THE WEEKLY REPORT
A decision sheet a CMO can forward. An appendix the analyst can rebuild.
The PDF is the product. It is not a dashboard, not a slide dump, and not a model notebook. Six models run on the same journeys. A rubric ranks them. Every file is approved by a marketing professional with extensive experience in marketing analytics before it is sent.
Comparison engine output: last-click, first-click, linear, time-decay, Markov chains, and Shapley values, scored for whether the numbers add up, whether this week looks like a pattern, and whether the ranking would actually change a budget meeting. Simple rules bound the story. Path models argue from the journeys in BigQuery.
What the report is for
Whether an assist is being missed
First-click and Shapley values often lift Email, Organic, and Referral against last-click. The decision sheet says whether that lift is large enough — and stable enough — to stop cutting those channels on last-touch ROAS.
Whether a closer is carrying the story
Last-click loves Direct and branded Paid Search. Markov chains ask if removing them actually drops conversions. If the removal effect is small, they are a closer, not a cause.
Whether this week is real
A share swing with a consent or export flag is data quality. A swing that still looks the same after eight weeks is a pattern. The report will not let those two look the same.
Whether to test, not argue
When last-click and the path models disagree past the threshold for two weeks, page six proposes a geo or pause test. Labeled as a recommendation. The product does not touch the ad account.
01
Decision sheet
Monday-morning moves, not a data dump.
The model recommended for planning this week, the three budget implications that actually change a meeting, and the items held for professional review. One page that forwards.
- Recommended planning model for this week, with the scoring reason in one sentence
- Three budget implications: hold, move, or test — each tied to a named channel
- Items held for review: new channel, a share that jumped, a consent drop, a gap in the export
- A single line a CMO can read without opening the rest of the file
The planning model for the week is chosen by a scoring checklist — do the numbers add up, is this week a pattern, would it change a meeting — not by a language model. Commentary on this page may only cite figures that appear later in the same PDF. If no model beats last-click on being useful, last-click stays the lead and the report says so.
02
Six-model comparison
Last-click and the path models in one grid.
Every planning channel across all six models, with the gap versus last-click. Shares add back up to your conversion count, so finance can trust the arithmetic.
- A channel-by-model grid: last-click, first-click, linear, time-decay, Markov chains, Shapley values
- Gap versus last-click, in percentage points, for every planning channel
- Conversion totals that add back up: credited conversions match the key-event count
- Channels where first-click and last-click disagree — introducers versus closers, named
Shares are computed on the same journeys and the same conversion definition, so disagreements are about the method, not the data prep. The list of planning channels is frozen for the week. Simple rules bound the story; Markov chains and Shapley values are the path estimates the checklist is allowed to promote.
03
Markov chains
See which channels the path actually depends on.
If a channel vanished from converting journeys, how many conversions would you lose? That is the removal effect.
- Removal effect per planning channel, ranked
- The path steps that actually carry converting traffic
- Channels that show up often but do not change how many people convert
- A short note on Direct: closer versus cause, based on the removal, not the last-click share
Built on converting and non-converting paths from the BigQuery export. Removal effect is how many conversions disappear if that channel is taken out of the path. A large last-click share is not treated as load-bearing unless the removal effect agrees.
04
Shapley values
Fair credit, with a range you can stand behind.
Each channel’s share plus a 90% range. Exact when the set is small; an estimate when it is not.
- Share per planning channel, ranked next to last-click
- A range on every share — width is a feature, not a footnote
- Exact versus estimated, stated, when the set of channels is large
- How many planning channels were in the mix, kept stable week to week
Fair split of credit across every mix of channels on converting paths. Exact when there are eight channels or fewer; otherwise an estimate with a range. This is path-based fairness, not a regional test. When the range overlaps last-click, no move is recommended.
05
Stability
Know if this week is a pattern or a blip.
How ranks have moved versus the last eight weeks. A channel whose share jumps is flagged, so you do not reallocate on noise.
- How this week’s ranking compares to the last eight weeks
- Channels that jumped in share, with the previous rank
- Consent-tracking and export-freshness flags that would explain a swing
- A hold recommendation when the pattern is too young to act on
Compares this week’s ranks to the last eight weeks, plus a per-channel swing threshold. A swing that lands with a consent or export gap is labeled data-quality, not a budget signal. A model that looks tidy but failed stability does not lead the PDF.
06
A test worth running
When the models disagree, a next step — not a debate.
If the gap is expensive enough, a geo or pause experiment. Labeled as a recommendation. The product does not bid, pause, or publish tags.
- The disagreement that is expensive enough to be worth a test (channel pair + models)
- A proposed design: a regional split, a pause cluster, or a holdout — with a suggested window
- What would count as confirmation, in one sentence
- Explicit non-actions: no bidding, pausing, or tag-manager writes from this page
A test is proposed only when the last-click versus path-model gap stays past the threshold for two weeks and the range does not overlap. This is not a mix model and not a causal claim. It is a cheap check on an expensive argument.
How a run is produced
01
Journeys in your project
Paths are built in your BigQuery. Event-level rows do not leave.
SQL over the GA4 export, with a cap on how much is scanned. Read-only access on that dataset only. Conversion definition, how far back a journey is counted, and the channel list are your settings.
02
Six models, in parallel
Simple rules and path models run on the same journeys, the same week.
Last-click, first-click, linear, time-decay, Markov chains, Shapley values. Jobs on AWS. Simple rules bound the story; path models argue from how people actually move.
03
Scored, then drafted
A rubric ranks the models. A language model writes the sentences.
Whether the numbers add up, whether this week is a pattern, whether it would change a meeting — numbers first. The language model may only mention figures already in the file. A check rejects unseen figures before professional review. The system is built that way.
04
Professional review, then the inbox
Every PDF is approved by a marketing professional with extensive experience in marketing analytics. A ranked PDF, not a login.
Items held for review: new channel, share swing, consent drop, export gap. Professional approval before send. Delivery over email or Slack. Monthly 30-minute readout on the retainer.
Why teams keep the PDF
One PDF. Not forty tabs.
The product is a document you can send to a CMO. Commentary is checked against the tables in the same file. Unseen figures do not ship.
Approved before send.
An agent drafts. A marketing professional with extensive experience in marketing analytics approves, edits, or holds. Target review time is under 20 minutes.
Last-click stays on the page.
Everyone already has that number. Printing it next to Markov chains and Shapley values is how a budget meeting moves, instead of stalling on a new methodology.
Your definition of a conversion.
Lookback, channel list, and the key event are yours. The appendix states them every week, with a data-quality score and how much BigQuery was used for the run.
Appendix, every week
So the analyst can rebuild the number without a call.
- Conversion event, how far back a journey is counted, and the channel list used for planning
- How many planning channels were in the mix, so Shapley values stay comparable week to week
- How fast older touches fade, if time-decay is in the grid
- Data-quality score: freshness, consent tracking, export gaps
- How much BigQuery was used for that run
- Whether the Shapley split was exact or estimated
- Commentary is checked against the tables; unseen figures do not ship
- Items held for a person before send