Skip to content

Decides

Search terms triage

Sorts every search term into keep, negative or new ad group, totals the waste by word and phrase, and reviews AI Max matches on their own.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick search-terms-triage.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
AI Max reporting, the Source column and brand lists
Needs
A search terms export from Google Ads or Microsoft Ads, ideally 90 days · The most a lead may cost · Brand terms, competitor names and your category words
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedPaid media
SkillPaid media

Search terms triage

Example output

Triage 90 days of search terms
PhraseSpendCall
free$1,326Negative
for hubspot$520New ad group
excel template$216Watch

Sorts every search term into keep, negative or new ad group, totals the waste by word and phrase, and reviews AI Max matches on their own.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

search-terms-triage/SKILL.md

In the folder

  • search-terms-triage/SKILL.md
  • search-terms-triage/references/negative-categories.md
  • search-terms-triage/scripts/ngram_triage.py
---name: search-terms-triagedescription: >-  Sorts every search term in a Google Ads or Microsoft Ads export into keep, negative, new keyword or new ad group,  bid down, or watch, totals the zero-conversion spend by n-gram so recurring waste shows up, and reviews AI Max  matches in their own section. Every threshold is tied to the company's own target cost per lead, and every negative  passes a false-positive check before it is listed. Returns ACT, CLEAN, INSUFFICIENT DATA, NEEDS DECISION or BLOCKED.  Use it weekly on any search account, or before a negative keyword cleanup. Not for setting the cost per lead  ceiling (use lead-value-math) or for writing the ads of a new ad group (use ad-variants).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Uses a connected ad account or reporting  connector if the agent has one, otherwise an uploaded CSV export or pasted rows. Python 3 is optional, for the  n-gram pass.metadata:  author: SHAKEUP  version: "1.0.0"  stage: decide  plugin: paid-media  homepage: https://www.shakeuphq.com/skills/search-terms-triage--- # Search terms triage Finds the words that waste search budget across every term they appear in, and the converting terms the account isn't bidding on yet. ## Purpose - One job: label every search term in one export and turn recurring waste into negatives with a match type, a level and a reason.- Stage: decide. It ends in a verdict for the run, and a label for each term.- It never adds, removes or uploads a negative, a keyword or a bid. It writes the list and the change log; the reader applies them (Google Ads Editor or Microsoft Advertising Editor, so there is a reviewable diff). ## Use when / not | Situation | Use || --- | --- || Weekly search term review, or a first cleanup of an inherited account | this skill || Sizing the waste before asking for budget | this skill (the waste total is the headline figure) || You don't know the most a lead may cost | `lead-value-math` first; this skill reads its `max_cpl` || A converting term needs its own ads and landing page | this skill labels it; `ad-variants` writes the ads || Checking an agent's access before it applies negatives | `agent-guardrails` || Choosing a keyword for an organic page | `keyword-selection` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || Search terms export: term, match type, added or excluded status, campaign, ad group, clicks, impressions, cost, conversions | yes | a connected ad account or reporting connector (pull the full window in 7-day slices if the connector caps rows, then merge); a CSV export; pasted rows | ask for it; `BLOCKED` || Window length | yes | the export's date line; the reader | ask; 90 days is the default request because shorter windows flag words that haven't had time to convert || Threshold T: the target cost per lead | yes | `lead-value-math` handoff `max_cpl`; `## Targets` "Target cost per lead"; the account's own non-brand cost per conversion (weaker, labelled) | use the weaker source and say so; with no conversions at all, `INSUFFICIENT DATA` || Brand terms and competitor names | yes | `## Company` "Brand terms"; `## Competitors`; the reader | ask once; without brand terms, brand spend hides inside non-brand || Core category words and product names | yes | `## Company` "Category"; `## Product`; `## Buying moments` search phrases | ask once; without them the core-category check can't run, coverage `partial` || Qualified outcomes per term (SQL or opportunity column from the CRM) | no | a CRM join on click ID; a column the reader adds | judge on platform conversions and say so || Pricing (free plan or trial?) | no | `## Company` "Pricing" | treat "free" as phrase-only until confirmed || AI Max columns: Source, landing page | no | the Search terms and ad combinations view; the export | review AI Max on match type alone | ## Profile read From `company-profile.md`: `## Company` (brand terms, category, pricing, business model), `## Product`, `## Buyers` (ideal customer and bad fit, for the plausible-buyer test), `## Buying moments`, `## Competitors` (names and mention policy), `## Targets` (target cost per lead), `## Accounts and data`, and every `## Learnings` entry for `search-terms-triage` with status `noted`, `fact` or `accepted`. No profile: ask in one message for brand terms, competitor names, the category in the buyer's words, whether there is a free plan, and the most a lead may cost. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings.** A past negative that cost conversions, or a term the reader protected, overrides the rules below. Name each one that fires. 2. **Load and merge.** Normalize terms (lowercase, trimmed, single spaces). Merge the same term across campaigns and slices: sum clicks, impressions, cost and conversions; keep the campaigns, match types and the strongest status (added, then none, then excluded). Drop "Total" rows. Report rows in and unique terms out. 3. **Split the pools.**    | Pool | Rule | Why |   | --- | --- | --- |   | Brand | the term contains a brand term (spaces ignored for 5+ letter names) | Brand converts at a different rate and hides non-brand waste |   | AI Max | match type or Source says AI Max | Reviewed in step 9 with its own rules |   | Competitor | contains a competitor name | Inside non-brand totals, labelled separately |   | Non-brand | everything else | The n-gram pool | 4. **Size the waste.** Waste = cost of non-brand terms with 1+ click and 0 conversions (excluded terms left out). Waste rate = waste / non-brand cost.    | Waste rate | Read |   | --- | --- |   | under 15% | healthy |   | 15% to 25% | watch |   | over 25% | investigate: the account is buying traffic it can't convert | 5. **N-gram the non-brand pool.** Split every term into 1, 2 and 3-word phrases (each counted once per term; single stopwords dropped; "what is", "how to" kept). Sum clicks, impressions, cost and conversions per phrase across every term it appears in. Flag, with T the threshold:    | Flag | Rule | Why |   | --- | --- | --- |   | negative strong | 0 conversions, cost at least 2 x T | Enough spend to have converted at the target cost twice |   | negative candidate | 0 conversions, cost at least 1 x T | It has spent what one lead may cost, and produced none |   | safe category | a jobs, education or consumer word, 0 conversions, any spend | The lowest false-positive categories; they need no threshold |   | over ceiling | converts, but cost per conversion above T | It converts: bid down, qualify or route to a cheaper offer, never negate |   | protected core | the phrase is, or sits inside, a core category phrase or product name | A negative on it would block the company's own category |   | in converting term | a single word that appears in any converting term | Never a broad negative | 6. **Run the false-positive checks** on every flagged phrase before it becomes a negative. All 4 must pass:    | Test | Question | Fails when |   | --- | --- | --- |   | Converting term | Does any converting term contain it? | yes, for a broad negative (a phrase or exact negative on the bad combination may still pass) |   | Plausible buyer | Could the ideal customer type it while looking to buy? | yes ("template" can be a buyer word; "salary" can't) |   | Core category | Is it the category or a product name? | yes: never negate; reclassify (step 7) or hold |   | Sample size | 20+ impressions and 5+ clicks behind it? | no: watch, unless it is a safe category | 7. **Classify and place.** Put each passing phrase in one of the 11 categories in `references/negative-categories.md` and take its match type and deployment group from there. Choose the narrowest match that blocks the waste: a single word broad only when it disqualifies on its own and appears in no converting term; otherwise phrase on the 2 or 3-word phrase; exact for one bad query. Add plurals as their own negatives. A phrase that fits no category is either wrong-industry waste (Off-ICP list) or a targeting gap (step 8): say which. 8. **Label every term.**    | Label | Rule (first match wins) |   | --- | --- |   | already excluded | status excluded: note only |   | competitor evaluative | competitor plus alternative, vs, pricing, reviews, switch: keep or route to a competitor campaign, if the mention policy allows |   | competitor navigational | competitor plus login, support, status, careers: negative in competitor campaigns only |   | negative | 0 conversions, cost at least T, enough sample, and the step 6 checks pass (or covered by a phrase negative from step 7) |   | bid down | converts above T: lower its bid toward max CPL x its conversion rate, or qualify it |   | new keyword | not yet a keyword, 2+ conversions at or under T, and the matched ad group's ad and page answer it honestly: add it as exact in that ad group |   | new ad group | the same, but the matched ad group's ad and page don't answer it (a different promise or buyer): it needs its own ad and page |   | watch | spend under the thresholds |   | keep | converts at or under T |    With a qualified-outcome column, judge on cost per qualified lead instead of cost per conversion. 9. **Review AI Max separately.** From the AI Max rows (Google: add the Source column to the search terms report and filter match type "AI Max"; the "Search terms and ad combinations" view adds headlines and landing pages):    | Check | Rule |   | --- | --- |   | Too early | AI Max on for under 14 days: no negatives yet, only obvious safe-category ones (Google's guidance is to wait at least 2 weeks) |   | Brand and competitor leakage | AI Max serving brand or competitor terms in non-brand campaigns: set brand exclusions with a brand list before anything else |   | Landing pages | final URL expansion choosing careers, support, blog or other pages that don't convert: add URL exclusions |   | Cost per conversion | compare with the non-AI-Max non-brand pool, and read AI Max on cost per opportunity at account level once the judgment window has passed |   | Negatives | add only for terms that underperform consistently across campaigns, and say they came from AI Max | 10. **Run the script.** `python3 scripts/ngram_triage.py --csv terms.csv --brand ... --competitor ... --protect "<core phrase>" --threshold <T> --days <window>`. It returns totals, the flagged n-grams with a suggested match, term labels, brand navigational terms, the AI Max section and a first verdict. Then apply steps 6 to 9 to its candidates. The script can't see the reader's requests: a negative the reader asks for that fails a step 6 test turns the verdict to `NEEDS DECISION`. Without Python, do steps 2 to 5 by hand on the top 200 terms by cost (say so, and say what share of cost they cover), then steps 6 to 9. 11. **Write the change log.** One row per negative or keyword: date, level (account, list name, campaign), match type, the phrase, the source terms and their cost, the reason (flag and category), and how to reverse it. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No export with terms and cost || `INSUFFICIENT DATA` | Window under 30 days, or 0 non-brand conversions: only safe-category negatives are listed, everything else is watch || `NEEDS DECISION` | A flagged phrase fails the core-category check, or the reader asks for a negative that fails a step 6 test || `ACT` | At least 1 negative, new keyword or new ad group clears the rules || `CLEAN` | Nothing clears the rules and the waste rate is under 15% | Named stops: `BLOCKED` asks for the export. `NEEDS DECISION` stops on the conflicting phrase only, showing its spend, the converting terms that contain it and 2 or 3 options (phrase negatives on the bad pairs, the informational list on mid and bottom of funnel only, a cheaper offer such as a guide for that intent). Every negative that passed is still listed as ready. No other stops. ## Learn Append to `## Learnings` when: - the reader rejects a negative ("we sell to schools"): a `fact` that protects the word for this company;- a past negative is found to block converting traffic (impressions on a converting keyword fell after it): a `fact` naming the negative, and a `proposal` to loosen the rule that produced it;- a promoted term or new ad group reports cost per opportunity after the judgment window: a `fact`;- the reader sets a different threshold: a `proposal` with their reason. Example: `- 2026-10-24 | search-terms-triage | fact | Never negate "template": buyers search "attribution report template" before demos | Why: 3 opportunities from that term in Q3 (reader's CRM) | Status: noted` ## Output ```Verdict: <ACT | CLEAN | INSUFFICIENT DATA | NEEDS DECISION | BLOCKED> · Coverage: <complete | partial (what's weak)> · N-gram pass: <script | by hand, top n terms, x% of cost>Window: <dates, days> · Threshold T: <value> (<source>) Waste: <amount> of <non-brand cost> non-brand spend (<rate>, <band>) across <n> termsBrand: <cost>, <conversions> · Competitor: <cost> · AI Max: <cost>, <conversions> Negatives ready| Phrase | Match | Level or list | Category | Cost | Terms | Reason | Held for a decision| Phrase | Cost | Converting terms that contain it | Options | New keywords and new ad groups| Term | Conversions | Cost per conversion | Label | Why | Bid down| Term | Cost per conversion | Max CPC at its rate | AI Max<checks from step 9, each with what it found> Change log| Date | Level | Match | Phrase | Source terms | Reason | Reverse by | Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: search-terms-triageversion: 1.0.0verdict: ACTcoverage: completefor: <account, window>outputs:  waste: 6830  waste_rate: 0.21  negatives_ready: 14  held: [marketing attribution]  new_ad_groups: [attribution software for hubspot]  ai_max: brand exclusions missingnext: [ad-variants, lead-value-math]open: [reader applies negatives in Editor and logs them]``` `ad-variants` reads `new_ad_groups`. `lead-value-math` reads `waste` for the recovered-pipeline estimate. ## Failure modes | Problem | Do this || --- | --- || The export has no conversions column | `BLOCKED` for negatives; list the waste candidates by cost only and ask for the column || Conversions are form fills on a long sales cycle | Say so; prefer a qualified-outcome column; never promote on a single form fill || Low-volume terms are hidden by the platform | Block patterns (phrases) rather than single queries: suppressed queries are covered by the phrase || A brand term is also a common word | Ask the reader for the brand terms that are unambiguous; match the rest only with the category word next to them || The reader asks to push the negatives live | Recommend the Editor upload with the change log; this skill never applies changes || Lead volume and sales-accepted leads both fall after a cleanup | Over-negated: list the negatives added in the window and test removing the newest first | ## Influenced by - SHAKEUP's B2B Google Ads playbook: waste sized on 90 days of zero-conversion spend, the cost-weighted n-gram pass, the 11 negative categories and their deployment, and the false-positive checks.- Google's guidance on AI Max reporting (the Source column, the search terms and ad combinations view, brand lists, waiting 2 weeks before adding negatives): https://support.google.com/google-ads/answer/16470459?hl=en

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

Waste rarely sits in one term, so we total it by phrase and tie every threshold to your own cost per lead. We added the 4 checks because a wrong negative leaves no trace, and AI Max gets its own review because its matches behave differently.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Read your learnings first

    Words your team protected, or negatives that once blocked buyers, change the rules before the first row is read.

  2. 02

    Split brand and AI Max out

    Brand converts cheaply and hides waste, so it gets its own totals. AI Max matches get their own review.

  3. 03

    Size the waste

    Zero-conversion spend on non-brand terms, as a figure and a share. Over 25% means the account buys traffic it can't convert.

  4. 04

    Total the cost by word and phrase

    A word like "free" may cost little in each term and a lot in total. Its script adds up every 1, 2 and 3-word phrase across the export.

  5. 05

    Hold every threshold to your cost per lead

    A phrase becomes a candidate once it has spent what one lead may cost with nothing to show. Jobs, course and consumer words need no threshold.

  6. 06

    Check before anything is blocked

    Each candidate must pass 4 tests, covering converting terms, a plausible buyer, your own category and sample size. Then it gets a match type and a list.

  7. 07

    Make the call

    Every term gets a label and the run gets a verdict, with a change log you apply in the editor yourself.

What you need

  • Required

    A search terms export from Google Ads or Microsoft Ads, ideally 90 days

    A connected ad account or reporting connector → A CSV export you upload → Rows you paste

  • Required

    The most a lead may cost

    The handoff from lead value math → Your company profile → The account's own cost per conversion, marked as the weaker source

  • Required

    Brand terms, competitor names and your category words

    Your company profile → Your answer

  • Optional

    Qualified outcomes per term from your CRM

    A CRM join on click ID → A column you add to the export

What it can return

  • ACT

    At least one negative, new keyword or new ad group clears every rule.

  • CLEAN

    Nothing clears the rules and waste is under 15% of non-brand spend.

  • INSUFFICIENT DATA

    Under 30 days of data or no conversions, so only the safest negatives are listed.

  • NEEDS DECISION

    A candidate would block your own category word, or a requested negative fails a check.

  • BLOCKED

    There is no export with terms and cost.

Where it stops for you

  • When there is no search terms export
  • When a negative would block your own category or a converting term

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run on 90 days of search terms for a made-up B2B SaaS company that sells attribution software, with no free plan. Its max cost per lead is $285.

Verdict

ACT

Coverage: complete

Waste

$4,018 of $10,288 non-brand spend (39%, over the 25% line) across 10 terms. Brand spent $190 and AI Max $830, both reviewed apart.

PhraseMatchWhereSpendWhy
freeBroadAccount$1,3260 conversions across 2 terms, and there's no free plan
what isPhraseInformational list, mid and bottom of funnel$960Learn intent on a demo campaign
open sourcePhraseOff-ICP list$336Self-hosting buyers, which the product doesn't serve
jobs, salary, courseBroadAccount$702Job seekers and students, safe at any spend
Not blocked

"attribution" and "software" appear in every converting term. "excel template" spent $216, under the $285 line, so it waits.

New keywords and ad groups

"marketing attribution software" (3 conversions at $217) goes in as a new exact keyword. "attribution software for hubspot" (3 at $173) gets its own ad group and the HubSpot integration page.

AI Max

it served the brand's own reviews query and a competitor's pricing query. Set brand exclusions with a brand list before adding any AI Max negatives.

The handoff block the next skill reads
skill: search-terms-triage
version: 1.0.0
verdict: ACT
coverage: complete
for: Google Ads, last 90 days
outputs:
  waste: 4018
  waste_rate: 0.39
  negatives_ready: 7
  held: []
  new_ad_groups: [attribution software for hubspot]
  ai_max: brand exclusions missing
next: [ad-variants, lead-value-math]
open: [apply negatives in Editor and log them]

What good looks like

  • A waste figure with its share of non-brand spend, brand kept apart
  • Negatives as phrases with a match type, a list and a reason, not a pile of single queries
  • No negative on a term that converts, or on your own category word
  • Converting terms you don't bid on yet, each marked as a new keyword or a new ad group
  • A change log you could reverse line by line

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Why tie the threshold to cost per lead?

Because a phrase that has spent what one lead may cost, with no lead, has had a fair chance. A fixed spend figure would be too strict for one account and too loose for another.

Will it block words that some buyers type?

It checks first. A word in any converting term never becomes a broad negative, and your own category words are never blocked. Anything that fails a check is held for your decision.

How does it handle AI Max?

Separately. It checks whether AI Max serves your brand or competitors' names, which landing pages it picks and its cost per conversion. AI Max negatives wait for the first 2 weeks, as Google advises.

Does it change the account?

No. It writes the negatives, the match types, the lists and a change log. You upload them in Google Ads Editor or Microsoft Advertising Editor, so you see the diff first.