Overview

01 · Activation

Social Proof

02 · Substrate

Demand Intelligence Cohorts

03 · Engine

Engine Signal AI

04 · Measurement

Attribution Portal & Analytics

06 · Integrations

Integrations overview Commerce platforms Customer Engagement Customer data platforms Ad platforms Data warehouse Custom & developer

Pricing

Contact

Book a demo

The measurement layer

Defensible by construction. Honest about its limits.

Three lift measurements, computed from the same session-level substrate, derived from real purchase events. The model is observational — not a randomised experiment — and the page reports it that way.

Two-group modelProduct-scopedLive mode required

Measurement model

attribution.v2· two-group session model30d window

01Session substrate

PDprimary

12,000

BDbaseline

50,000

Compute lift

02Conversion rate lift

Primary4.0%

Baseline1.5%

Lift+167%

Monetise

03Incremental revenue

~£25,500/ mo

From300 incremental orders·£85 AOV

The four counts

Four counts. Two splits.

Sessions sort into one of two cohorts. Orders sort by whether the purchased product matches the clicked product.

Sessions in the date rangeAll sessions

bots filtered

splits into

PDPrimary

Flockr click sessions

Sessions where any product was clicked with a Flockr message visible on a browse surface.

PLP · search · predictive search · recommendation

BDBaseline

All other sessions

Every other session in the same period.

Mutually exclusive with PD

Orders in the date rangeAll orders (deduped)

product-scope test

splits into

PNPrimary order

Same product, same click

The purchased product equals the product Flockr-clicked on browse in this session.

Subset of PD-session orders

BNBaseline order

Everything else

Including “wrong-product” purchases by primary sessions.

Catches the over-claim cases

Why split orders separately

A primary session that clicked product A on the PLP and then bought product B is in PD, but the order is BN. This stops Flockr from claiming credit for purchases it didn’t influence.

Three lifts

Three lifts. One substrate.

All derived from the same session data. None of them collapses into the others — each tells a different story about Flockr’s influence.

1Session level

Conversion rate lift

How much more often does a Flockr-influenced session end in an order, vs. an unexposed session?

Formula

(PN/PD−BN/BD)÷(BN/BD)

Worked example+167%

Headline figure on the portal page

2Product impression level

Product CTR lift

For the same product, how much more often is its tile clicked when a Flockr message is visible vs. not?

Paired comparison

with-message CTRvswithout-message CTR

tile in viewport— both sides
same product set— appeared in both conditions

Like-for-like, controls for product appeal

3Page view level

Page CTR lift

For a given browse page view, how often does it end in clicking a Flockr-influenced product vs. a non-Flockr product?

Share comparison

Are Flockr products winning the click?

None of them collapses into the others.

The full derivation

Six steps. End to end.

From PD/BD/PN/BN to incremental revenue per month. Worked with illustrative round numbers — same example that flows through the hero.

attribution.walkthrough· worked example30-day period · AOV £85

Primary order rateorders ÷ sessions

PN÷PD=480÷12,000

4.0%order rate

Baseline order rateorders ÷ sessions

BN÷BD=750÷50,000

1.5%order rate

Conversion rate liftprimary vs baseline

(4.0%−1.5%)÷1.5%=2.5%÷1.5%

+167%lift over baseline

Incrementality multiplier1 − (BN/BD ÷ PN/PD)

1−(1.5%÷4.0%)=1−0.375

0.62562.5% of primary

What this is: the fraction of primary orders that are extra — above what the baseline rate would have produced anyway.

Incremental orders & revenuemonetise the multiplier

orders =480×0.625=300 orders

revenue =300×£85 AOV=£25,500

£25,500period total · 30 days

Monthly run rateextrapolate to 30 days

(£25,500÷30 days)×30

£25,500/ mo · estimate

Why a multiplier, not all primary orders

Some of the orders in the primary group would have happened at the baseline rate anyway. The multiplier credits only the share above baseline — the implicit Flockr contribution.

Methodological locks

Two locks. No over-claim.

Two constraints that prevent the model from claiming credit for things it didn’t influence.

01· Product-scoped

The clicked product must equal the purchased product

A primary-group session that clicked product A on the PLP and then bought product B is still in PD — but the order goes into BN, not PN. Without this constraint, the model would over-claim every purchase made by a Flockr-exposed shopper.

Effect — in practice

clickedA· boughtA→PN✓

clickedA· boughtB→BN

02· Browse surfaces only

PDP, cart and checkout don’t qualify a session for primary

Browse surfaces are where Flockr influences product selection. Downstream surfaces confirm decisions already made. Restricting attribution to browse makes “Flockr was involved in the choice” the explicit claim.

Effect — in practice

Qualifying

PLPsearchpredictiverecommendation

Not qualifying

PDPcartcart drawercheckout

Operational layer

Built to stay honest.

Three operational properties that keep the model defensible at run-time — not just on paper.

01· Bot filtering

Excluded from both groups

Sessions classified as bots are dropped before the model runs — they don’t inflate the baseline or the primary count. Detection runs against the standard exclusion list maintained for the analytics pipeline.

BotsReal sessions · primary + baseline

Same exclusion list as analytics

02· Paired comparison

Like-for-like for Product CTR

Product CTR lift only includes products that appeared in both conditions — with and without a Flockr message. Same product, both genuinely seen.

same product settile in viewportboth conditions

Controls for product appeal

03· Live mode

Required for attribution figures

Attribution requires Flockr in live mode — rendering messages, firing visibility events, recording browse-surface clicks. In data mode the conversion fields are empty by design, not misrepresented.

Liveattributionpopulates

Dataattributionempty by design

Honest at run-time. Not just on paper.

The empirical anchor

Every engagement starts with a holdout.

Every new client begins with a true 50/50 holdout test. Half of shoppers see Flockr, half don’t — run to statistical significance, reported as a full causal picture. The observational model is what runs after. The randomised test is what anchors it.

holdout.experiment· 50/50 splitEvery new client

1Setup

50/50 random split

Every shopper deterministically assigned at session start. Half see Flockr — messages render, signals act. Half don’t — genuine holdout, never exposed to a Flockr message during the test.

50%Holdout

50%Treatment

2Run

Until statistically significant

Duration isn’t fixed — the test runs until lift figures clear a p < 0.05 significance threshold. Larger catalogues hit significance faster; smaller stores take longer. No arbitrary stop dates.

Confidence95%· significant

50%p < 0.0599%

3Report

Full causal report

Not a single number. The result returns the full causal picture — conversion lift, revenue per visitor, AOV impact, and more — each with the confidence interval and significance flag attached.

Conv. rate lift+ 41%sig

Rev / visitor+ 38%sig

AOV impact+ 4%sig

After significance

Option AFold the holdout into live — full traffic on Flockr from there.

Option BKeep the holdout running — ongoing causal truth, calibrating live figures continuously.

Observational from there. Anchored from start.

Honest about what this is

The ongoing model is observational. The anchor is randomised.

The 50/50 onboarding holdout is the empirical anchor — that’s the causal truth. Once it ends and the holdout folds into live (Option A), steady-state attribution returns to observational measurement.

The primary group is then self-selecting — clickers are higher-intent than average shoppers. Some share of ongoing reported lift is intent-driven, not Flockr-driven. The product-scoped constraint partially compensates by stopping the model from claiming downstream unrelated purchases. It does not eliminate it.

Ongoing lift figures should be read as commercial-grade estimates anchored to the empirical holdout result — directionally reliable, useful for tracking change over time, but not themselves a randomised result.

Clients who want continuous causal truth keep the holdout running (Option B) — the randomised comparison stays live indefinitely, calibrating attribution against ground truth on a rolling basis.

Attribution in the portal

See it on your data.

Attribution figures live on the Conversion & Attribution page in the Portal — alongside daily charts, message-quality validation, and per-product diagnostics.

Book a discovery call→Or read the pricing

Conversion & Attribution pageLive

From this page4 KPIs

Conversion rate lift

Incremental revenue / mo

Product CTR lift

Page CTR lift

Diagnostics also in portal3 views

Daily orders & revenue split

Message quality validation

Per-product attribution table

Live data ·1-hour cache