CAVEMANAIQ all four storylines
storyline 02 · the auction

Should we believe the platform’s budget advice?

GA
ME
TM
G4
SC
MC
DM
DB
01
01 · one console

Seven dashboards to one console

the console · all seven servers, one login
ask anything…
7 servers docked · each holding its own login
Campaigns, ad groups and advertisements each have their own on and off.
02 · you ask · on the record

The question is the whole interface

“ The platform keeps telling us to raise the budget, and our campaigns seem to contradict each other. What does that recommendation actually mean, and should we believe it?
you
the console
on the record
7 servers docked · each holding its own login
handed to the agent, word for word · nothing runs unasked
▸ we write the request down before we send it
▸ and the answer when it comes back
Nobody could say what the recommendation was optimising for.

Learning how the auction actually behaves

Before touching a budget, the vendor’s own documentation was read on what the automated campaign type does, what it will report and what it will never expose

what had to be understood first

what the advice isa recommendation to raise a budget is a spend projection, not a profit forecast · it answers what more money would buy, never whether it is worth buying
what the automation hidesthe automated campaign type reports at campaign level and does not expose what sits beneath it · so its overlap with a named campaign is invisible from its own report
where overlap showsonly the search-terms report, over a long window, shows one search reaching two campaigns · a short window flatters everything
what a label decidesthe automation can only be divided along product labels that are actually correct, so the labels were verified before anything was split
03 · dispatch

The agent decides who to ask

THE MCP FLEET
THE PLATFORMS · reached by API, not by login
the agent
reasoning: what do I need, in what order?
DB the database 01
what actually sold, same weeks
GA Google Ads 02
what is really running · and the terms people typed
MC Merchant Center 03
which products belong to which lane
02
GA
Ads
spend · conversions
Google Ads API
your ad account · Google’s cloud
ME
Meta
stands down
Meta Marketing API
your ad account · Meta’s cloud
TM
GTM
stands down
GTM API
the tag container running on your site
G4
GA4
stands down
GA4 Data + Admin APIs
your site’s events · browser + server
SC
GSC
stands down
Search Console API
how Google search sees your site
03
MC
MC
products · feed
Merchant Center API
your product feed · Google’s cloud
DM
DM
stands down
audience · conversion upload
your audience lists · Google’s cloud
YOUR OWN RECORDS · READ THROUGH THE DATABASE MCP, NOT A VENDOR CLOUD
01
DB
the database
your own records
the database · your own records
settled orders · matched windows
The account is asked what is actually running, not what it is supposed to be running.
data plane · no human · always running

THE DATA PLANE · runs whether or not anyone is watching
a shopper · not you
the site · code running
every visit writes the record · no one at the wheel
Google’s servers · ads runtime
no change here
Meta’s servers · ads runtime
no change here
04 · gather

The evidence converges on the session

GA
Ads
entity state per level · search terms, long window
MC
MC
product labels, per lane
DB
the database
settled orders · matched windows
one payload · assembling
Ads
entity state per level · search terms, long window
MC
product labels, per lane
the database
settled orders · matched windows
all 3 answers in · ready to combine
The first correction lands immediately: a campaign that reported as live had its only advertisement paused.
05 · the readout

The same searches reached two campaigns

what had to be understood first
Before touching a budget, the vendor’s own documentation was read on what the automated campaign type does, what it will report and what it will never expose
what the advice isa recommendation to raise a budget is a spend projection, not a profit forecast · it answers what more money would buy, never whether it is worth buying
what the automation hidesthe automated campaign type reports at campaign level and does not expose what sits beneath it · so its overlap with a named campaign is invisible from its own report
where overlap showsonly the search-terms report, over a long window, shows one search reaching two campaigns · a short window flatters everything
what a label decidesthe automation can only be divided along product labels that are actually correct, so the labels were verified before anything was split
the overlap finding
built from GA MC DB
the overlaphundreds of searches appeared in both campaign reports over the same long window · the wider-catalogue campaign won most of that spend and returned more on it
people who know youa campaign that catches your own name first · so demand excluded elsewhere has somewhere to go
people browsing productsone automated campaign kept, its learning intact · the products split inside it rather than across new ones
Each kind of demand gets a defined campaign role · profitable searches stay eligible, and exclusions follow the longer-window evidence.
A long window of real search terms showed hundreds of them appearing in both campaign reports, with the wider-catalogue campaign winning most of that spend and returning more on it than the specialist route did.
06 · the loop closes

You decide

you, again
“Yes · apply the bounded set.”
applied · and read back
The name-catcher is genuinely serving. The product lanes are read back item by item. The approved exclusions were applied and checked. Profitable terms remained eligible. The brand campaign turned out not to be short of budget at all · it had hit its bid ceiling. The ceiling was raised first, and the budget was doubled a fortnight later alongside a separate change · and it is still the ceiling, not the budget, that caps what this campaign spends. One category route is paused · and a paused campaign’s past spend is not presented as current activity.
The small bounded set is applied through a preview tied to the exact state a person reviewed · a preview that had gone stale was refused rather than applied · and then what now exists is read back rather than assumed.
07 · how it ended

How it ended

how it ended

a conclusion was corrected
Causal lift: unknown.
The name-catcher and the automation changed in the same window, and an internal source contaminated the labelling of one of them · so the movement is an observation over matched windows, never a proven cause.
It is stated here, beside the numbers it qualifies, rather than at the bottom.