Every candidate action gets scored before it reaches you. High confidence is the default filter — actions below that are visible if you toggle the filter, but the system never recommends them by default.

Bid & negate — the patience budget

Account CPO = Total Spend ÷ Total Orders. Multiplied by your target ACoS, this is one conversion's worth of spend at goal efficiency. Think of it as "one unit of patience" for a single keyword.

Harvest confidence — search terms to promote

Different products have different margins. A premium SKU at 15% target ACoS shouldn't be judged against a discovery SKU at 40% target. BSP looks up the target ACoS for each row in this order, taking the first one that matches:

1. SKU rule — most specific. Set per-SKU on the Settings page.
2. Keyword rule — applies to a specific keyword or product target.
3. Campaign-type rule — applies to all SP / SB / SD campaigns.
4. Account-level target — the universal default.

Placements and keywords with sparse data can produce wildly misleading metrics — a placement with 22 clicks and 0 orders looks like infinite ACoS, but that's mostly bad luck rather than real signal. Three reliability layers pull noisy metrics toward sensible defaults so the rec engine doesn't churn on statistical artifacts.

1. Bayesian CVR smoothing on placements

The prior CVR is the campaign's own CVR if the campaign has ≥ 50 clicks AND ≥ 5 orders; otherwise it's the account-wide CVR. The prior weight = 25 means the placement's own data starts to dominate around 30 clicks. A placement with 25 clicks and 0 orders gets smoothed to roughly the campaign's CVR rather than 0% — preventing infinite-ACoS recommendations on early data.

2. Minimum orders gate for placement validity

A placement only participates in the diagnostic gate's routing decision if it has ≥ 20 clicks AND ≥ 2 orders. The clicks floor was already in BSP; the orders floor is new. Placements below this don't generate recs and don't influence the bid-vs-placement routing.

3. CPC safeguard on bid increases

The math-efficient ceiling for a keyword's CPC. If the current CPC already exceeds this ceiling, BSP will never recommend bidding higher — even if the keyword's ACoS reads as "below target" due to a few cheap clicks. In that case, the safeguard caps the new bid at currentBid × 0.90 (a 10% pullback) instead.

Bids and placement multipliers interact mechanically — final bid in any auction is bid × placement multiplier. Recommending both at once at full strength tends to overshoot, because each adjustment is sized as if it's the only thing changing. The gate routes intelligently to prevent compounded overshoot.

Tier 2 safety rules — run first, in every campaign

1. Safety A — "Don't scale bad traffic": if total campaign ACoS > target × 1.10 AND every placement is also above target, the problem is the campaign as a whole, not allocation. Fire bid only (downward).
2. Safety B — "Don't reduce good traffic": if total campaign ACoS < target × 0.90 AND every placement is also below target, the campaign is broadly underspending. Fire bid only (upward).

If neither safety rule fires, routing branches based on how many enabled keyword/product targets the campaign has.

Single-keyword-target campaigns — gap comparison

When a campaign has just one keyword/target, both levers are judged against the same target. BSP measures how broken each lever is by computing a gap — the worst ACoS deviation from target — and routes to whichever is bigger:

• No qualifying placement → bid is the only available lever
• Tied gaps (within 10pp) → bid wins by default. Exception: if a placement is bleeding (consuming ≥ 10pp more spend share than its sales share), the imbalance signal flips it — placement wins
• Otherwise the larger gap wins; the other lever is suppressed

Multi-keyword-target campaigns — bids default, placements only when clearly bad

When a campaign has 2+ keyword/product targets, the placement multiplier affects all of them simultaneously, and they may have different performance. The routing becomes more conservative about touching placements:

• If any placement is clearly bad → both levers fire together, with each nudge scaled by 0.5 to prevent compounded overshoot
• If no placement is clearly bad → fire bid only, suppress placement
• No tie band, no imbalance signal in this branch — the strict 1.5× test is the only trigger

The 0.5× scaling means a bid that would have moved $1.50 → $1.30 now moves $1.50 → $1.40, and a placement that would have moved 80% → 70% now moves 80% → 75%. Each lever contributes equally to the fix instead of compounding. The CPC safeguard always fires at full strength regardless of dual-fire scaling.

Worked example — single-keyword campaign

Worked example — multi-keyword campaign

Most ad optimizers treat each rec independently — they don't model how your queued actions affect each other. BSP does. When you queue a search-term negate, the parent keyword's metrics will change once that negate applies (lower spend, possibly lower clicks, same or higher sales). BSP recomputes bid and placement recs against those projected metrics so you don't end up double-correcting.

What gets adjusted

For each negated search term, BSP subtracts its full clicks/spend/sales/orders from:

• The parent keyword/target that produced it (matched by campaign + ad group + keyword text — covers all target types: keyword, ASIN, expression, audience)
• The placement aggregates for that campaign, allocated proportionally by each placement's share of campaign spend (since search term reports don't expose which placement a click came from, proportional allocation is the best available model)

After subtraction, ACoS, CPC, and CVR are re-derived from the adjusted totals. Every bid rec, every placement rec, and the diagnostic gate's routing decision uses these projected metrics instead of raw metrics.

What this means in practice

• When you queue a bad search term, the parent keyword's projected ACoS drops
• Its bid rec gets gentler — closer to current bid, since the negate cleanup already does some of the work
• If projected ACoS lands within the 5% tolerance band, the bid rec disappears entirely
• Placement aggregates also shift, which can change the diagnostic gate's routing decision for that campaign

Worked example

Scope — only Negate Terms rec actions feed projections

The projection layer only models negates queued from the Negate Terms → ⚡ Rec Actions tab, where the workflow naturally happens. Other negate sources are intentionally excluded:

• Pauses — multi-SKU ad groups can't be cleanly attributed; single-SKU pauses make optimization moot anyway
• Bulk Negate / Bulk Negate Phrase — outside the normal optimization routine
• Cross-campaign negation — small effect at the per-rec level, complexity not worth it
• Search Term Patterns — outside the normal optimization routine

Workflow protection — the going-back guardrail

If you've already staged bid or placement recs and then go back to the Negate Terms rec actions tab to queue more negates, BSP shows a modal asking how to handle the inconsistency:

• Apply negates, keep bid/placement recs as-is — accept that recs may now be slightly off (they were sized against the pre-negate world)
• Remove queued bids & placements, apply negates — wipe the bid and placement queues so you can re-stage from scratch with fresh projection-aware recs
• Cancel — close modal, don't apply

Following the natural workflow (negates → placements → bids) avoids this modal entirely. It only fires when you go backward, which is the moment friction is appropriate.

A search term qualifies for negation when its ACoS is far above target, OR when it's accumulated meaningful spend without any orders. The thresholds are deliberately strict — negation is destructive (the term can't show again in that ad group), so BSP only negates when the case is unambiguous.

Triggers (either condition)

• Zero orders, spend over patience: orders = 0 AND spend > 7 × patience budget. The 7× multiplier matches "seven units of patience burned" — a clear signal the term won't convert.
• Has orders but ACoS too high: orders > 0 AND ACoS > target × 4. Even with conversions, ACoS this far above target is unrecoverable through bid tuning alone.

Action

Add the search term as an exact-match negative keyword in its ad group. The term stops showing in that ad group; same term can still appear in other ad groups unless cross-campaign negation also fires (Action 2).

Some search terms appear in multiple campaigns at once — usually because broad-match keywords pulled them in independently. When a term is wasting money across ≥ 2 campaigns, negating it in just one ad group leaves the bleed running elsewhere. This rec catches that.

Eligibility

• Same search term appears in ≥ 2 distinct campaigns
• Triggers either: zero orders + spend > 7 × patience budget, OR ACoS > target × 4 (same triggers as Action 1, but evaluated on combined cross-campaign totals)

Action

Add as an exact-match negative in every ad group where the term has appeared. One queued action covers all the duplicate negation work.

Bid adjustments cover any enabled keyword or product target with a bid — across Sponsored Products, Sponsored Brands, SB Multi Ad Group, and Sponsored Display campaigns. Only entities with actual spend, a recorded ACoS, and a non-trivial deviation from target are considered. Keywords within 5% of target (either side) are skipped — that tolerance band prevents micro-adjustments around already-good positions.

The Gentle Nudge formula

• Sensitivity is the Bid Nudge slider (default 0.25 = "Moderate")
• The move is proportional — a keyword far above target gets a larger absolute move than one slightly above
• Bids are floored at $0.02 (Amazon's minimum)
• The CPC safeguard caps any bid increase if current CPC already exceeds the math-efficient ceiling (see Reliability Layer)

Why not just set the "correct" bid directly?

Amazon's auction is dynamic. A theoretically-correct bid today can overshoot tomorrow as competition shifts. Nudging in proportional steps converges gradually — preserving "sweet spot" positions where a single large jump could collapse impression share that took weeks to build.

Placement adjustments operate on all placement types present in your bulk sheet, including AMC audience placements. The math is type-agnostic — any placement record with the right shape (multiplier, clicks, orders, ACoS) is eligible. New placement types Amazon adds in the future will be picked up automatically.

Eligibility — the validity floor

• ≥ 20 clicks AND ≥ 2 orders — below this, the ACoS reading is too noisy to act on (see Reliability Layer for the smoothing math)
• Not in the legacy "Sponsored Brands Campaigns" sheet (Amazon doesn't accept placement adjustments via that sheet)
• Resolved target ACoS exists for the placement's campaign

The Gentle Nudge formula (mirrors bids)

• Multipliers are clamped to 0% – 900% (Amazon's allowed range)
• Result is rounded to the nearest whole percent

Bid × multiplier interaction

When a search term is consistently outperforming target inside a broad/auto campaign, the smart move is to break it out as its own exact-match keyword in a dedicated campaign. That gives you direct control over its bid and isolates its budget from the noisier parent campaign.

Eligibility

• Search term has at least 2 orders
• ACoS ≤ target × 0.5 (performing at half target or better)
• OR: orders ≥ 2 with ACoS unattributed (still harvest-worthy on click pattern alone)
• Not in an exact-match or negative-exact ad group (already promoted)

High confidence requires ≥ 5 orders on the search term.

Action

Create a new Sponsored Products campaign targeting the search term, with SKU pre-filled from the best-performing ad group for that term. You set the campaign name, budget, and bids before the action queues.

Sometimes the issue isn't a single search term — it's a phrase that appears in dozens of variations, all of them bleeding (or all of them converting). BSP analyzes 2-word and 3-word patterns across the search term report and flags both ends.

Pausing a product ad stops the product from showing in that campaign entirely. The thresholds here are deliberately stricter than for negate/bid recs — only products that almost certainly won't recover from term-level tweaks alone are flagged.

Triggers (either)

• Zero orders AND spend > 10× patience budget (vs 7× for negation)
• Has orders AND ACoS > target × 6 (vs 4× for negation)

"Safe to apply" doesn't mean every metric will improve overnight. It means:

• Statistically grounded — every high-confidence action cleared both a click threshold and a spend threshold scaled to your account economics. Plus the reliability layer (Bayesian smoothing, min-orders gate, CPC safeguard) prevents acting on noise.
• Proportional, not aggressive — bid changes use a nudge model, not full corrections. No single change craters impression share.
• Conservative thresholds — negation requires 4× overshoot (not 1.1×), harvest requires 50%-of-target ACoS (not just any conversion).
• Smart lever routing — the diagnostic gate decides whether to fire bids, placements, or both, based on how many keyword targets the campaign has and how broken each lever is. When both fire, each is scaled by 0.5 to prevent compounding.
• Recs react to queued negates — the projection layer recomputes bid and placement recs against the post-negate world, so you don't stack corrections that overshoot.
• Reversible — every change goes into the Optimization Queue first. You review, undo individual items, export when satisfied. Nothing reaches Amazon until you upload the file yourself.