Break down sales contribution with N-grams

These prompts break search terms down into N-grams (word / phrase units) and use DuPont decomposition to identify which phrases generate sales and where efficiency drops. This is the main analysis tool for separating “sales are coming in but efficiency is poor” from “efficiency is good but exposure is too thin”—two situations you can never tell apart by looking at ACoS alone. It’s the natural entry point when you want fresh hints for your next move.

When you use this category

• “I want to analyze which terms are driving sales”
• “I want to see N-gram performance by brand name or use case”
• “I need an interpretation guide that accounts for Auto campaign anonymization”

N-gram premise

N-gram analysis decomposes Search terms—not the Keywords you registered in the ad. It takes the strings shoppers actually typed into Amazon’s search box and splits them into 1-word / 2-word / 3-word / 4-word units to see how each phrase contributes to sales.

---

Prompt 1: Identify N-grams with high sales contribution

🎯 Purpose

Decompose Search terms into mono / bi / tri / 4-grams and surface the top phrases by Sales Contribution and Imp Efficiency.

📋 When to use

• You want to find the terms driving sales
• You want to analyze which phrases are creating sales
• You want a list of Search terms with high Sales Contribution

📝 Prompt

Using the most recent finalized weekly data, decompose Search terms
into N-grams (mono / bi / tri / 4-gram) and run performance analysis
with DuPont decomposition (Sales = Imp × CTR × CVR × AOV).
Run with top_n {{TOP_N}} and max_pages {{MAX_PAGES}},
focusing on Sales Contribution and Imp Efficiency for the top entries.

Placeholders:

Placeholder	Description	Default source
{{TOP_N}}	Number of top entries to show per N-gram type (e.g. 20)	Default 20
{{MAX_PAGES}}	Maximum number of pages to fetch (200 rows / page, e.g. 10)	Default 10

⚙️ What the AI does

1. Decomposes Search terms into N-grams — pulls Search terms from the finalized weekly data and splits them into four types: mono-gram (1 word) / bi-gram (2 words) / tri-gram (3 words) / 4-gram (4 words)
   • Why finalized weeks: unfinalized weeks still have impressions and clicks in mid-aggregation, which makes the denominator unstable for judging phrase-level sales contribution
   • Why four types: shorter phrases produce more rows but interpretation is coarse; longer phrases produce fewer rows but the intent is clearer. Looking at each level separately lets you isolate “what is happening at which granularity”
2. Evaluates each phrase with DuPont decomposition — expands into the four factors of Sales = Imp × CTR × CVR × AOV
   • Why DuPont: separating the components of sales lets you isolate the next move—“impressions are sufficient but CVR is low → LP / pricing issue” vs. “CVR is high but exposure is thin → bidding / budget issue”
3. Calculates Sales Contribution and Imp Efficiency — extracts the top entries centered on Sales Contribution and Imp Efficiency (= CTR × CVR)
   • Why look at Imp Efficiency: even at the same ACoS, the action changes depending on whether CTR or CVR is driving it. The product gives you a composite score for efficiency
4. Returns an approval table — for each N-gram type, lists phrase / Sales / Spend / ACoS / CTR / CVR / AOV / CPC / Imp Efficiency along with medians and totals
   • Why include the median: averages get pulled by outliers; showing the median alongside makes it easier to see how skewed the top phrases are

📊 Example Output

N-gram analysis (most recent finalized week) — top 20 per N-gram type

Top bi-grams

Phrase	Sales	Spend	ACoS	CTR	CVR	AOV	Imp Efficiency
amino-acid shampoo	¥182,400	¥27,000	14.8%	0.42%	8.1%	¥3,800	0.034
refill shampoo	¥98,200	¥18,200	18.5%	0.31%	5.6%	¥3,500	0.017
… (18 more) …

Top tri-grams

Phrase	Sales	Spend	ACoS	CTR	CVR	AOV	Imp Efficiency
amino-acid shampoo refill	¥64,800	¥8,200	12.7%	0.51%	9.4%	¥3,800	0.048
… (19 more) …

Medians: ACoS 19.4% / CTR 0.28% / CVR 4.2% Totals: Sales ¥1,248,000 / Spend ¥232,000 / ACoS 18.6%

Mono-grams and 4-grams have been output in the same format.

🔒 Safeguards

• Analysis only—no automatic application to bids or negative Keywords (no proposal step)
• Unfinalized weeks are automatically excluded — aggregation goes back to the most recent finalized week
• Actual Search terms in Auto campaign rows may be anonymized — Auto-sourced rows can be anonymized by the Picaro backend (Auto campaign anonymization), so they cannot be classified as precisely as Manual campaigns
• N-gram-level totals include duplicate counting — they do not match the Search term-level total (the same term is broken into multiple N-grams)
• Rows where the actual Search term cannot be obtained fall back to the registered Keyword

🚦 Execution mode

Phase	State	Conditions
Phase 1 (now)	One-shot	Available with Picaro connection only
Phase 2 (Q3 2026)	Weekly run	Convertible into a “saved prompt,” delivered weekly at the start of the month
Phase 3 (Q4 2026)	Monthly auto (bundled in reports)	Top N-grams auto-attached to the monthly report
Phase 4 (2027)	Monthly auto	Slack notification with the key points only when anomalies are detected

👉 Next steps

• See the market search funnel with SQP — look at the same phrases on Amazon’s market search data, not Picaro’s ad data
• Design label taxonomy — attach meaningful labels to the top phrases and move on to semantic analysis
• Why looking only at ACoS isn’t enough — the philosophy behind DuPont decomposition and multi-dimensional judgment

---

Prompt 2: N-gram analysis filtered by Taxonomy

🎯 Purpose

Filter N-grams by a pre-designed Taxonomy (brand / use / competitor, etc.) and compare sales contribution and efficiency at the level of meaningful groups.

📋 When to use

• You only want to see N-grams related to brand names
• You want to focus the analysis on competitor-category terms
• You want to analyze within a specific label category

📝 Prompt

Using the most recent finalized weekly data, run N-gram analysis
filtered to Taxonomy category "{{TAXONOMY_CATEGORY}}".
With top_n {{TOP_N}} and max_pages {{MAX_PAGES}},
output Sales Contribution / Imp Efficiency / negative Keyword candidates /
promotion candidates.

Placeholders:

Placeholder	Description	Default source
{{TAXONOMY_CATEGORY}}	Taxonomy category used for filtering (e.g. brand / use / competitor)	Taxonomy registered via label management; falls back to a built-in demo value if not registered
{{TOP_N}}	Number of top entries to show per N-gram type (e.g. 20)	Default 20
{{MAX_PAGES}}	Maximum number of pages to fetch (e.g. 10)	Default 10

⚙️ What the AI does

1. Filter targets by Taxonomy — uses the Taxonomy registered beforehand via Design label taxonomy to split N-grams into semantic clusters
   • Why filter by Taxonomy: looking only at portfolio-wide average ACoS hides structural skews like “over-investment in defensive Keywords.” Slicing by semantic clusters is the only way to see things like “the efficiency of offensive KWs is dropping” or “too much budget is concentrated on brand KWs”
2. N-gram decomposition + DuPont decomposition — same as Prompt 1: mono / bi / tri / 4-gram × Sales = Imp × CTR × CVR × AOV
3. Surfaces negative Keyword candidates and promotion candidates together — within each group, picks out phrases with zero sales but accumulating clicks as negative Keyword candidates, and Auto-campaign-sourced phrases with strong performance as promotion candidates
   • Why surface both: once you’ve narrowed down to a semantic group, the next action—“what to cut and what to grow in this group”—becomes clear, so the report goes straight from analysis to the next move in one pass
4. Behavior when Taxonomy is not registered — runs on built-in demo values and prepends a note encouraging the user to register a Taxonomy

📊 Example Output

N-gram analysis — Taxonomy competitor (most recent finalized week)

Top bi-grams (under competitor)

Phrase	Sales	Spend	ACoS	Imp Efficiency	Labels
Competitor-A shampoo	¥84,600	¥21,200	25.1%	0.018	competitor / brand-A
Competitor-B refill	¥42,300	¥12,800	30.3%	0.014	competitor / brand-B
… (18 more) …

Negative Keyword candidates: 5 (clicks ≥ 10 / sales = 0) Auto → Manual promotion candidates: 3 (Auto-sourced phrases with high sales contribution)

Competitor group totals: Sales ¥248,000 / Spend ¥68,400 / ACoS 27.6%

Detailed usage: Design label taxonomy

🔒 Safeguards

• Analysis only—negative Keywords and promotions are not applied automatically (candidate list only)
• Unfinalized weeks are automatically excluded
• Actual Search terms in Auto campaign rows may be anonymized (Auto campaign anonymization)
• N-gram-level totals include duplicate counting (do not match the Search term-level total)
• If no Taxonomy is registered, it runs on built-in demo values and results are treated as reference

🚦 Execution mode

Phase	State	Conditions
Phase 1 (now)	One-shot	Picaro connection + Taxonomy registration (optional)
Phase 2 (Q3 2026)	Weekly run	Convertible into a “saved prompt” for regular per-group checks
Phase 3 (Q4 2026)	Monthly auto (bundled in reports)	Top phrases per group auto-attached to the monthly report
Phase 4 (2027)	Monthly auto	Slack notification only when between-group skews look abnormal

👉 Next steps

• Design label taxonomy — define the semantic groups here if Taxonomy is not yet registered
• Add negative keywords — execution destination for the negative Keyword candidate list
• Promote Auto to Manual — execution destination for the promotion candidate list

---

Prompt 3: N-gram analysis with interpretation guide

🎯 Purpose

Annotates a standard N-gram analysis with data-constraint notes so the report can be read while keeping Auto campaign anonymization and total mismatches in mind.

📋 When to use

• You’re not sure how to read the results
• You want to know why Auto campaign Search terms aren’t visible
• You want analysis that accounts for the data constraints

📝 Prompt

Using the most recent finalized weekly data, run N-gram analysis and
state the following caveats at the top of the report:

1. Data scope: ad target detail data
2. Rows where the actual Search term cannot be obtained fall back to the registered Keyword
3. Actual Search terms in Auto campaign rows may be anonymized
4. N-gram-level totals include duplicate counting and do not match the Search term-level total

Run with top_n {{TOP_N}} and max_pages {{MAX_PAGES}}.

Placeholders:

Placeholder	Description	Default source
{{TOP_N}}	Number of top entries to show per N-gram type (e.g. 20)	Default 20
{{MAX_PAGES}}	Maximum number of pages to fetch (e.g. 10)	Default 10

⚙️ What the AI does

1. Shows the four constraint notes at the top — summarizes “data scope,” “fallback,” “Auto campaign anonymization,” and “duplicate counting” at the very top of the report
   • Why put them at the top: if the reader sees only the result table, they will be confused by “the numbers don’t match” or “Auto’s actual Search terms aren’t appearing.” Sharing the premises up front prevents that
2. N-gram decomposition + DuPont decomposition — runs the same aggregation as Prompt 1
3. Adds an Auto / Manual provenance column — flags each N-gram row as Auto-sourced or Manual-sourced
   • Why expose provenance: Auto-sourced rows carry anonymization risk and cannot be classified as precisely as Manual-sourced ones. With provenance visible, the reader can adjust their confidence in the interpretation accordingly
4. Restates the reason totals don’t match at the end of the body — re-explains “N-gram-level total ≠ Search term-level total” with the formula

📊 Example Output

N-gram analysis (with interpretation guide) — most recent finalized week

Premises before reading

1. Scope: ad target detail data (finalized week)
2. Rows where the actual Search term cannot be obtained fall back to the registered Keyword
3. Auto-sourced actual Search terms may be anonymized in the Picaro backend
4. Because the same Search term decomposes into multiple phrases, N-gram-level totals do not match the Search term-level total

Top bi-grams

Phrase	Provenance	Sales	Spend	ACoS	Imp Efficiency
amino-acid shampoo	Manual	¥182,400	¥27,000	14.8%	0.034
(anonymized phrase)	Auto	¥58,200	¥14,100	24.2%	0.012
… (18 more) …

Mono / tri / 4-grams have been output in the same format.

Why the totals don’t match: the Search term “amino-acid shampoo refill” expands into bi-grams “amino-acid shampoo” and “shampoo refill” and the tri-gram “amino-acid shampoo refill,” so the same sale appears on multiple rows.

🔒 Safeguards

• Analysis only—no automatic application
• Unfinalized weeks are automatically excluded
• Auto-sourced rows are displayed with an anonymization flag
• Rows that fell back to the registered Keyword are explicitly shown with a fallback indicator
• The difference between N-gram totals and Search term-level totals is always restated at the end of the body

🚦 Execution mode

Phase	State	Conditions
Phase 1 (now)	One-shot	Available with Picaro connection only
Phase 2 (Q3 2026)	Weekly run	Convertible into a “saved prompt” as a regular analysis for beginner users
Phase 3 (Q4 2026)	Monthly auto (bundled in reports)	Auto-attached to the monthly report with the interpretation guide
Phase 4 (2027)	Monthly auto	Constraint notes are always included so the reader never loses the premises

👉 Next steps

• Identify N-grams with high sales contribution — having understood the constraints, move on to whole-portfolio analysis without filtering
• N-gram analysis filtered by Taxonomy — narrow down to semantic groups to raise interpretation accuracy
• Create campaigns — if you want to avoid Auto campaign anonymization, consider converting to Manual campaigns

❓ FAQ

Q: Why are Auto campaign actual Search terms anonymized? A: With Auto campaigns, some of the Search terms that Amazon automatically matched are no longer obtainable as individual terms by the time the data reaches Picaro. This is a constraint on the data source side, not Picaro’s. If you want to track the sales efficiency of anonymized phrases in detail, the standard move is to convert the campaign to Manual.

Q: Is it an aggregation error that the N-gram-level total and the Search term-level total don’t match? A: It is not an error. N-grams are designed to decompose one Search term into multiple phrases, so the same sale is counted on multiple N-gram rows. The Search term-level total is the correct sales figure; the N-gram total is only for relative comparison between phrases.

---

Prompt quick reference

What you want to do	Prompt to use
Sales contribution analysis across all N-gram types	Prompt 1
Analysis filtered by Taxonomy	Prompt 2
Interpretation guide that accounts for data constraints	Prompt 3

Related categories

• Design label taxonomy — design the Taxonomy (prerequisite for Prompt 2)
• Add negative keywords — process the negative Keyword candidates found via N-grams
• Promote Auto to Manual — process the promotion candidates found via N-grams
• See the market search funnel with SQP — look at Amazon’s search data instead of Picaro’s ad data
• Why looking only at ACoS isn’t enough — the philosophy behind DuPont decomposition and multi-dimensional judgment