Adobe put out a fresh launch post for Marketing Campaign Analytics yesterday. It’s the product that used to be called Adobe Mix Modeler, renamed at Summit in April, and the October 29 post leads with a line every Q4 media buyer wants to hear: “The model rescores in minutes as new data comes in, so you can adjust creative, targeting, or delivery while the campaign is in flight.” The same day, the release notes marked full funnel modeling as generally available.
This is a docs-based review. I have not had access to a Marketing Campaign Analytics org, built a model or moved a dollar on one. Everything here comes from Adobe’s April and October launch posts, the product page and the Experience League documentation (every page I used was updated October 29), checked against an independent product map kept by an Adobe Champion. When the docs don’t answer something, I say so.
My read: the modeling is serious and the docs are unusually honest. But the minutes in the pitch are scoring minutes. What the model actually believes about each channel changes when it retrains, and you pick that schedule from Monthly, Quarterly, Yearly or None. Some of those beliefs can be typed in by hand. Four weeks before Black Friday, the questions that matter are when your model last learned something, and who typed what into it.
What it is
Marketing Campaign Analytics is Adobe’s measurement and budget planning app inside CX Analytics, built on Adobe Experience Platform. The overview docs describe a marketing mix model, which works top down on aggregate spend and outcomes, and an optional multi-touch attribution layer, which works bottom up on event data. Adobe joins the two with what it calls bi-directional transfer learning, so the person-level and channel-level answers are supposed to agree.
The workflow is ingest, harmonize, build, train and score, then plan. You bring in summary data like walled garden spend and TV flights, plus event data if you have it, map everything to common fields, and build a model around one conversion such as revenue, orders or leads. Plans then take a trained model, a date range and a budget, and recommend a split by channel. You can set minimums and maximums per channel and per flight.
The October 29 post lists what’s new on top of Mix Modeler: a conversational experience, data quality checks before a run, full funnel attribution, channel synergies, planning across multiple conversions, and results flowing into Customer Journey Analytics. Natural language questions run through Adobe CX Enterprise Coworker, the agent layer Adobe is putting across its apps.
What changed this week
Two things. First, the relaunch post, which shifts the pitch from “measure” to “optimize mid-flight.” Second, full funnel modeling. According to the models overview, you can now model one upper funnel KPI, like newsletter signups, alongside one lower funnel KPI, like revenue. The upper funnel model’s weights feed the lower funnel model, so channels that build demand get credit for the sales they help cause, and Adobe says the contributions add up “with no double-counting.”
That’s a real problem worth solving. Adobe’s own docs admit that “upper funnel tactics often get lesser credit in a conversion focused MMM model,” and every brand team has sat through that meeting. One catch on availability, which I’ll come back to: the release notes say generally available, while the models overview page still carries a Limited Testing banner over the full funnel section.
What works
You can see the knobs. Most mix models arrive as a vendor deck. Here, the build docs show every setting: the lookback window (1 to 52 weeks), up to 30 internal and external factors with a positive, negative or automatic effect, channel adstock with lag and carryover per channel, and a spend share option for channels with sparse data. If you’ve wanted an in-house model without a team of PhDs maintaining open source code, this is the closest a big suite has come.
You can feed it your lift tests. The prior knowledge step lets you enter a contribution percentage for any channel, with an optional confidence level. If you ran a geo holdout on paid social last spring, that result can shape the model instead of sitting in a slide. That’s the right way to run a mix model.
The diagnostics are there. The model insights page shows R squared, MAPE and RMSE on training fit, plus out of sample error. You get marginal response curves with a break-even point for each channel, which is the chart your CFO actually wants. There’s also a channel synergy heatmap scaled 0 to 10, and everything downloads to CSV or Excel.
The fine print is honest. The product page FAQ says the app is “designed to inform business judgment, not replace it,” and that incremental impact “is a model-based estimate, so review it alongside its inputs, assumptions, and model quality.” The build docs flag the Super Bowl problem themselves: one-off investments should go in as a factor, not get smoothed into spend share. That’s more candor than most measurement vendors manage.
What breaks
Scoring and learning run on different clocks. The train and score docs separate the two. Training is where the model learns how each channel drives your conversion. You retrain on at least one year of data, and Adobe suggests doing it when “market dynamics have changed.” Scoring runs new marketing data through what the model already learned. The scheduler offers Daily, Weekly or Monthly scoring, but training only Monthly, Quarterly, Yearly or None. So “rescores in minutes” means fresh numbers run through last month’s beliefs, or last quarter’s, or whatever the person who set the schedule picked.
That matters this quarter. If your model last trained in August, November traffic gets read through summer behavior. The launch post says H&R Block’s model “refreshes weekly.” The scheduler has no weekly training option, so ask which of the two “refresh” means before you copy the story.
The drift alert may not be in your org yet. The launch post says “an automatic alert flags any drift in performance.” On the model insights page, model drift detection sits under a Limited Testing banner that says it “might not be available yet in your environment.” Under Adobe’s release strategy, features go from Limited Testing to full rollout “over the course of a couple months,” switched on at the organization level. The alarm the pitch leans on is not guaranteed to ring for you before Cyber Week. And when it does ring, the dialog offers “Remind me later.”
Your plans don’t follow the model. This one gets an Important box in the docs and no mention in the pitch. The plans overview says that if the model behind a plan is scored on new data, “you need to create a new plan to take into account the rescored data.” The train and score page repeats it: rescoring “does not change any Plans that are already created.” So the in-flight loop is rescore, then rebuild the plan, then compare, then decide. The budget split your agency is executing will sit still until a human rebuilds it.
Priors are a steering wheel. The same box that takes your lift test also takes “past stakeholder experience.” It’s a contribution percentage and a confidence level, entered by whoever has edit rights. Nothing in the docs I read requires a source for a prior. The launch post sells exactly this: teams can “adjust the factors and underlying assumptions, and add their own business context.” That’s power and risk in one sentence. A channel lead who believes in their channel can type that belief in at 80 percent confidence and get a model that agrees.
The holdout is short. Out of sample error uses “the last quarter of training data as a holdout set.” That’s a reasonable default, but if last quarter was summer, it tells you little about how the model handles a peak season it may only have seen once or twice in its window. Adstock is capped at one eighth of the training window, which works out to 13 weeks on a two year window. Long brand effects past that get cut off by design.
The agentic layer is mostly promise. The independent CX Enterprise atlas, written by an Adobe Champion, lists proactive optimization as “coming soon” and the collaboration workspace as “in development.” It also notes the conversational interface wasn’t documented in Experience League when it last checked on October 27, and that in-flight optimization rests on a dashboard Adobe marks as beta. The full funnel GA label next to a Limited Testing banner fits the same pattern. Read the docs page, not the launch post, for what is switched on.
Who it’s for
Enterprise brands already on Adobe Experience Platform, with a couple of years of clean weekly spend and outcome data (the default retrain window is two years) and a real analyst who will own the model. Ideally they also have a few lift tests to use as priors. Retail, subscription and seasonal businesses that live or die on one window will get the most out of the planning side, if they retrain before the window rather than during it.
It’s also for CMOs who want to stop renting their mix model from a consultancy once a year. Owning it means owning the schedule, the priors and the change log too.
Not for teams without Platform, without a data owner, or hoping an agent will replace an analyst. Also be careful with the “partner evaluations” use case in the overview, which pitches judging “the effectiveness of agencies and partners.” If the agency being judged can also edit the priors, you don’t have an evaluation.
Pricing and limits
No public price. The product page sends you to a demo request, and says existing Mix Modeler customers should “contact your Adobe account team for details about naming, included capabilities, and access.” Connectors are partial: “Not every platform has a built-in connector, so confirm supported ingestion options for each source with your Adobe team.” Multi-touch attribution “cannot measure most offline touchpoints.” Touchpoints can’t overlap, and every model needs at least one channel with spend. How much history you need “depends on your data coverage and model configuration,” which is a polite way of saying ask.
What to do this week
Put two dates on every readout. The last training date and the last scoring date, side by side on every slide that moves money. If anyone says the model “updated,” ask which one.
Decide your peak plan now. Either retrain before Black Friday on a window that includes last year’s peak, or add a peak factor so the spike doesn’t get read as channel performance. Don’t let a model trained in summer make the November call on its own. This matters even more if Meta’s on-device conversion default starts shifting platform-reported purchases on November 12. An outside model is only a useful check if it’s current.
Make priors earn their place. Every prior needs a source (which test, which date), an owner, and a line in a change log. No source, no prior. Keep agencies out of edit rights on the models used to grade them.
Build the plan rebuild into the cadence. Since rescoring doesn’t touch existing plans, schedule the rebuild and the comparison right after scoring, and make a named person approve the new split before it goes to the agency.
Get four answers from Adobe in writing. Is drift detection on in our org, and when? Is full funnel modeling on, given the banner? Which of our ad platforms have prebuilt connectors? What changes in our contract and access under the new name?
Adobe built a mix model you can actually open up and inspect. That’s the good news. It also means the model’s answers are partly made of decisions someone on your team typed in. Before November, make sure you know who that was and when they did it.
