Automate the Google Ads search terms report with a model trained on your own decisions

The search terms report's Added/Excluded status can help identify candidate training labels. It describes targeting status, not proof that a person reviewed the term or a dated record of what they decided.

Google's statuses include ADDED, EXCLUDED, ADDED_EXCLUDED and NONE. “None” means neither targeted nor excluded; it does not mean “reviewed and approved to leave”. Confirm the intended decision and campaign context before using a row for training. Review both-added-and-excluded cases separately.

Combine the export with your team's decision log where available. This gives you useful starting data while keeping the training labels aligned with your actual policy.

Ask ChatGPT whether "free crm software" should be a negative, and it needs your account context and instructions. For example:

• your account sells to enterprises, so "free" is always a negative;
• "crm for small teams" converted well last quarter, so it stays;
• your team excludes competitor names in one campaign but bids on them in another.

Those rules live in your past decisions. A model trained on them applies them the same way every week, at the volume of a large account, without writing a new classification prompt for each batch. Compare it with your current process on held-out terms.

It reads the search term, plus the context your team looks at when deciding:

It answers in your own categories, for example:

• Add as keyword
• Add as negative
• Leave

Every answer comes with a confidence, to prioritize review. Confidence is not a guarantee that excluding a term is the right business decision.

1. Download past search terms with their decisions

In Google Ads, open the Search terms report. Pick the last 6–12 months and add the Added/Excluded column, along with campaign, ad group and keyword. Download it as a spreadsheet.

2. Build labels from confirmed decisions

Keep “Leave” labels only where your team confirms that decision. Click volume does not establish that a term was reviewed. Confirm the scope and intent of added or excluded terms, resolve Added/Excluded combinations, and omit unresolved rows. Rename the reviewed answers to your chosen categories.

Remove personal information and unrelated identifiers from the export. Keep repeated terms with the same context together when splitting data, so duplicates do not inflate test results.

3. Train

In StayCharted, or from the sheet with the Google Sheets add-on, train an AI Classifier Model:

• Inputs: the search term, campaign, ad group and keyword.
• Answer: your decision column.

The report shows how often it agreed with your team on terms it never saw.

Test it on a later period too: train on terms before the test month, fill that later month’s terms, and compare with what your team actually decided.

4. Run it every week

1. Download the last 7 days into a fresh tab, preserving your prior results.
2. Fill an empty decision column with the add-on; existing answers are not overwritten. Each filled term gets a suggestion and a confidence.
3. Sort by confidence. Review negative-keyword wording, match type, scope, spend and conversions before applying approved negatives in bulk: in the Google Ads interface, in a negative keyword list, or with Google Ads Editor. Prioritize unsure answers and sample confident ones too.
4. Add the terms your team corrected to the training sheet, and retrain every month or two.

• It doesn't change your account. You apply what's suggested, with the tools you already use.
• This model uses the input columns you selected. It does not fetch live account metrics or optimize bids. Check spend and conversions before excluding anything expensive.
• It doesn't see every search. Google's report leaves out terms searched by very few people.
• It follows your past decisions, including the mistakes. If your team excluded terms that later proved valuable, correct those rows before training.