Choosing an attribution window length for a show with a long buying cycle

A visitor scans into the March edition, walks the floor, and goes home. In September she opens a marketing email about next March, clicks nothing, and forgets it. In January she searches the show name, lands on the site, and registers.

Ask any reporting system which channel found her and it will say organic search, because organic search is the only touch that survives inside a default window. The visit that actually created the intent happened ten months earlier and, as far as your attribution is concerned, never happened at all.

Attribution window length is the setting that decides which touches exist. It is chosen once, usually by whoever set up the tag, and it silently outranks every model decision made afterwards.

What the defaults actually are

Google's Analytics Help documentation, as it stands in 2026, sets the lookback window for acquisition key events, meaning first visits and first opens, to a default of 30 days with a 7 day alternative, and for all other key events to a default of 90 days with 30 and 60 day options. Google Ads help documentation, checked in the same year, sets the click through conversion window to a default of 30 days, adjustable from 1 up to 30, 60 or 90 days for Search and Display campaigns, with a view through conversion window that defaults to 1 day.

Those numbers were chosen for businesses whose customers decide in a week. Nothing about them is wrong. They are simply parameters set for a different buying cycle, and an exhibition sitting on them is accepting somebody else's assumption about how long a decision takes.

The ceiling matters as much as the default. If your true journeys run to five months, a platform capped at 90 days cannot represent them whatever you set, which is one of the reasons the platform number and the file number will never reconcile without a worksheet.

Measure the distribution before you argue about the setting

The query is not hard. For every registration in a completed edition, find the earliest recorded touch, take the difference in days, and count the results into buckets.

Here is the shape from a file of 5,000 journeys, which is typical enough to work with.

The 90th percentile is the 4,500th journey in that sorted list. Cumulative counts reach 4,200 at day 90 and 4,700 at day 180, so the 4,500th sits inside the 91 to 180 bucket. It is 300 journeys into a bucket holding 500 spread across 90 days, so interpolating gives 90 plus 0.6 times 90, which is 144 days.

Round it to 150 and you have a defensible window with a stated basis: nine registrations in ten have their entire journey inside it.

What does a 30 day window discard?

On this distribution, 2,750 of 5,000 journeys start within 30 days of the registration. The other 2,250, which is 45 per cent, have a first touch the 30 day window cannot see.

Credit is not lost when that happens, which is what makes the effect hard to spot. It moves. The registration still exists, so whatever touch remains inside the window collects the credit that would have gone to the truncated one. Deadline email, retargeting and brand search sit late in the journey by construction, so they are the beneficiaries in nearly every case.

A 90 day window discards 800 journeys, which is 16 per cent, and it is a much smaller distortion. Going from 90 days to 150 days recovers a further 300 journeys, or 6 per cent of the file. That is the honest way to present the choice: each extension buys you a measurable number of journeys, and the marginal return falls quickly.

There is a second number to compute alongside it, and it is the one that usually settles the argument. Split the truncated journeys by first touch channel. If the 2,250 journeys that a 30 day window would truncate are 48 per cent paid social and 22 per cent partner referral, then the window is systematically moving credit away from two named channels and towards email, which makes it a budget decision taken by a default setting.

Why a show's distribution has two humps

Most exhibition files do not produce a smooth decay. They produce a spike near the deadline and a second, flatter cluster at whatever interval separates editions.

The spike is the fortnight before doors, where journeys start and finish inside a few days because the person was prompted by an exhibitor, a colleague or a deadline email and acted immediately.

The second cluster is people whose relationship with the show began at the last edition, or at the point the previous edition's post-show communication went out. On an annual show that puts real mass somewhere near 300 to 365 days, and no window short of a year will ever capture it.

That second cluster is the argument for handling returning attendees separately rather than stretching the window to twelve months. A 365 day window will attribute nearly every returning registration to whatever the previous edition's post-show email happened to be, which is technically consistent and analytically useless. Report registrations from people who attended the last edition as their own group, then run the window logic on everybody else. The number that matters for the returning group is retention. Measure what share of last edition's attendees registered again, and how early in the cycle they did it.

What do you do when the platform will not go past 90 days?

Set the platform to its maximum, then stop treating the platform as your attribution system.

The platform window governs which conversions the platform claims, which is what it bills against and what its bidding algorithm optimises towards. That is a legitimate job and 90 days is enough to do it, because a bidding algorithm needs a signal within days rather than months.

Your own window governs how credit is assigned to rows in the registration file, and it lives in your warehouse where the touch table has no ceiling. A 150 day window there costs a wider date filter on a join.

Running the two at different lengths is fine as long as nobody is surprised by it. Write both numbers on the reconciliation: platform window 90 days, reporting window 150 days, difference in claimed registrations 214. Once the gap has a line and a cause, it stops being an argument and becomes a note.

The failure to avoid is setting your reporting window to 90 days because the platform does. That imports a constraint from a system that has a different job into a system that does not share it, and it is the single most common reason a show's reporting understates everything that happens before the summer.

The cost of a long window

Windows are not free, and the costs run in the opposite direction to the truncation problem.

A longer window credits more touches per registration, so the average journey gets longer and every channel's raw touch count rises. Channels with high volume and low intent, such as broad display or a heavily sent newsletter, gain the most, because their chance of appearing somewhere in a 150 day span is close to certain.

Identity drift is the second cost. Over five months people change devices, clear cookies and change jobs. The further back the window reaches, the more of its matches rest on a weaker key, and a match made on a stale key is worse than no match because it is confidently wrong.

Storage and query cost is the least interesting of the three and the one most often cited. A touch table for a large show over 150 days is millions of rows, which is unremarkable for any warehouse built this decade.

My own preference is the 90th percentile, computed per show rather than per portfolio, revisited once a year, with the returning attendee group excluded from the calculation. Shows in the same portfolio genuinely differ: a regional trade fair with a two month campaign and an international congress with a nine month one should not share a window because it is convenient for the reporting template.

Where this stops

The distribution you measure is a distribution of recorded touches. Real journeys start earlier than the record does. Anything before your tracking picked the person up is invisible, so the measured 90th percentile is a lower bound on the true journey length, and the gap is largest for exactly the channels that reach people who are not yet in your database.

Windows also interact with everything downstream. A decay model inside a 150 day window behaves very differently from the same model inside 30 days, since the half life and the window bound each other. A sequence model needs whole journeys, so a short window truncates the sequences a transition matrix is built from and biases the transition probabilities towards late channels.

And the window will not make your numbers agree with the platform's. Different windows are one of the four standard causes of that gap, and walking the platform number down to the file is a separate exercise. The acquisition and attribution pillar covers where these decisions sit together.

Run the first touch to registration distribution on your last completed edition this week and find the 90th percentile. Then compare it with whatever your reporting currently uses, and if the two are more than a month apart, you have been reading a channel ranking that was decided by a default rather than by your audience.