire point of doing MMM properly and honestly for links and PR.

The read only matters if it changes a decision. A recalibrated 9% is not a trophy — it is a marginal-return question: at current spend, is the next pound into digital PR still buying more incremental conversion than the next pound into paid search? MMM’s response curves answer exactly that, which is why the calibrated model earns its keep at planning time, not just in the quarterly review. The disciplined move is to reallocate toward earned media only until its marginal return meets the alternatives, then hold — not to swing the whole budget on a single flattering decomposition. A model that shifts budget sensibly and survives the next experiment is worth more than one that produces a dramatic headline and cannot be reproduced.

Reading the model without fooling yourself

MMM fails in ways that look authoritative, which makes its failures dangerous. Watch for these five:

• Trusting an uncalibrated earned-media number. The headline error. High-collinearity, low-variation channels are exactly where an uncalibrated model is most confidently wrong. No experiment, no finding.
• Baseline leakage. If the base looks enormous and stable, some of it is almost certainly earned media misfiled as “free” demand. Interrogate the base before you accept the media splits.
• A carryover set too short. A paid-length half-life on earned media silently transfers its credit to whatever paid channel was live when the delayed conversions landed. Check the decay assumptions explicitly.
• Collinearity masquerading as precision. A clean-looking coefficient on a variable that moves in lockstep with another is a guess dressed as a measurement. Check variance inflation; combine or nest correlated inputs; lean on priors.
• Treating the model as a one-off. A single model is a snapshot that ages every week. Without a monthly refresh and a recurring calibration experiment, you will optimise toward channels that were merely present in good periods, not the ones that caused them.

The discipline that prevents all five is the same one that governs a good geo holdout: decide in advance what the model must show to change a budget, calibrate it against a real experiment, and report contributions as ranges the finance team can stress-test. Write down, before you look at the output, which decision each contribution figure will drive — a model you are unwilling to act on is a hobby, and a number you can only interpret after seeing it is a story rather than a measurement. For the benchmark context that keeps any contribution figure honest, our link building statistics for 2026 is the reference to keep open beside the output.

Three questions teams always ask

“We are a single small brand — is MMM even for us?” The open-source tools lowered the floor, but the data bar did not move: you still need two to three years of varied weekly data and an analyst who is comfortable with the modelling. If you have that, model at the organic-clicks level rather than revenue and calibrate hard. If you do not, skip MMM for now and run geo holdouts — they need far less history, cost far less to stand up, and answer the causal question directly.

“How is this different from the geo holdout in the companion piece?” They are complements, not rivals. A geo holdout is a bottom-up experiment that proves one campaign caused a lift with high confidence but narrow scope. MMM is a top-down model that covers the whole mix continuously but cannot, alone, identify a collinear channel. The best programmes use the holdout to calibrate the model — experiment for causation, model for coverage and currency.

“The model says PR is 2% of revenue. Should I cut it?” Not on that number. First ask whether the model was calibrated, whether the carryover was long enough, and whether PR was modelled through brand search and organic. A naive 2% is almost always an artefact of the three biases in Section 3, not a verdict on the channel. Fix the model before you touch the budget.

The Monday-morning version

You can start without a data scientist. This week, pull two to three years of weekly data — your outcome, your paid spend by channel, and a first earned-media input variable (PR placements or referring domains per week will do). Getting that history into one place is the hardest and most valuable step. Then plan the one thing that makes the whole model trustworthy: a geo holdout you can run this quarter to calibrate the earned-media effect.

The prize is real. Done naively, MMM is the tool most likely to talk a finance team out of funding earned links and digital PR — the very channels that compound hardest over time. Done properly — with the right input variable, a carryover that respects how slowly authority pays off, a nested structure that credits the mediators PR actually moves, and an experiment to pin it all down — MMM becomes the tool that finally proves earned media’s worth in the currency budgets are decided in. Understanding what link building is and how authority accrues is the foundation; giving it a defensible seat at the budget table is what this cluster is built to do. For teams expanding across borders, the same logic scales to country-level models covered in our guide to international link building.