# Test and evaluate n8n marketing workflows

> Test n8n workflows before they go live: pinned data, test sets, AI evaluations, debugging past executions and Insights for time saved.

Source: https://mentionx.ai/learn/n8n/n8n-testing-evaluations  
Author: Dipuck Jones, founder of MentionX  
Updated: 4 November 2026  
Level: Practitioner

**Testing in n8n means pinned data and manual runs for logic, test sets and evaluations for AI steps, and Insights after launch. AI steps need the most attention, because the same prompt can give different answers and a model change can quietly make results worse.**

> **Key takeaways**
>
> - Pin data and test edge cases before publishing.
> - Score AI steps on a test set before and after changes.
> - Use test channels so tests don't message real people.
> - Track production health and time saved with Insights.

*Checked 4 November 2026 (CET) against n8n documentation and the n8n pricing page. n8n ships updates often and node options, plan limits and prices change, so confirm in your own instance and on the pricing page. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.*

## Why marketing workflows need testing

Marketing automations touch leads, customers and brand channels. A wrong branch can route a hot lead to nurture, send an unapproved post or overwrite CRM fields. The n8n docs on why to test explain that AI outputs are variable, so you need evidence a change helps before you ship it.

## Before publishing

Checklist:

- **Pin data** for normal cases and edge cases: empty fields, odd characters, duplicates, very long text.
- **Run each branch** of IF and Switch nodes at least once.
- **Point outputs at test destinations:** a test Slack channel, a copy of the sheet, a sandbox CRM if you have one.
- **Check the timezone** in workflow settings.
- **Set the error workflow** (lesson 16).

## Testing AI steps

For classifiers, extractors and drafts, build a small test set: real inputs with the answer you expect. For classification, matching is simple. For drafts, score with a checklist (uses only proof facts, right length, right tone) or use a model to grade against criteria, checked by a person.

Rerun the test set whenever you change a prompt, a model or a node version.

## After publishing

The executions list shows every production run. You can open a failed execution and load its data into the editor to debug with real inputs. Workflow settings control which executions are saved (lesson 20 covers storage).

Insights shows production execution counts, failure rates and run time. Setting estimated time saved per workflow turns that into a figure leadership understands. Be honest with time-saved estimates: count only work people actually stopped doing.

*Our view:* the test set is the most valuable thing you build for an AI workflow. Models change; a test set tells you whether to change with them.

> **Pro move:** add every real failure to the test set. Over time it becomes a record of the cases your workflow must handle.

## Testing tools in n8n

| Tool | Use |
|---|---|
| Pinned data | Fixed inputs so you test the same case repeatedly |
| Manual executions | Run and inspect each node's output |
| Executions list | Debug production runs and load their data into the editor |
| Evaluations | Run a dataset through AI steps and score outputs |
| Insights | Track production executions, failures, run time and time saved |

## Evaluations

The n8n docs describe two kinds. **Light evaluations** happen before deployment: run a handful of test cases and compare outputs by eye. **Metric-based evaluations** run after deployment on a larger dataset with scores, such as correctness or whether a category matched the expected one, so you can compare prompt or model changes.

## Insights

According to the Insights docs, a summary banner with the last 7 days of production executions shows for all plans. The full dashboard is available on Cloud Pro and Enterprise and self-hosted Business and Enterprise. You can set estimated time saved per workflow, fixed or dynamic. Insights excludes manual and sub-workflow executions.

## A pre-launch checklist

- Pinned data covers normal, empty, duplicate and long inputs.
- Every branch ran at least once.
- Test destinations used; no real messages sent.
- Workflow timezone set.
- Error workflow set.
- AI steps scored on a test set.
- Owner and backup named in the description.
- Estimated time saved set honestly.

## Building a good test set

| Include | Why |
|---|---|
| Typical cases | Most of your volume |
| Edge cases | Empty fields, odd formats, other languages |
| Past failures | Stop repeats |
| Ambiguous cases | Check the "unclear" route works |
| Adversarial inputs | Text that tries to override instructions |

Label expected outputs with the person who owns the decision, such as a sales manager for lead routing. Their labels define correct.

## Setting up an evaluation, node by node

| Step | Node and settings |
|---|---|
| Dataset | A Google Sheet or data table with `input` and `expected` columns |
| Trigger | Evaluation trigger, pointing at the dataset |
| Workflow under test | Your classifier or extractor steps, unchanged |
| Record outputs | Evaluation node set to write the output back to the dataset row |
| Metric | Evaluation node set to calculate a metric, such as a categorization match comparing output to `expected` |

Run the evaluation from the editor. Each run gives a score you can compare across prompt or model changes. The docs describe both deterministic metrics and AI-judged ones; use deterministic matches wherever there is a single correct answer.

## Debugging a production failure

1. Open **Executions** and filter by failed.
2. Open the execution and find the red node.
3. Read the input that node received. Most failures are unexpected input.
4. Use the option to debug in the editor, which loads the execution's data so you can rerun with the real input.
5. Fix, then add that input to your pinned data or test set.

## Cost math

Evaluation runs are manual and use model calls for every row. A 40-row test set run five times while tuning is 200 model calls. That's cheap for classification with a small model, but check costs before running large datasets through expensive models.

## Ledgerline walkthrough: comparing two models

*Ledgerline is a fictional accounts-payable software company. Details are illustrative.*

When a newer, cheaper model became available, marketing ops ran the 40-row lead test set through both. Scores were close, but the cheaper model mislabelled two German leads. They kept the existing model for German-language forms using a Switch node on a detected language field, and used the cheaper model for the rest.

## Worked example: Ledgerline's lead classifier evaluation

*Ledgerline is a fictional accounts-payable software company. Details are illustrative.*

Ledgerline's lead classifier sorted form answers into "demo-ready", "researching" and "not a fit". Sales complained about misrouted leads.

Marketing ops built a 40-row test set from past leads, labelled by a sales manager. The current prompt matched 31 of 40. Adding category descriptions and an "unclear" category raised matches to 36, and unclear leads went to a person. They also found two failures caused by forms in German, and added those to the test set.

### Walkthrough

1. Pick the lead classifier from lesson 8 or the routing workflow from lesson 6.
2. Make a test sheet with 30 real (anonymised) inputs and the correct expected category for each.
3. Use the evaluation trigger to run each row through the workflow.
4. Compare output with expected and record match or no match as a metric.
5. Note the score. Change the prompt or model, rerun, and compare.
6. Publish only changes that score at least as well.
7. In workflow settings, set estimated time saved, and check the Insights banner after a week.

## Try it yourself

Build a 30-row test set for one AI step. Score your current prompt, then try one change. Keep whichever scores higher, and write down the score in the workflow description.

## Quick quiz

**1. What is pinned data for?**

*Answer:* Testing with the same fixed input each time.

**2. What do metric-based evaluations do?**

*Answer:* Score AI outputs on a dataset so you can compare changes.

**3. Does Insights count manual executions?**

*Answer:* No. Manual and sub-workflow executions are excluded.

## What should you check before the next lesson?

- Edge cases are pinned and tested.
- AI steps have a scored test set.
- Time saved is set and estimated honestly.

Next: [Security, credentials and governance](https://mentionx.ai/learn/n8n/n8n-security-governance). Or return to the [n8n for Marketers track](https://mentionx.ai/learn/n8n/n8n-for-marketers).

## Common questions

**How many test cases do I need?** Start with 20 to 30 real examples covering normal and edge cases.

**Can I test without sending real messages?** Yes. Point Slack and email nodes at a test channel or disable them during tests.

**How do I show the value of automation?** Set estimated time saved per workflow and use Insights.

## Sources

- n8n Docs, [Understand why to test](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test)
- n8n Docs, [Track usage with Insights](https://docs.n8n.io/administer/observe-and-log/track-usage-with-insights)
- n8n Docs, [Manage execution data](https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/manage-execution-data)
- n8n Docs, [Configure workflow settings](https://docs.n8n.io/build/manage-workflows/configure-workflow-settings)
- All accessed 4 Nov 2026.

## Frequently asked questions

### How many test cases do I need?

Start with 20 to 30 real examples covering normal and edge cases.

### Can I test without sending real messages?

Yes. Point Slack and email nodes at a test channel or disable them during tests.

### How do I show the value of automation?

Set estimated time saved per workflow and use Insights.

---

Written and reviewed by Dipuck Jones for MentionX Learn. The HTML version of this page: https://mentionx.ai/learn/n8n/n8n-testing-evaluations