Why a time decay attribution model fits a deadline driven show

Registration for the autumn edition opened in January. By the end of August the file held 6,800 rows. Over the last fourteen days before doors, it took another 4,900.

Every attribution scheme has to decide what to do with that shape, and most of them decide by accident. A time decay attribution model decides on purpose, with one parameter you choose, defend and write on the report.

The formula, and what the parameter does

Each touch gets a weight of two raised to the power of minus t over h, where t is the number of days between the touch and the registration and h is the half life. Sum the weights across the journey, then divide each weight by that sum so the credit adds to one.

The half life is the whole model. A touch exactly one half life before the registration is worth half as much as a touch on the day of registration. Two half lives back and it is worth a quarter. The curve never reaches zero, which is the property that separates decay from a hard cut off.

Google's Analytics Help documentation, as it stands in 2026, defines the model the same way and sets a default half life of seven days, so a touchpoint seven days before a conversion receives half the credit of one on the conversion day, with the decay continuing inside a lookback window that defaults to 30 days. Seven days is a sensible default for a retailer. It is close to indefensible for an exhibition that sells for three quarters of a year.

The same documentation records that the first click, linear, time decay and position based models were removed from Google Analytics as of November 2023, leaving data driven attribution and two last click variants. Whatever decay you run from here is one you built yourself, against your own touch table, which makes the half life yours to justify and yours to publish.

What does a fourteen day half life do to one journey?

Take a registration with four recorded touches, at 120, 60, 21 and 3 days before it completed, and run the arithmetic with a half life of 14 days.

The touch at 120 days gets two to the power of minus 120 over 14, which is two to the minus 8.571, or 0.00263. The touch at 60 days gets two to the minus 4.286, which is 0.05127. The touch at 21 days gets two to the minus 1.5, which is 0.35355. The touch at 3 days gets two to the minus 0.214, which is 0.86198.

Those four weights sum to 1.26943. Divide each by the sum and the credit split is 0.2 per cent to the touch at 120 days, 4.0 per cent to the touch at 60 days, 27.9 per cent to the touch at 21 days and 67.9 per cent to the touch at 3 days.

Now run the same journey with Google's seven day default. The weights become 0.0000069, 0.00263, 0.125 and 0.74301, summing to 0.87065. The split is effectively nothing for 120 days, 0.3 per cent for 60 days, 14.4 per cent for 21 days and 85.3 per cent for the final touch.

Same journey, same data, and the last touch moves from 68 per cent to 85 per cent. The parameter is not a detail. At a seven day half life the model is a slightly softened last touch, and if that is what you want, use last touch and save yourself the explanation.

Run it the other way for completeness. At a 45 day half life the weights are 0.1575, 0.3969, 0.7236 and 0.9548, summing to 2.2328, so the same four touches split 7.1, 17.8, 32.4 and 42.8 per cent. Stretch the half life to 180 days and the split flattens to roughly 19, 24, 28 and 30 per cent, which is nearly the linear model. One parameter moves the output across the whole space between last touch and equal credit, which is why quoting the result without quoting the half life is meaningless.

How do you pick the half life from your own file?

Do not pick it from a blog post, including this one. Pick it from the distribution you already have.

The measurement is one query. For every registration with at least two touches, compute the number of days between each touch and the registration, then take the median across all of those touch level gaps. On a nine month show with a heavy deadline, that median usually lands somewhere between ten days and three weeks, because the touch table is dominated by late email clicks and retargeting even though the calendar is long.

Setting the half life to the median gap has a clean interpretation. Half your recorded touches sit closer to the registration than the half life and are worth more than half of a same day touch. Half sit further away and are worth less. The model is then centred on your own show's behaviour rather than on a default written for online retail.

Two adjustments are worth making after that.

If your show has a price break structure, check the median against the gap between the last price increase and the show. A half life much longer than that gap will smear credit across a boundary your audience genuinely responds to. If the early bird closes eight weeks out and the final rate applies from four weeks out, a half life of 28 days will treat the two periods as nearly interchangeable when your registration curve says they are not.

If you run a genuinely long lead programme, such as onsite rebooking of attendees at the previous edition, a short half life will bury it. That is an argument for reporting those registrations separately rather than for lengthening the parameter, because one channel with a twelve month lag should not set the weighting for the other eleven.

What decay fixes that a fixed weight split does not

A position based rule such as 30, 30, 30 and 10 treats the third touch of nine as the third touch of nine whether it happened yesterday or last November. For a business where the deadline does the selling, that discards the most informative column in the touch table.

Decay uses it directly. Two registrations with identical channel sequences but different timing get different credit splits, which is correct: a paid search click 90 days out did something different from the same click 3 days out.

Decay also degrades gracefully. A journey with one touch gets 100 per cent to that touch under any half life. A journey with forty touches, which happens with heavy email senders, spreads credit smoothly instead of concentrating 90 per cent on three touches picked by position. The fixed weight approach and its trade offs sit with the W shaped split, and the honest summary is that decay is better suited to a deadline and harder to explain in a room.

There is one place the fixed split wins. Position based rules are stable across editions in a way decay is not, because a change in send frequency changes the shape of the touch table and therefore the decay output without any change in performance. If your email team doubles the number of deadline reminders, decay will report email up and you will need to say why.

Where the model stops being useful

Time decay is a rule about recency and nothing else. It has no view about whether a touch did anything, and it will confidently credit a channel that appears late in every journey for the sole reason that it appears late in every journey.

Retargeting is the clearest case. It fires after somebody has visited the registration page, which is to say after they have already decided, and it will therefore sit close to the registration in almost every journey it touches. Under any short half life it collects a large share, and Li and Kannan, writing in the Journal of Marketing Research in 2014, found the opposite of that flattering picture: their model of channel consideration and visits over time, validated by a field experiment, identified cases where retargeting actually decreased conversion probability. A decay model cannot see that. It sees a recent touch.

The second limit is the window. Decay never reaches zero mathematically, but your lookback window does, and every touch outside it contributes exactly nothing whatever the half life says. Google Ads help documentation sets a default click through conversion window of 30 days, extendable to 60 or 90 days for Search and Display campaigns, which means an out of the box setup discards touches from before the previous quarter regardless of your decay parameter. Choosing that boundary is a separate decision with its own arithmetic and it bounds this one.

The third limit is that the parameter is still your opinion. It is one number instead of four, which is progress, and it is not evidence. If you want the data to set the weights rather than a parameter you chose, a transition matrix over the observed sequences is the next step up, at the cost of needing your non converting journeys as well as your converting ones. Where any of this sits in a wider acquisition picture is the subject of the acquisition and attribution pillar.

Compute the median touch to registration gap on your last edition this week. If it comes back near seven days, the Google default happens to suit you and you should say so in writing. If it comes back at 18 days, as it does on most shows I have seen, then every time decay number your platform has ever shown you has been weighted for somebody else's business.