Most marketing budgets are still judged on the last click, even though customers now meet a brand on TV, in search, on social, through affiliates and in AI assistants before they buy. This article explains where each touchpoint sits in the journey, why a single channel rarely wins alone, which traps make naive analysis mislead, and how a probabilistic model measures online, offline and AI touchpoints together.
Where each touchpoint sits in the 2026 customer journey, including AI assistants; why channels win as a team; three traps that make naive analysis mislead; how a probabilistic model measures online, offline and AI touchpoints; how it compares with media mix modelling; which channels it covers; and how to put it into production.
Your customers meet you in more places than your reports show
Before and after a purchase, customers pass through several phases. Each phase has its own touchpoints, and many of the important ones never produce a click: a TV spot, an out-of-home (OOH) poster on the way to work, a catalogue on the kitchen table, a conversation in the store. And since 2024 a new one has joined them.
- Linear TV & Connected TV (CTV)
- Radio
- Podcasts & audio
- Out-of-home (OOH) & digital out-of-home (DOOH)
- PR
- Social & creators
- Word of mouth
- AI assistantsNew
- AI search & AI answersNew
- Search
- Price comparison sites
- Websites & reviews
- Catalogue
- Personal recommendations
- Online shop & retail media
- App
- Store / branch
- Sales staff
- Phone & direct contact
- Customer service
- Newsletter & CRM
- App notifications
- Complaints handling
- Reviews
- Loyalty programmes
- Surveys & feedback
Goals are scored by teams
When a striker scores, the scoreboard shows one name. But the goal started with the defender who won the ball and the midfielder who played the killer pass. Last-click attribution is the scoreboard: it credits whoever touched the ball last. Our model works like modern sports analytics. It compares the moves that led to a goal with the moves that didn't, and credits every player for what they actually added. The result: you stop cutting your playmakers and overpaying for tap-ins.
TV, radio, out-of-home (OOH), display. They win the ball and create the demand everything else builds on.
Price comparison sites, generic search, catalogue, social, AI assistants and AI search. They set up the chance.
Brand search, direct, email click. They finish what the others prepared.
One channel doesn't win the game. Combinations of channels do.
Why naive analysis misleads
- aEvening conversion peaks can be almost entirely driven by brand search. Remove it and the effect disappears.
- b“Cookie bombing”: some networks serve large volumes of cheap, barely visible impressions. These appear as the last “view” before many purchases and collect credit under last-touch, without influencing anyone. A data-driven model exposes this.
- cDirect visits get no credit: in Google Analytics, an order that follows a direct visit is credited to the previous click, very often a paid search ad, unless there was no other touchpoint at all. The same provider sells that search ad. Relying on its measurement alone lets the seller grade its own homework.
All of these rule-based methods look only at journeys that converted, and only at clicks. They're hypotheses, not analysis.
How the attribution model works
User-level attribution, including offline channels and AI assistants
Learns from your data
A probabilistic model compares journeys that converted with journeys that didn't, estimates how much each channel really raises the probability of conversion, and splits each converting user's credit across the touchpoints they saw.
Combinations, not single channels
Journey combinations with a higher response rate weigh more, so you see which channel pairs work as a team, for example TV and brand search, or catalogue and store.
Stable and cross-checked
The probabilistic model is calculated directly from your journeys, not trained, so there is nothing to overfit. We recalculate it on many resampled sets of journeys, measure how much each channel's credit moves, and check it against the fitted regression model. A KPI has to be reproducible.
Rich touchpoint features
Beyond channel presence: frequency of exposure, creative, placement or site category, time of day and week, and recency.
Complete journeys
Better cross-session recognition stitches returning visitors into longer journeys, so early touchpoints aren't lost.
Your business rules on top
A post-processing layer applies your rules, such as excluding service contacts, capping brand search, or handling returns and cancellations.
Report on any dimension
Credit is assigned per user and per touchpoint, so it rolls up by channel, network, campaign, creative, funnel phase, region or segment. Media mix models can't do this.
Built for scale
Designed for billions of impressions and tens of millions of users, running in production on your cloud.
Beyond the click: offline channels and AI assistants
This is where we go further than purely digital attribution. Channels that never produce a click are brought in at the user level.
TV and radio
Spot-level response: the traffic and conversion uplift in the minutes after each airing, against a baseline, assigned probabilistically to the journeys that were active.
OOH, print and catalogues
Regional exposure plus voucher and promo codes, catalogue and mailing IDs and QR codes, matched to journeys.
Stores and call centres
Offline sales and calls joined to online journeys through loyalty IDs, call tracking or hashed identifiers, privacy-safe and GDPR-compliant.
AI assistants
Identifiable AI referrals as their own channel, post-purchase survey answers, and the lift in brand search and direct traffic.
Same journeys, two views
Illustrative numbers. Last touch undervalues display, overcredits Network A (see cookie bombing above) and brand search, cannot see TV at all, and books AI assistants as direct traffic.
Where media mix modelling fits
Media mix modelling is a valid tool, and we use its strengths. The difference is the level of detail it can see.
| Rule-based attribution last click, linear, etc. | Media mix model | ProDataAI user-level MTA + cross-media optimisation | |
|---|---|---|---|
| Uses non-converting journeys | No | Not applicable (aggregate data) | Yes |
| Learns from your data | No | Yes | Yes |
| User-level journey insight | Partly | No, aggregated weekly channel data | Yes |
| Measures how channels work together | No | Usually not: each channel is modelled separately and added up | Yes: channel combinations and cross-channel effects |
| Any reporting dimension (network, creative, segment) | Partly | No | Yes |
| Covers offline channels | No | Yes | Yes |
| Forward-looking budget recommendation with scenarios | No | Yes | Yes |
Media mix models work on aggregated weekly data and treat each channel separately. We work at the journey level and measure how channels amplify each other.
We combine the strengths of both, an approach now often called unified marketing measurement: the offline coverage and forecasting of a mix model with the journey-level detail of attribution.
Every channel you spend money on
Online
- SEO and paid search (SEA), brand and generic
- AI assistants and AI searchNew
- Display networks and trading desks
- Social media and creators
- Online video and Connected TV (CTV)
- Retail media
- Affiliate and price comparison sites
- Newsletter and CRM email
- Direct and app
Offline
- Linear TV
- Radio and audio
- Out-of-home (OOH) and digital out-of-home (DOOH)
- Catalogues
- Direct mail
- Call centre and outbound calls
- Retail stores
- Events and trade fairs
How ProDataAI can help
We have built and run this kind of attribution in production for advertisers in retail and fashion, telecom, travel and online marketplaces. The first conversation is free: tell us your channel mix and how you measure today, and we tell you honestly whether data-driven attribution would change your budget decisions.
See it applied to your channels. Our Multi-Touch Attribution & Budget Optimisation solution puts this into production on your data and turns it into a budget recommendation. Explore the solution →