The Marketing Intelligence Layer
An AI-operated marketing system I build inside your business. It connects every platform you spend or measure on, reads them live the moment you ask, and hands back finished work: answers, audits, forecasts, media-mix models, reports, and fully built campaigns. You sign every change.
01 // Connect
Every connector is built on the same pattern: a live read, a data-health check, and a standing audit. Because the pattern is fixed, any platform with an API is connected, tested, and reporting within a day. On day one you get a verified connection, a health report, and a first audit of the account.
It is attribution-agnostic: whichever tool you trust plugs in the same way, and every analysis downstream works unchanged. Nothing is copied into a warehouse. The question goes to the source, so restated conversions and late-arriving orders are always in the answer.
Search
Social
Programmatic
Measure
Attribution
Lifecycle
Affiliate
02 // Run
Ask
Mention the layer in a private Slack channel and ask in plain language. Why did organic sessions drop on one brand last week? How does this month’s cost per new customer on Meta compare to last month’s? Build me last week’s paid-media summary. The answer is computed from live data across every platform and every brand, and comes back as text, a chart, a PDF, or a workbook. Anyone on the team can ask, not only the analyst.
Watch
Standing audits score every account against a best-practice checklist: searches landing in the wrong campaign, spend with no return, efficient campaigns starved by budget caps, duplicate conversion actions inflating results, placements that absorb money without orders, ads missing their tracking tags, analytics events that quietly stopped firing. Every finding ships with the file that fixes it. Anomalies land in Slack mid-week, and every alarm is checked against reporting lag and an independent signal before anyone is told something broke.
Decide
A monthly reallocation engine moves budget toward the sources that bring in new customers worth the most over their lifetime, inside hard guardrails: protected brand campaigns, a minimum return on acquisition cost, and limits on how far any line moves in one cycle. Kill criteria are written before the first dollar. Every idea runs as an experiment with halt rules set in advance; the winner rolls out, the loser retires. A quarterly growth review steers the next round.
Act
The layer writes back to every platform’s API: bids, budgets, keywords, negatives, targeting, audiences, ad copy, and tracking tags. Every change is previewed as a dry run, refuses to overwrite anything a person changed since the plan was made, and saves a rollback file before the first write. Bulk work goes out through the ads editors, and anything without an API is driven through the platform’s own interface. Nothing that moves money or changes live copy goes out without a person’s sign-off.
Cadence //
03 // Measure
Most marketing reports quietly disagree with one another. The layer settles that before anything else, with house rules written into the code instead of left to memory.
Attribution is the floor. Your attribution tool’s click credit is each channel’s minimum value. What a platform reports about itself is a health check for its bidder, never revenue truth.
Cost comes from the platform. Spend is read from the ad platform itself. No channel gets a cost per customer or a return on spend without it.
Clicks, not views, for social. View-through credit stays out of social reporting unless you ask for it.
Graded when the sale lands. Fresh campaigns are judged on conversion date, not click date, so recent winners aren’t mistaken for losers.
Three lenses before a verdict. Linear, first-click, and platform numbers are reconciled before any conclusion is drawn.
One alarm is never a conclusion. A drop is checked against reporting lag and an independent signal before anyone hears about it.
Media-mix model //
Clicks can’t see what video, upper-funnel social, and brand lift are worth. The model can: weekly, across years of every channel’s spend, with market conditions, holidays, and every promotion as controls. Two independent modeling libraries are fit to the same data and their answers triangulated.
It only ever adds to attribution. It never lowers a channel’s click credit, and a channel the data can’t move is flagged, not quoted. Where national data can’t read a channel, the layer designs the experiment that can: a randomized geo holdout, sized in advance, whose result then calibrates the model.
04 // Ask for
05 // Build
The layer researches demand, competitors, and coverage gaps; writes the strategy in new customers and revenue; lays out the structure and naming; builds keyword packs, negatives, and audiences; writes every ad; produces the creative; launches across every platform; verifies the tracking; then reads results at the right maturity date and keeps optimizing.
Hundreds of ads written or rewritten in one pass. Duplicates caught by word stem, casing and compliance rules enforced, every headline slot filled, and ads updated in place so they keep their history.
Scripts written, images and video generated, voice-over added, and finished cuts rendered in every aspect ratio each platform asks for.
A sale goes live across search, shopping, social, video, and email in one build, and comes down just as cleanly the moment it ends.
06 // Remember
Every decision, business rule, prior result, and lesson learned is written into a structured knowledge base inside the project: what your goals are, which campaigns are off-limits, why a change was made, what went wrong last time.
Around it sits the discipline that keeps it trustworthy: a register of every capability, a log of every decision, and a library of every platform quirk ever found, so a trap is only fallen into once. Any procedure run twice becomes a script, so the third run takes minutes.
Each new session starts with full context instead of a blank slate. Over time it becomes the institutional memory of your marketing, and it stays with you when people don’t.
Read at the start of every session. Written at the end of every one.
07 // Report
Recurring reports regenerate on their own: living workbooks updated without losing a single formula, chart packs, goal scorecards, campaign dashboards, executive PDFs, and decks where a verifier recomputes every number on every slide before you see it.
Four of them are below, redrawn with the numbers and names removed. What survives is the shape of the insight: a target tracked to the week, a customer base cross-checked across sources, every account change marked on the spend it changed, and a quarter reviewed channel by channel.
PF // Weekly executive brief
ON TRACKPF // Customer base report
CROSS-CHECKEDPF // Account changes & impact
SIGNEDPF // Quarterly growth review
COMPOUNDING08 // Trust
0
copies stored
Nothing is copied into a warehouse. Every figure is read from the source at the moment of the question.
0
customer contact details read
Personal data is stripped inside the connector before anything leaves the source, or access to it is never granted at all.
100%
of writes logged and reversible
Previewed as a dry run, guarded against overwriting a newer change, and undoable from a rollback file.
1
human sign-off, every change
Recommendations are automatic. Pushing them live never is. A person approves each one first.
▸ READ-ONLY UNTIL A WRITE IS APPROVED · CREDENTIALS NEVER IN THE CODE · NO NUMBER RELEASED UNTIL ITS VERIFIER PASSES
09 // Start
You provide read access to each platform, your definitions (what counts as a new customer, which attribution model you trust, gross or net revenue, your goals), and one owner who approves every write. The rest is mine.
Day one
Week one
Month one
Built inside your business. Yours to keep.
Two emails and one call. I’ll tell you honestly whether the layer would pay for itself in your business, and how fast.