How Recommendations Are Generated

Prism recommendations are built from your account's own performance data, not from a fixed script or an outside signal — this page explains what feeds them, how much evidence they require, how long a window of history they look at, and how often each type actually runs.

Fast Path

Every recommendation reads recent performance data for the entity it's evaluating, checks that there's enough evidence to act on, compares that evidence against a target threshold, and attaches a confidence score before it reaches your queue. How often a given type gets a chance to run at all depends on your plan's cadence — see How Often Recommendations Run below.

How To Use This

Four things determine whether — and how — a recommendation gets generated:

1. Data — the specific metrics Prism pulls for that entity (clicks, spend, orders, ACOS, and similar).
2. Timeline — how many days of history it looks at.
3. Threshold — the minimum evidence and the target it's comparing against.
4. Confidence — a score that reflects how much the first three actually support the suggested change.

The sections below walk through each one, by recommendation type. None of these numbers are contractual — they're current defaults that Prism refines as it learns more about typical account behavior, so treat them as "how this works today," not a permanent spec.

What Data Feeds a Recommendation

Every recommendation type covered on this page is built entirely from performance data synced from your connected Amazon Ads account: clicks, impressions, spend, orders, sales, ACOS/ROAS, CPC and CVR, budget utilization, placement-level performance, and hour-of-day metrics for scheduling.

Nothing here comes from a manual survey, a static rule you can't inspect, or a source outside your own account activity. If a recommendation cites evidence, that evidence is queryable in Command Center. All five types on this page currently evaluate Sponsored Products campaigns only.

How Often Recommendations Run

Recommendation types are grouped into three cadence tiers, and how often each tier runs depends on your plan:

Your plan	Real-time risk alerts	Daily tactical work (bids, budgets, negative keywords)	Weekly strategic work (dayparting, placement)
Free	Not scheduled	Not scheduled	Not scheduled
Explorer	Not scheduled	Weekly	Weekly
Starter	Daily	Daily	Weekly
Growth and higher	Hourly	Daily	Weekly

Every plan — including Free — gets one full run covering all three tiers during onboarding, so a newly connected account sees an initial pass at every recommendation type regardless of plan. After that, ongoing cadence follows the table above.

Bid Changes

Data: keyword-level clicks, spend, orders, sales, CPC, and ACOS.

Timeline: roughly the last 30 days of history per keyword.

Threshold: a keyword typically needs around 8 clicks in that window before Prism proposes a bid change — below that, there usually isn't enough evidence to act on confidently.

How the bid is calculated depends on the keyword's situation:

• Keywords with spend but zero orders get cut aggressively, close to the platform's minimum bid.
• Keywords running well above your target ACOS get a formula-based reduction — the further above target, the larger the cut, with built-in floors so a bid doesn't drop too far in a single step.
• Keywords running efficiently below your target ACOS get a formula-based increase — the increase gets larger (up to roughly 50% in the strongest cases) the more efficient the keyword is relative to target, and Prism won't suggest an increase at all if the resulting bid wouldn't meaningfully beat the keyword's current cost-per-click.
• Every proposed bid stays within a $0.02–$10.00 range, and reductions never propose a bid higher than the keyword's current one.

Scope: only manually-targeted keywords (exact, phrase, broad match) on Sponsored Products campaigns — auto-targeting isn't evaluated by this recommendation type.

Budget Changes

Data: daily spend, sales, and budget utilization per campaign.

Timeline: roughly the last 30 days of daily history, with the most recent few days set aside since order attribution can lag.

Increase eligibility: the campaign needs to be hitting its budget cap on the large majority of recent days (around 80% of them) while ACOS stays within target. The size of the increase (10–20%) scales with a blend of how much budget pressure the campaign is under, how often it hits the cap, and how much ACOS headroom it has — role-based caps can also limit how large an increase can be for certain campaign types.

Decrease eligibility: ACOS needs to be meaningfully above target — on the order of 1.5x target or more — sustained for roughly two consecutive weeks, not just a single bad day. The size of the cut (20–30%) scales with how far over target the campaign has drifted.

Faster path: campaigns can also get a proactive budget-increase recommendation within a day or two of a predicted budget exhaustion, once at least a week of spend history exists — this doesn't wait for the cap to actually be hit first.

Negative Keywords

Data: search-term-level clicks, spend, and conversions.

Timeline: roughly the last 60 days, with the most recent few days set aside for attribution.

Threshold: a search term typically needs around 8 clicks with zero conversions in that window before Prism suggests negating it — that floor rises to around 20 clicks for tightly-scoped campaign roles (for example, exact-match scaling or focus campaigns), since those campaigns are more sensitive to accidentally blocking a term that just hasn't converted yet.

Relevance context: every suggestion is scored for semantic relevance to your brand and product as supporting context you can review — it's informational, not a hard filter that silently excludes candidates.

Priority: labeled high, medium, or low based on how much spend is behind the term, to help you triage — this doesn't change whether a term is suggested, only how it's flagged for review.

Placement Optimization

Data: placement-level clicks, spend, and sales, split across Top of Search, Product Pages, and Rest of Search.

Timeline: roughly the last 30 days.

Threshold: a campaign needs at least $50 in combined spend across its placements, and at least one placement already performing better than target ACOS, before Prism proposes a change.

How the multiplier is calculated: Prism computes the theoretical optimal placement multiplier for an outperforming placement, then proposes roughly half of that step as a conservative move rather than jumping straight to the calculated optimum. Amazon placement multipliers can only go up, never negative, so placement recommendations only ever propose increases — up to a maximum multiplier of 900%.

Dayparting

Data: hour-of-day performance (clicks, conversions, spend) per campaign.

Timeline: a rolling 60-day window — not a minimum-history requirement. The window is aggregated into 24 hour-of-day buckets (all 9am observations summed together, all 2pm observations summed together, and so on) before analysis, so a bucket's evidence comes from every matching hour across the whole window, not from a single day.

Threshold: a campaign needs at least $100 in combined 60-day spend to be analyzed. Each hour-of-day bucket is classified as off-peak, active, or exceptional relative to your account's own average conversion rate, and a bucket only gets flagged once it has accumulated enough clicks on its own — typically a handful, scaled to your account's overall conversion rate.

Scope: Sponsored Products campaigns only, and schedule changes only apply to blocks of at least 4 consecutive hours, to avoid fragmented on/off toggling.

How Confidence Is Calculated

Most recommendations carry a confidence score built from several signals, not a single guess. Bid, budget, and placement recommendations share one model that blends:

• Sample size (up to 35% of the score) — more clicks and conversions behind the recommendation raise confidence.
• Data stability (up to 25%) — a steady day-to-day conversion rate raises confidence more than a volatile one.
• Size of the change (up to 20%) — a small, cautious adjustment carries less risk (and often more confidence) than a large one.
• Recency (up to 20%) — recommendations built on data from the last few weeks carry more confidence than ones stretching back further.

Negative-keyword confidence uses a related but separate formula weighted toward click volume relative to the threshold and spend behind the term, with its own floor so a negative-keyword suggestion is never scored below a baseline confidence level.

A low-confidence recommendation typically means a smaller, more conservative suggested change, or a flag for closer manual review, rather than a suppressed recommendation. This score is also what automation gating reads — see Turn On Auto-Apply and Use Trust and Smart Tuning for how confidence connects to what runs automatically.

What This Isn't

These thresholds, timelines, and cadences are today's defaults, not a fixed promise. Two similar-looking keywords can get different treatment because of campaign role or account-specific data volume — not because the system is inconsistent. As Prism's models improve, specific numbers here can shift; the underlying principle — act only when the evidence supports it — does not.

FAQ

Does Prism ever act on too little data?

No. Every recommendation type has a minimum evidence bar. Below that bar, Prism either withholds the recommendation or waits for more data instead of guessing.

Can I see the specific evidence behind one recommendation?

Yes. Open the recommendation in Command Center to see the underlying data, the reasoning, and the confidence score behind it.

Why don't I see certain recommendation types as often as I expected?

Cadence depends on your plan. Bid, budget, and negative-keyword recommendations run daily on Starter and above but weekly on Explorer; dayparting and placement recommendations run weekly on every paid plan. See How Often Recommendations Run for the full breakdown.

Do these recommendation types cover Sponsored Brands or Sponsored Display?

Not currently. Bid changes, budget changes, negative keywords, placement optimization, and dayparting all evaluate Sponsored Products campaigns only. Support for Sponsored Brands and Sponsored Display is being planned for the next major release.

Why did two similar keywords get different recommendations?

Usually a campaign role difference, or a difference in how much recent data each one has — not inconsistent logic.

Do these thresholds change over time?

Yes. They're current defaults, but we may refine them as we learn more about typical account behavior, so exact numbers can shift between reviews of this page.

Next Steps

Go to Work Command Center to see this evidence applied to real recommendations, or read Manage Bids if bid-specific thresholds are what you're evaluating right now.