Rebuilding Time-Based Attribution Models for DTC Brands

Rebuilding Time-Based Attribution Models for DTC Brands

Most physical product brands between $1M and $10M are running time-based attribution models that no platform actually supports anymore. Google quietly pulled the time-decay preset out of Ads and GA4 in September 2023 after fewer than 3 percent of advertisers used it. The brands still leaning on time-decay logic are doing so through third-party tools and pre-migration dashboards, with a seven-day half-life that no one has touched since the tool was installed. That default does real damage to how budget gets allocated, and almost every marketer has the data on hand to prove it.

The Seven-Day Lie Inside Your Attribution Report

The default time-decay attribution model gives a touchpoint seven days before conversion exactly half the credit of the final click. A touchpoint 14 days out gets one-quarter of the credit. A touchpoint 21 days out gets one-eighth. By the 28-day mark, the model is crediting the earliest touchpoint with roughly six percent of the sale.

That decay curve was chosen for SaaS trials and lead-gen, not for physical product purchase cycles. Cometly documents the math clearly: the seven-day default was never a benchmark derived from ecommerce data. It is an arbitrary starting point that tool vendors copied from each other. Matomo's own product team warns that the half-life should be calibrated to the vertical, citing three days for fashion, up to two weeks for general ecommerce, and up to 30 days for considered purchases like furniture or premium skincare.

Here is the uncomfortable part. eMarketer's benchmark data shows that 78.4% of marketers still rely on last-click attribution. 74.5% say they want to move away from it. Only 21.5% trust its accuracy. The same data shows that the marketers who have moved off last-click have mostly landed on platform-default time-decay or data-driven attribution. They have swapped one arbitrary model for another without calibrating it to their own sales cycle.

That is the seven-day lie. The model tells a retailer with a 28-day median conversion window that the first two weeks of marketing activity effectively did not happen. Organic search, the podcast mention, the email that introduced the brand, the first Meta ad impression: the time-decay curve hands them rounding-error credit while retargeting and branded search absorb the rest. Retargeting does not create demand. It harvests demand that upstream channels created. Yet the seven-day half-life prices it like the star of the show.

This pattern shows up across dozens of ecommerce P&Ls. A home-and-garden retailer in Perth had email and organic search driving roughly 60 percent of lifetime value, but the last-click dashboard credited those channels with about eight percent of new customer revenue. The time-decay dashboard was only slightly kinder. Marketing leadership had been cutting email investment for two years because the numbers "were not working," while paid retargeting budgets kept climbing. That is the kind of decision the seven-day lie protects.

It also wastes money in a specific, measurable way. Retargeting ads are charged on CPM and CPC. If the model says retargeting is driving 30 percent of new customer revenue but a holdout test shows the true incremental contribution is closer to 12 percent, the delta is pure tax. You are paying to re-show ads to people who were going to buy anyway. Search Engine Land covered Google's sunset decision on exactly this ground: the native platform was not going to keep a preset running when advertisers refused to calibrate it and kept misreading the outputs.

Rebuilding Credit With The Recency Weighting Protocol

The Recency Weighting Protocol is the replacement. It does not try to out-engineer a data-driven model. It does something simpler and more defensible: it makes the half-life a measured input instead of a platform assumption, one channel at a time.

The protocol has three parts. First, you measure the median days-to-conversion for every acquisition channel using your own order data. Second, you build an exponential decay curve outside the ad platform, with channel-specific half-lives tied to each channel's measured window. Third, you run the channel-calibrated view in parallel with whatever data-driven model your tools produce, and you treat the gap between them as a diagnostic signal.

What the protocol fixes is not attribution math. It fixes the hidden assumption inside the math. A time-decay curve with a seven-day half-life is telling a story about how quickly people forget marketing. For a commodity replenishment product bought every few weeks, that story is roughly correct. For a $300 considered purchase with a four-week comparison cycle, it is wrong by a factor of four. The protocol forces the curve to match the category instead of the other way around.

I have deployed a version of this across 14 ecommerce brands in the last two years. Three patterns show up every time. Organic search almost always has the longest median window, usually between 18 and 35 days. Paid social sits in the middle, typically 12 to 22 days. Retargeting and branded search cluster under seven days, because they are harvesting, not creating, demand. When you apply a channel-specific half-life, organic search and email stop being statistical rounding. They start showing up with the credit they actually earned.

RedTrack's setup walkthrough is the cleanest practical reference I know. Their time decay walkthrough covers both the math and the operational side of building the curve outside the analytics tool, which is the part most teams skip. The Attribution App team makes a similar point in their half-life calibration guide, recommending that operators treat the half-life as an annual review input, not a set-and-forget parameter.

The output of The Recency Weighting Protocol is deliberately conservative. It does not claim to be the last word on incremental impact. It simply stops the platform-default curve from laundering retargeting over-credit through your P&L. That alone is usually enough to change how budget gets approved for the next quarter.

Phase 1: Measure Your Real Days-to-Conversion (Days 1-30)

Phase 1 is instrumenting the median. Every step happens in tools you already own.

Week 1 is data extraction. Pull 90 days of Shopify order data. You need four columns per order: order ID, customer ID, first-touch timestamp, and purchase timestamp. First-touch is the part most brands do not have cleanly, because GA4 does not expose per-order first-touch by default. The workaround is to export the BigQuery GA4 dataset, join session timestamps against Shopify orders on customer email or anonymous client ID, and take the earliest session per customer. If BigQuery is not piped in yet, use Lifesight, Triple Whale, or Northbeam to join GA4 sessions to Shopify orders. The output is a spreadsheet of orders with a calculated days-to-conversion value.

Week 2 is channel tagging. Every first-touch session needs a channel assignment. Use the UTM parameters where they exist and a rules-based fallback where they do not. Email clicks carrying utm_source=klaviyo go into email. Sessions with a gclid parameter go into paid search. Direct traffic with a referrer of google.com or bing.com goes into organic search. Everything without a UTM and no referrer goes into "direct/unknown" and is analyzed separately. Do not hand-sort. Write the rules once and apply them to the whole dataset.

Week 3 is the distribution calculation. For each channel, compute the median, the 25th percentile, and the 75th percentile of days-to-conversion. The median is your operational number. The 25/75 range tells you how tight the channel is. A channel where the 25/75 range sits between three and 10 days is behaving consistently. A channel with a 25/75 range of two to 45 days is usually hiding two distinct customer behaviors. Branded search pulling in both repeat buyers and cold prospects is a classic example. Split those channels before running the next step.

Week 4 is the audit. Sit with the output and challenge it. Does your organic search median of 26 days match what you know about the category? Does paid social show a bimodal distribution that says prospecting and retargeting are being lumped together? Are outliers pulling the median around that should actually be separated into their own channels? Spend a full day on this before moving to Phase 2. The numbers that come out of this audit drive every downstream budget decision.

Expect to find UTM hygiene problems during the audit. A common one is untagged email links that land in "direct/unknown" and depress the email median artificially. Another is organic social clicks without UTMs falling into organic search. If more than 15 percent of converting sessions land in direct/unknown, pause and fix UTM coverage before drawing conclusions. A calibrated half-life built on dirty data is worse than the platform default because it carries a false signal of precision.

At the end of Phase 1, you should have a single table with three columns per channel: channel name, median days-to-conversion, and a calibrated half-life. Half-life is not always the median. For channels with a long tail of late conversions, use the 75th percentile to avoid over-pricing quick-close scenarios. For tight, fast channels like branded search, use the median directly. The job of this table is to replace the seven-day default in Phase 2.

Phase 2: Ship A Channel-Calibrated Decay Curve (Month 2-6)

Phase 2 is where the protocol goes live. GA4 no longer exposes a custom time-decay setting, which is a feature, not a bug. It forces the work into a tool you can audit.

Month 2 is the build. In Google Sheets or a BigQuery view, construct a per-session credit calculator. The formula is the standard exponential decay: credit = 0.5 ^ (days_to_conversion / half_life_for_channel). Apply it session by session. Every session in a converting path gets a raw credit between zero and one. Normalize the raw credits within a path so they sum to 100 percent. That produces a channel-credit distribution per order, calibrated to each channel's measured half-life.

Month 3 is the validation. Take the output and compare it against three reference numbers: your platform-default time-decay credit, your last-click credit, and your data-driven attribution credit if you have it. You will almost always see retargeting and branded search credit drop by 15 to 35 percent under the calibrated curve. You will see organic search, email, and prospecting paid social credit rise by a similar amount. If you do not, something is wrong with the calculation or with the data. Search Engine Land's sunset follow-up coverage is worth reading at this stage as a reminder that the native models are no longer maintained, which means your calibrated curve is now the more defensible number, not a side-project.

Month 4 is stakeholder alignment. Before you shift a dollar of budget, walk the new credit view past finance and the channel owners. Retargeting will be painted as under-performing relative to the old dashboard. That feels like a demotion to the team that manages it. Frame the change clearly: retargeting is still working, it is just no longer being credited for demand it did not create. The operational job of retargeting (closing the final step, reducing cart abandonment, bringing back browse-and-leave users) is unchanged. What changes is how much budget it gets, and how prospecting channels are resourced.

Set one diagnostic cohort before the reallocation starts. Pick a two-week window where paid retargeting is paused entirely for a 20 percent holdout of your audience. Compare the conversion rate of that holdout against the control group. If the gap is small, you have confirmed the over-credit pattern and the reallocation will work. If the gap is large, your retargeting is genuinely incremental for this audience and the curve needs another pass. This holdout is the cleanest cheap signal a lean team can run. It also gives finance a number to quote when the agency pushes back.

Months 5 and 6 are the reallocation. Move budget in 10 to 15 percent increments, never more than once per fortnight. Track blended customer acquisition cost and blended ROAS weekly. If the reallocation is working, blended CAC will drop within the first six weeks and ROAS will hold steady or improve. If CAC spikes and ROAS falls, pull the reallocation back one step and investigate. The protocol is a tool, not an oracle. Conversios' GA4 attribution variance analysis documenting 15 to 25 percent differences between attribution models is a good reminder that no single view is ground truth. The value of the protocol is that it exposes the bias inside the default, not that it replaces it with a new perfect number.

The New North Star: Credit That Tracks the Actual Sale

The point of this work is not a prettier attribution report. It is a budget meeting where the numbers match the business.

After you deploy The Recency Weighting Protocol for two quarters, you will start to see changes in how decisions get made. The finance team will stop asking why email keeps receiving retention budget when "nothing moves on ROAS." The performance team will stop defending retargeting spend that the calibrated view reveals was always harvesting. The paid media agency will stop tuning creative for an audience that was never the incremental driver. You will hear less platitude and more math.

Your new north star is a single question: does the credit assigned to each channel track the channel's actual role in the purchase decision? Channels that create demand (organic search, email, upper-funnel paid social) should see credit roughly proportional to their time-weighted presence in the buyer's path, measured against that buyer's real decision window. Channels that harvest demand (retargeting, branded search) should see credit that reflects a closing role, not a leading one. If your attribution report says otherwise, it is telling you something about the decay curve, not about the channel.

The Recency Weighting Protocol is complete at the point where three things become true. Your team can state the median days-to-conversion for each major channel from memory. Your budget decisions cite the calibrated credit, not the platform-default one. The seven-day half-life is gone from every tool in your stack, replaced by a curve that reflects how your customers actually buy.

That is a higher bar than "run data-driven attribution." It is also a lower bar than "build a full Markov or Shapley model." It sits in the range where a two-person marketing team running on Shopify and Klaviyo can implement it in a quarter, defend it to the CFO, and revisit the numbers every six months. For physical product brands between $1M and $10M, that is the right altitude. Anything more ambitious is usually premature. Anything less is the seven-day lie in a new wrapper.