← All articles How it works

Six attribution models, side by side

One customer journey, six ways to share the credit. What last touch, first touch, linear, time decay, position based and data-driven attribution each assume, and how to read them in InsightAdScale.

A customer sees your connected-TV spot, clicks a social ad a week later, scrolls past a display banner, and finally searches for your brand and buys. Which channel earned the sale? Every attribution model is a different answer to that question, and none of them is the truth. Knowing what each one assumes is how you stop arguing about the number and start using it.

In short

Rule-based models (last touch, first touch, linear, time decay, position based) apply a fixed rule to every journey. A data-driven model learns the split from your own paths. InsightAdScale shows each channel's conversions, revenue and credit under whichever model you pick, so you can see where the models agree and where they don't.

The six models

ModelThe ruleAssumesTends to favour
Last touchAll credit to the final interactionThe closer did the workSearch, retargeting
First touchAll credit to the first interactionDiscovery is what mattersCTV, video, display
LinearEqual shares to every touchEvery touch helped equallyChannels with many touches
Time decayMore credit the closer to the saleRecent touches matter moreLower-funnel channels
Position based40% first, 40% last, 20% shared in betweenOpening and closing both matterThe ends of the journey
Data drivenCredit learned from converting and non-converting pathsYour data can show what changes the oddsDepends on your data

Two cautions apply to all of them. First, every model shares out credit for conversions that happened, so none of them says whether a conversion would have happened anyway; that is what incrementality testing is for. Second, a model can only credit touches it can see. A CTV impression that cannot be tied to the later visit gets no credit under any rule, which is why cross-device matching matters.

Worked example: one order, five rules

Put a value on the journey in the diagram. Say the purchase on day 15 is a made-up $600 order. Under each rule-based model the $600 is shared like this (time decay with a seven-day half-life, rounded to the dollar):

ModelCTV (day 1)Social (day 8)Display (day 12)Search (day 14)
Last touch$0$0$0$600
First touch$600$0$0$0
Linear$150$150$150$150
Time decay$62$125$186$227
Position based$240$60$60$240

The time-decay figures come from weighting each touch by one half for every seven days between it and the purchase: 14 days for the CTV view gives a weight of 0.25, 7 days for the social click 0.5, 3 days for the display view about 0.74 and 1 day for the search click about 0.91. Each touch's share is its weight divided by the total. Nothing about the customer changed between rows; only the rule did. CTV's credit for the same order runs from nothing to all of it.

Across a real portfolio the same thing happens thousands of times, and the per-channel totals on the Attribution screen are the sum. That is why a channel's number only means something next to the name of the model that produced it.

Reading it in InsightAdScale

On the Attribution screen, the model picker at the top switches every panel at once. The Channel credit breakdown lists, for each channel, its attributed conversions, attributed revenue and share of the model's credit; the chart beside it draws the same shares. Change the model and watch which channels move.

The Attribution screen with the model picker set to Data Driven, a channel credit table of conversions, revenue and credit share, and a bar chart of credit by channel.
Attribution under the data-driven model for a made-up portfolio: credit share, conversions and revenue for six channels.

Further down, the Attribution funnel lays out impressions by journey stage (awareness, consideration, intent, conversion, loyalty) and the channels that drive each stage. The CTV-to-digital mapping follows connected-TV views to the mobile, desktop and tablet visits that came after, with the cross-device match rate that makes those links possible. In the current build these two panels show built-in sample figures; the channel credit breakdown above them reads your loaded results.

The attribution funnel from awareness to loyalty and the CTV-to-digital mapping table with CTV impressions, digital conversions attributed and cross-device match rate.
The journey funnel and CTV-to-digital mapping, further down the same screen.

Choosing a model, practically

  • Pick one model for reporting and stick to it. Switching the reporting model mid-quarter makes every trend line meaningless. Use the others to explore.
  • Look at the spread, not just the number. If a channel's share barely moves between first touch and last touch, its credit is robust. If it swings from 5% to 40%, its value depends on your assumptions, and it is a good candidate for a lift test.
  • Match the model to the objective. An awareness flight judged on last touch will always look poor. Judge it on reach and lift, or at least on a position-based or data-driven view.
  • Keep data-driven honest. A learned model needs enough converting and non-converting paths to be stable. On a small advertiser or a new channel, a rule-based model is easier to explain and harder to over-fit.
What InsightAdScale stores. Attribution results are held per channel and per model, so switching models is instant and every user sees the same numbers. The model results themselves come from your attribution data; the screen does not re-run a model when you switch.

What you load, and what the screen does with it

Each attribution result is a record tied to a campaign, with the model type, the channel credited, the credit value, attributed conversions and attributed revenue, plus optional fields for the conversion ID and the touchpoints in the path. The screen sums conversions, revenue and credit by channel for the selected model, and the model picker lists the model types present in your data. In practice that means:

If you…You will see…
Load only last-touch resultsOne model in the picker; the others cannot be compared yet
Load the same conversions under several modelsA like-for-like comparison: switch models and the totals stay the same while credit moves
Load models for different periodsTotals that differ between models for reasons that have nothing to do with the rules; avoid this
Spell a channel two waysTwo rows for one channel; agree channel names before loading

A quick spread check

  1. Pick the period and the channels you report on.
  2. Note each channel's credit share under last touch and under first touch.
  3. Note it again under the model you report on.
  4. Mark any channel whose share moves by more than you would accept in a budget conversation. Your team sets that line.
  5. Those channels are the candidates for a lift test before their budget is cut or grown.

Questions analysts ask

Does InsightAdScale run the attribution models?

No. It stores and compares results produced by your attribution process, whether that is a vendor, a platform export or your own data science team. That keeps the numbers auditable: what is on screen is what was loaded.

Which model should leadership see?

One, agreed in advance and kept for the year, with a footnote naming it. Show the spread only when a decision depends on it.

Can data driven be compared with the rule-based models?

Yes, as long as it was run over the same conversions and period. It is often the most useful view and the hardest to explain, so keep a rule-based model alongside it.

Next: how a measurement team uses attribution, lift and identity together.

See it on your own campaigns

Bring one quarter of campaign data

We'll walk through the command center, attribution and recommendations with your channels and your team, as a hosted trial or in your own cloud.

Request a demo