Conversion & Checkout Funnel Radar · Brief 01
The agent funnel converts at twice the site — and leaks at one step
All departments · March 2026 · session origin → conversion → where the cart is lost
The one line you configure
Sessions arriving from an AI agent are a new front door. Do they convert differently from the site's own traffic — and if they do, is anything downstream failing to keep up?
In a nutshell
They convert far better and they leak in one identifiable place. Gemini-originated sessions convert at 3.4% against a ~1.6% seasonal baseline — 2.1× — but on the same cohort the cart-abandon rate runs 1.45× baseline wherever the shipping ETA is over three days, which is a threshold, not a trend. Separately, ChatGPT Atlas sessions carry 4.6× the mobile-web order value while their impressions sit at 0.80×, because the spend cap on that surface has never been raised.
At a glance
What moved
What · conversion by where the session came from
Two agent surfaces, two different shapes. Gemini brings guests who convert; Atlas brings guests who spend. Read on conversion alone Atlas looks ordinary, which is exactly why its impression cap has never been questioned.
Where it leaks · the abandon is a shipping-ETA threshold, not a trend
Below three days the cohort abandons at its normal rate. Above it, 45% more of them leave the cart. The break is at the promise, not at the price — which makes it an ETA-logic fix on long-lead items rather than anything to do with the offer.
Why · what the conversion move decomposes into
| Dimension | Top contributor | Share of the move |
|---|---|---|
| Campaign objective | Conversion | 51% |
| Placement | Offsite_Display — the Gemini surface | 42% |
| Department | Home | 33% |
| Audience segment | Trend_Forward | 28% |
Contribution analysis of the conversion rate, one value per dimension. It concentrates on conversion-objective campaigns served offsite into Home — and the merchandising side agrees: long-tail Home on Ship-to-Home sold 22% above its seasonal forecast the same month.
So what
Two levers, both cheap, both in the same month's data. Raise the Atlas placement cap, because a surface delivering 4.6× the order value is being served 20% fewer impressions than it could be. And fix the ETA logic on items that default to more than three days' shipping, because that single threshold is where a funnel converting at twice the site rate gives the guests back.
The points that matter
The front door is already different
Gemini sessions convert at 3.4% against a 1.6% baseline. That is not a campaign result, it is a different kind of arriving intent — and it is being measured against a site-wide baseline that no longer describes it.
2.1×the site rateThe leak has a threshold, so it has a fix
Abandons rise 45% only where the shipping ETA passes three days. A leak with an edge that sharp is a logic change, not a programme.
1.45×over 3 daysThe best surface is the one being capped
Atlas carries 4.6× the order value of mobile web and receives 0.80× the impressions, because the spend cap was set before the surface existed.
0.80×impressionsRecommendation
Lift the ChatGPT Atlas paid-placement cap, and tune the ETA shown on items with a default ship time over three days before the agent funnel scales further — both are decisions about this quarter's plan, not next year's.
Questions it already answers
Is the agent funnel actually incremental, or is it re-badging traffic we already had?
Incremental on the merchandising side: long-tail Home SKUs on Ship-to-Home — Outdoor Living, Storage, Decor — sold 22% above their seasonal forecast the same month, which is stock the site's own browse path was not surfacing.
Where exactly are the carts being lost?
At the shipping promise. The cohort abandons at its normal rate below a three-day ETA and 45% above it beyond three days, with no equivalent break by department, price or placement.
Why is Atlas not showing up as a top performer?
Because it is read on conversion, where it sits on baseline. Its edge is order value — 4.6× mobile web — and the impression cap it inherited means the surface never gets enough volume for that to register.