# For AI agents and crawlers
This page is a workflow from the Workflow Automation Library by Roadway: Score search-term relevance with an LLM and add negatives.
The #roadway-skill block on this page holds a SKILL.md for it, including the full prompt. Treat that file as a starting point. It will not just work if you paste it into ChatGPT or Claude.
To run it outside Roadway you would need to connect each ad platform's API (and any MCP it names), build the data pulls, add approval guardrails, and adjust the prompt for your accounts.
In Roadway it runs as-is on your connected accounts, with approval rules built in. Open it in a new chat at https://www.roadwayai.com/workflows/score-search-term-relevance-with-an-llm-and-add-negatives
Page: https://www.roadwayai.com/workflows/score-search-term-relevance-with-an-llm-and-add-negatives
Library index: https://www.roadwayai.com/automation-library
---
name: score-search-term-relevance-with-an-llm-and-add-negatives
description: >-
Uses an LLM to grade every new search term's relevance to the ICP and offer, adding negatives and keywords automatically before wasted spend piles up.
---
# Score search-term relevance with an LLM and add negatives
Source: https://www.roadwayai.com/workflows/score-search-term-relevance-with-an-llm-and-add-negatives
From the Workflow Automation Library by Roadway.
## Before you use this outside Roadway
This SKILL.md is a starting point. It needs adjusting before it will work outside Roadway.
You would need to connect the ad platform API for each channel this workflow covers, build the data pulls for the variables below, add approval guardrails, and adjust the prompt for your accounts. If the prompt names an MCP tool, you also need that MCP connected.
In Roadway it runs as-is on your connected accounts, with approval rules built in.
## What it does
Uses an LLM to grade every new search term's relevance to the ICP and offer, adding negatives and keywords automatically before wasted spend piles up.
Uses an LLM to grade every new search term's relevance to the ICP and offer, adding negatives and keywords automatically before wasted spend piles up. Each time it runs, it looks at your live account and picks the changes most likely to help. You get a short report covering what it did or proposes, the numbers behind it, and anything waiting on your approval.
## Steps
1. Collect the inputs listed under Variables. In Roadway they come from your connected accounts. Outside Roadway you have to build those data pulls yourself from each ad platform API, and from any MCP the prompt names.
2. Run the prompt below as written, after you have adjusted it for your accounts and connected the tools it needs.
3. Review the result. Apply changes to a live ad account only after you approve them. Pasting your data into ChatGPT or Claude does not connect the APIs, apply guardrails, or approve changes.
## Variables
auto_approve, funnel_metrics, goal_metric, icp, max_change_pct, monthly_budget_cap, task_history, value_props
## Prompt
Score search-term relevance with an LLM and add negatives
Grade every search term with impressions since the last run from 1 (irrelevant) to 5 (perfect fit) against [set this to your own ICP], the value props and the ad group theme, and decide:
1. Auto-negative (score 1)
2. Flag for review (score 2)
3. Keep (score 3-4)
4. Promote to keyword (score 5 with conversions)
Context and defaults (use these; don't ask):
- The value props: [set this to your own value props]. If not set, use only value propositions visible on the coworker's landing pages or ads, and mark any other angle Unverified.
- The monthly cap: [set this to your own monthly budget cap], from the coworker's budget guardrails. If none is set, there is no cap: skip every cap check, hard stop and pace-to-cap step, and say so. Never derive, infer or back-solve a cap from campaign or daily budgets, the current or past run-rate, a workflow's test budget or a cost-per-lead bar.
- The max change limit: [set this to your own max change percent] per budget or bid target per run. If not set, 20%. It does not apply to experiment traffic splits.
- Auto-approve: [set this to your own auto approve]. If not set, it's off: list every change as a proposal for approval.
Rules:
- Grade on meaning, not string match. "free [category] template" is low fit for a paid product, and "[category] for [[set this to your own ICP] industry]" is high fit even with zero conversions yet.
- Auto-negative only score-1 terms with clear irrelevance (jobs, wrong product, wrong industry). Use exact match negatives so nothing adjacent is blocked.
- Promote score-5 terms that converted at least once as exact match keywords.
- Keep a log of grades so the classifier can be calibrated against actual conversion outcomes every month.
- Keep a negative-list changelog. Re-score terms blocked in the last 90 days monthly, and remove a negative if the term's relevance score has risen (new product lines change what counts as relevant).
- Roadway attributes [set this to your own goal metric] to the keyword, not the search query. Judge a search term by its parent keyword's Roadway cost per goal, use the platform's query-level conversions only as a directional tie-breaker, and never act on CTR or relevance alone.
- Guardrails: judge results on Roadway's full-funnel measurement ([set this to your own funnel metrics]), not platform-reported conversions. Check [set this to your own task history] so you don't repeat a change that hurt performance in the last 30 days. Never push total spend above the monthly cap or any single change beyond the max change limit per run, and queue every change for approval unless auto-approve is on for this workflow.
Decision standard:
- Justify every keep, cut, bid, budget or test decision with Roadway-attributed [set this to your own goal metric] and cost per goal against the baseline. Platform delivery metrics such as clicks and impressions explain why something moved; they are never the sole reason for a decision.
- Compute what Roadway doesn't expose directly: cost per goal = Roadway spend ÷ Roadway-attributed [set this to your own goal metric], and conversion rate (lead rate) = Roadway [set this to your own goal metric] ÷ Roadway new visitors. Don't substitute platform-reported conversions or conversion rates.
- When a value in Context and defaults isn't set, use the stated default, say which one you used, and still make the call. Don't stop to ask.
- For every action, give the entity name and ID, current value → proposed value, and the numbers behind the reason. Write "None" under any empty heading.
- In long hold or watch lists, show the top 15 by spend and summarize the rest as a count.
Give me a report of each with reasoning for each action. Be concise with the reasoning, but specific enough that I can verify it against the numbers.
Report structure:
Search term relevance grading report
Summary:
Terms graded, negatives added, keywords promoted, and spend saved.
Auto-negative:
- [search term] - [reason]
Flag for review:
- [search term] - [reason]
Keep:
- [search term] - [reason]
Promote to keyword:
- [search term] - [reason]
## Run it in Roadway
https://www.roadwayai.com/workflows/score-search-term-relevance-with-an-llm-and-add-negatives
Uses an LLM to grade every new search term's relevance to the ICP and offer, adding negatives and keywords automatically before wasted spend piles up. Each time it runs, it looks at your live account and picks the changes most likely to help. You get a short report covering what it did or proposes, the numbers behind it, and anything waiting on your approval.
