Why last-click undervalues the upper funnel
The last-click attribution model is simple, and therefore widespread: the whole value of an order goes to the source of the final touch. The problem is that a purchase in e-commerce rarely happens in a single session. A typical chain looks like this:
Display ad → blog article from organic search → social →
brand search → direct visit → purchase
In a last-click model all the revenue goes to the direct visit or the brand search. Display, the blog and social get zero. If budget decisions are made from that report, the first channels to be cut are exactly the ones that create demand, and the remaining budget flows into brand search — that is, into harvesting demand that already exists and would largely have arrived anyway.
Assisted conversions are a way of seeing the missing part of the picture: how many times a channel took part in chains that ended in a purchase without being the last touch.
How to read the report
An assisted conversions report usually gives three numbers per channel: assisted conversions, direct conversions and the ratio between them. An example of the structure (figures are illustrative, to show the logic):
| Channel | Assisted | Direct (last-click) | Ratio | Role |
|---|---|---|---|---|
| Display | 410 | 60 | 6.8 | Opens the chain |
| Social | 320 | 90 | 3.6 | Prepares the purchase |
| Organic (informational queries) | 280 | 130 | 2.2 | Prepares the purchase |
| Paid search (category queries) | 190 | 210 | 0.9 | Mixed role |
| Paid search (brand queries) | 120 | 340 | 0.35 | Closes |
| Direct visits | 90 | 480 | 0.19 | Closes |
How to read it:
- A ratio well above 1 — the channel nearly always sits at the start of the path. Judging it on last-click ROAS is pointless: that ROAS will always look bad.
- A ratio around 1 — the channel opens and closes chains about equally often. Usually this is category paid search and remarketing.
- A ratio well below 1 — the channel harvests demand that already exists. It looks like the most efficient of all, but it largely inherits the work of earlier touches.
Assisted and direct conversions must never be added together. A single order produces one direct and several assisted conversions, so the sum across all channels comes out several times larger than the real number of orders. This is a metric for comparing the roles of channels, not for counting results.
How this changes budget decisions
The practical effect of the report is not a reallocation of budget “by formula” but a change in the decision procedure.
- A channel is not switched off on last-click ROAS alone. First you look at the ratio of assisted to direct and the share of chains in which it is the only non-converting touch.
- The upper funnel is judged on its own metrics — reach of new audiences, share of new users, entries into the chain — not on direct revenue.
- The attribution model changes. Moving from last-click to linear, position-based or data-driven redistributes revenue and usually inverts the profitability picture.
- The attribution window gets checked. If it is shorter than the real decision cycle — weeks in furniture, appliances and DIY — the upper touches simply never make it into the report.
- The hypothesis is tested by switching off. A controlled pause of the channel in some regions, compared against a control group, is the only way to learn what actually happens to revenue.
The same thing inside the site: personalization
The logic of assisted conversions transfers completely to on-site mechanics, and here it is ignored more often than in advertising.
The typical scenario: a user sees a product in a personal recommendation block on the product page, does not click, or clicks and does not buy, and two days later comes back, finds that product through search and places an order. Click-based attribution credits the order to search. The recommendation block looks weak in the report, even though it was what put the product in front of the shopper.
The opposite mistake is just as common: crediting the block with all the revenue of sessions in which it was clicked. Some of those purchases would have happened without the block — the person was already looking for that product.
This is why the contribution of personalization mechanics is not measured by attribution:
| Method | What it shows | Fit for a decision |
|---|---|---|
| Widget click → order in the session | Correlation, biased upwards | Operational monitoring only |
| Assisted conversions of the widget | The presence of a supporting role | Shows the effect is wider than the click |
| A/B test with a control group | Causal revenue uplift | A basis for a decision |
| Holdout group over a long horizon | Accumulated effect, including delayed | A basis for judging the tool overall |
That is precisely why a correctly designed experiment compares not “clickers versus non-clickers” but the whole test group against the whole control group, including everyone who never noticed the widget. The difference between the groups in revenue per visitor is the real effect, and assisted conversions are already included in it automatically.
Common mistakes
- Comparing channel ROAS figures calculated in different attribution models. Numbers from the ad platform and from web analytics almost never match, because the models and windows differ.
- Treating assisted conversions as proof of contribution. They are a signal to run a check, not the result of one.
- Forgetting cross-device behaviour. Without identity stitching the chain breaks and part of the assisting touches is simply invisible — more on this in end-to-end analytics.
- Judging the upper funnel on a short window. With a 7-day window, in categories with a long decision cycle most first touches never survive to the order in the report.
- Switching a channel off and not measuring the consequences. If overall revenue did not fall after the pause, confirm that by comparison with a control rather than by timing coincidence.