CAIMAN DATA × VOLTAGE
The connector + the operating framework · live demo

The AI does the analysis.
You make the calls. The ledger proves it.

This is the actual operating system we run on Amazon brands — the weekly operator console, the keyword intelligence behind it, and the client report with a measured dollar ledger. Click around. Everything works.

Caiman Data is the connector — first-party Amazon SP-API + Ads data in. Voltage is the operating framework on top — decisions, guardrails, and proof out.
Demo brand: "Apex Brands" (enameled cast-iron cookware) is fictional and every number is illustrative — the mechanics, guardrails, and report formats are the real system.

Monday, 8:04 AM — your console is ready Apex Brands · Week 27

Sunday night, the system read every report on the account and staged this week's moves — each with its reasoning, the data behind it, and the monthly dollar impact. Your job is the judgment: tick, untick, approve. Typically ~20 minutes.

24
reports read overnight (ads, SQP, business, inventory, fees)
41,832
rows analyzed across search terms, placements & queries
13
moves staged — 10 actionable, 1 held for your judgment, 2 sensors
$2,632/mo
at stake across actionable moves, labeled by type
~18 min
of your time. That's the whole job.

This week's moves tick = approve · tap a row for the why

Stop waste
Negate "enamel pot rack" negative exact
61 clicks, $312 spend, 0 orders over 60 days — shoppers want storage racks, not cookware.
$340/mowasted spend saved
why ▾
Evidence
61 clicks / 0 orders · $5.11 avg CPC · appears in 3 broad campaigns · CVR on the account's converting terms is 8.4%, so 61 clicks should have produced ~5 orders.
Action
Negative-exact in the 3 broad campaigns. Variations ("pot rack organizer", "rack for pots") blocked by the same exact negatives only if they also show zero-conversion — each judged separately.
Reversible
Negatives can be removed in one click. The change is logged with its prior state.
Negate "cast iron trivet set" negative exact
47 clicks, 0 orders — accessory intent, wrong product.
$285/mowasted spend saved
why ▾
Evidence
47 clicks / 0 orders · $6.06 avg CPC · trivets are a $12 accessory; our AOV is $89 — intent mismatch, not a bid problem.
Action
Negative-exact. "trivet" as a phrase negative is held — it appears inside 2 converting queries.
Negate "camping dutch oven lid lifter" negative exact
38 clicks, 0 orders — bare-cast-iron camping gear, we sell enameled kitchen cookware.
$122/mowasted spend saved
why ▾
Evidence
38 clicks / 0 orders · the n-gram "camping" is 1-for-167 clicks account-wide (see View 2) — this is one of 4 camping-intent terms staged this month.
"dutch oven" broad — held for a conquest decision your call
The head term converts at 52% ACoS — above target, but it's also 31% of new-to-brand volume. The system never auto-negates head or competitor terms; it surfaces them with the trade-off.
$187/moif trimmed
why ▾
Trade-off
Trim = save ~$187/mo but lose ~9 first-time customers/mo at the top of the funnel. With a 46% margin and repeat behavior unknown on this SKU, this is a judgment call, not a math call — which is why it's held for you.
Options
(a) keep as-is, (b) bid down 20% and watch 2 weeks, (c) move to a capped conquest campaign with its own budget.
Bid & placement
"oval dutch oven" exact: $1.42 → $1.18 −17%
Running 41% ACoS vs 30% target for 3 straight weeks — walk the bid toward target, don't slash it.
$255/mospend efficiency
why ▾
Math
RPC $2.87 × 30% target = $0.86 breakeven-to-target bid; we step −17% (max clamp −25%/week) instead of jumping straight down, protecting rank on a term where we hold organic #6.
Guard
If impressions drop >40% in 7 days, the next console stages a partial walk-back automatically.
"enameled cast iron skillet" exact: $0.98 → $1.14 +16%
Converting at 19% ACoS, well under the 30% target — a starved winner. Feed it.
$175/moincremental sales
why ▾
Evidence
22 orders @ 19% ACoS · impression share only 4.1% (SQP) · lost-to-budget 0% — the constraint is the bid, not the budget. Headroom to target supports +16%.
Inventory gate
Cover check passed: 143 days on the advertised SKU. Scale-ups on thin-cover SKUs get held automatically (see the blue sensor below).
Top-of-search modifier 45% → 25% — "Braisers · Exact" campaign placement
TOS ran 2.8× the ACoS of rest-of-search after the holiday spike ended. The premium stopped paying.
$410/mowasted spend saved
why ▾
Evidence
Placement report, trailing 4 weeks: TOS 68% ACoS / RoS 24% / product pages 31%. TOS CVR fell from 11.2% → 6.4% post-event while CPC held — classic post-spike premium decay.
Action
Modifier 45% → 25% (staged step, not zero). Weekly placement checks re-raise it when TOS CVR recovers above 9%.
Harvest & scale
Harvest "5.5 qt dutch oven red" → dedicated exact harvest
9 orders at 19% ACoS, buried in an auto campaign with no bid control of its own.
$210/moincremental sales
why ▾
Action
Launch single-keyword exact at $1.35 (RPC-derived), add negative-exact in the source auto so the two never compete for the same query. De-duped against all existing exact targets first.
Abort gate
If the new exact shows 0 impressions in 36h, the source negative is auto-reverted — the query never goes dark.
Harvest "bread oven dutch oven" → exact harvest
6 orders, 24% ACoS from phrase — sourdough-driven demand with rising volume (see View 2 whitespace).
$140/moincremental sales
why ▾
Context
Pairs with the whitespace finding on "dutch oven for sourdough bread" — same intent cluster. Listing bullet update also staged (View 2) so paid + organic push the same story.
Brand misspell → brand campaign defend
A recurring misspelling of the brand converts in auto-substitutes; competitors' ads sit on that SERP today.
$95/moincremental sales
why ▾
Evidence
SQP: we win 94% of purchases when shown on the misspell but hold only 6% of its impressions — rivals rent our own shelf. Cheap to reclaim.
Growth — one offensive move per brand, every week
Whitespace: "dutch oven for sourdough bread" — 64K searches/mo, you're in 0.5% of them whitespace
Shoppers already buy us on this query when they see us — they almost never see us. Dedicated exact + a bullet rewrite to index for "sourdough".
$600/moincremental sales (projected)
why ▾
Evidence
Search Query Performance: ~64,200 searches/mo · 0.5% impression share · 11 purchases already at 9.8% purchase share when shown — proven relevance, invisible listing. This is found by a scan that runs every week, not by someone remembering to look.
Label
Projected, not measured — it enters the ledger's estimated column until the 30-day baseline comparison moves it (or doesn't). We don't book wishes.
Sensors — the system watching itself informational · no action needed
Inventory gate held two scale-ups cover gate
APX-DO-6QT-BLU has 19 days of cover with the restock landing ~Jul 21. Its two staged bid-ups were automatically held — scaling ads into a stockout kills rank and momentum.
auto-resumes on PO
why ▾
Rule
Scale-ups are suppressed below 21 days of cover unless inbound stock arrives first. Green-lit SKUs (like the 310-day skillet) get pushed harder instead — spend follows stock.
Fee scan: size-tier overcharge candidate found margin
Amazon's assigned tier on APX-BR-35QT-CRM implies ~$0.68/unit more than its listed dimensions justify. A remeasure/refund case is drafted — filing it is a 2-minute copy-paste.
$96/moif refund lands → measured
why ▾
Loop
The fee scan runs monthly across every SKU. When a filed case lands, the reimbursement shows up in the ledger as measured recovered dollars — not before.
10 of 10 actionable moves approved · $2,632/mo selected impact · all reversible — paused, never archived

Bounded autonomy the part "AI tools" don't have

Routine, low-risk moves can graduate to pre-authorized classes — each with hard caps, floors, cooldowns, and stand-down conditions. Everything else stays preview-then-approve forever. One file on disk kills all autonomy instantly.

CLASS A
Zero-order negates
Proven-waste terms only · competitor & head terms excluded · 60-day evidence window required.
CLASS B
Placement trims
Off-target placements stepped down within ±10% weekly clamps — never zeroed in one move.
CLASS C
Bid walk-downs
Chronic over-target keywords walked toward RPC-derived bids · 14-day cooldown between touches.
CLASS D
Budget re-caps
Data-gated — fires only where spend pacing data is complete. Otherwise: console.
CLASS E
Winner bumps
Under-target converters nudged up · inventory cover-gate checked before every single move.
This week: 5 moves fired in-class ($3.1K/mo), 4 overflowed to this console, 6 stand-downs — the rails refused two brands with flagged audits and one with stale data. A system that knows when not to act is the whole point.

Whitespace — demand you're invisible on Search Query Performance · scanned weekly

First-party Brand Analytics data (no scraping): every query with real volume where shoppers already buy you when shown — but you're barely shown. Proven relevance × missing impressions = the highest-confidence growth there is.

QuerySearches/moYour impression sharePurchases alreadyVerdict
dutch oven for sourdough bread64,2000.5%11Launch exact + listing
enameled dutch oven 5 qt22,3003.0%17Fund — page-1 breakout
braiser pan with lid9,8502.4%8Scale — proven converter
dutch oven wedding gift4,1001.1%5Seasonal — stage for Oct
colorful cookware set18,7001.8%3Watch — relevance thin

Paid → organic handoff 8-week share trend per query

When a keyword's organic position holds on its own for 4+ weeks, paid eases off and the budget moves to the next target. That's how ad-dependence falls without sales falling with it.

QueryShare trend (8 wk)HeldPaid stateAction
oval dutch oven2.1% → 6.8% ▲5 wkexact $1.18 · TOS 25%Ease paid −20%, redeploy
enameled cast iron skillet1.9% → 4.1% ▲2 wkexact $1.14Keep funding — not held yet
braiser5.2% → 5.0% —7 wkexact $0.92Ease paid, watch 2 wk
dutch oven 7 qt3.8% → 2.2% ▼exact $1.05Push — losing shelf

Where the money leaks — word by word n-gram analysis · every search term, not the top 20

Every customer search term broken into word stems, aggregated account-wide. Words that spend without selling get caught even when no single term is big enough to notice on its own.

"rack" — 0 orders / 214 clicks · $1,090 spend "camping" — 1 / 167 · $840 "lid lifter" — 0 / 61 · $310 "replacement knob" — 0 / 44 · $212 "sourdough" — 31 / 380 · 18% ACoS ← feed this "5.5 qt" — 24 / 296 · 21% ACoS

Each flagged stem becomes staged negatives in View 1 — with head terms and competitor names always held for a human call, never auto-blocked.

Why this data is different first-party, not scraped

SOURCE
Amazon's own APIs
SP-API + Ads API via the Caiman connector — Search Query Performance, search terms, placements, business reports, inventory, fees. No browser plugins, no stale scrapes, no third-party estimates of your own account.
JOIN
Every report, one decision
A bid change checks SQP share, RPC, placement efficiency, inventory cover, and margin in the same pass. Tools that read one report at a time recommend one-report decisions.
CADENCE
Weekly loop, monthly & quarterly sensors
The weekly console carries the account. Fee scans run monthly, co-purchase & repeat-behavior refresh quarterly, and everything lands in the same queue — nothing depends on someone remembering.

Your Monday note plain language · every week

Five lines, no jargon, drafted from the same logs the system executes from — what happened, what it was worth, what we're watching, and what we need from you.

APEX BRANDS — week of Jun 22 Did: 11 changes applied — 4 waste cuts, 3 bid walks, 2 harvests, 1 placement trim, 1 listing update. Impact: $1,184/mo of measured waste removed to date (30-day verified); $2,632/mo staged this week, pending your console. Watching: the 6-qt blue dutch oven — 19 days of cover, restock lands ~Jul 21; its scale-ups are held until then. Need from you: one decision — the "dutch oven" head-term conquest question (options in the console, 2 minutes). Board: your live dashboard link · full change log with reasoning attached.

The realized-$ ledger measured ≠ promised

Every applied change gets a baseline captured before it fires. Thirty days later it's measured against that baseline. Only then does a dollar enter the measured column. Stale and superseded recommendations are excluded — the ledger never counts ghosts.

Measured · verified vs baseline
$1,184/mo
Waste removed & efficiency gains, 30-day verified: $612 wasted-spend saved · $402 spend efficiency · $170 incremental sales. Cumulative run-rate since the loop started: $5.2K/mo across the account's history.
Estimated · staged & inside 30-day windows
$2,632/mo
This week's staged moves + applied changes still inside their measurement window. These graduate to measured — or get written down honestly when the data says otherwise.
W20W21W22W23W24W25W26W27 $180$320$425$560$640$790$920$1,184 + staged
Measured $/mo (30-day verified, cumulative)Staged, not yet counted

The change log — memory with receipts append-only · every change, its reason, its outcome

DateChangeReason loggedSourceOutcome
Jun 08Negate "storage rack for pots"52 clicks / 0 orders / accessory intentconsoleverified · +$28/mo
Jun 08"braiser" bid $1.31 → $1.123 wks over target · RPC-derived walk-downconsoleverified · +$61/mo, rank held #8
Jun 15Harvest "cast iron braiser 3.5 qt" → exact7 orders @ 22% in phrase · de-duped · abort gate setconsoleverified · +$140/mo
Jun 22TOS modifier 45% → 25% (Braisers)TOS 2.8× RoS ACoS post-eventpreauth · Bmeasuring · day 10 of 30
Jun 22Negate "camping tripod"31 clicks / 0 orders / wrong product linepreauth · Ameasuring · day 10 of 30
Jun 29Scale-up on 6-qt blue — declinedCover gate: 19 days of stock, PO inbound Jul 21sensorheld · auto-resumes on PO

Next session re-opens this file and measures every move against its captured baseline. Strategy compounds; guesses don't reset.

How a dollar earns its way into the ledger

STEP 1 · SUNDAY
Baseline captured
Before any change fires, the system snapshots the 30-day trailing state — spend, orders, ACoS, rank proxy — for exactly what's being touched.
STEP 2 · MONDAY
You approve, it executes
Approved moves push via the API with preconditions re-checked live at write time. Every write logs its prior value — everything is walk-back-able.
STEP 3 · DAYS 1–30
The window runs
The change sits in the estimated column. Abort gates watch for damage (zero-impression harvests auto-revert). No credit claimed yet.
STEP 4 · DAY 30
Measured — or written down
Outcome vs baseline, conservatively. Wins book as measured $/mo; misses get logged as misses. That's the number we stand behind.

The whole system in one line

01 · THE CONNECTOR
Caiman Data — data in
First-party SP-API + Ads API access to your account: every report, every grain, on schedule. Garbage in, garbage out — so the "in" is Amazon's own data, not scrapes or estimates.
02 · THE FRAMEWORK
Voltage — decisions out
The operating loop on top: weekly console, phase playbooks, guardrails, bounded autonomy with a killswitch, and a human on every judgment call. The connector is the pipe; this is the brain.
03 · THE PROOF
The ledger — measured
Baselines before, verification after, honest write-downs in between. Not projections on a slide — a running number your CFO can poke.

See it on your account

Text the keyword CAIMAN to +1 417.765.0412 — or book a no-pressure discovery call to find out if your business qualifies to work with Caiman Data + Voltage Business Builders.

Discover · Build · Optimize · Exit — we help entrepreneurs build Amazon businesses that create wealth.