SOLUTIONS · MARKETING & SEO
Site search query classification: see what shoppers want, in your own catalog categories
Classify site search queries into your own catalog categories. Review zero-result searches and possible gaps in demand using your search history.
What shoppers type into your site's search box is the most direct demand data you have: "waterproof hiking boots wide", "gift for dad under 50", "the blue one from instagram". Site search query classification assigns searches into your own catalog categories. Then you can:
- see demand by category;
- prioritize searches that return nothing;
- investigate demand your catalog may not serve.
StayCharted AI Model Trainer learns your categories from the searches that already led to a product click or a purchase.
Why your own categories, and why a trained model
A general AI can tell you that "trail runners" are shoes. Without your catalog rules, it may not know that your store files them under Outdoor → Running, not Footwear. It also needs context to know that your customers say "puffer" for what your catalog calls "insulated jackets". Your catalog's boundaries are your own, often hundreds of categories deep.
A model trained on your shoppers' own searches learns those boundaries. It then sorts tens of thousands of searches a month using the same published model, with a confidence on each one.
What it reads, and what it answers
It reads the search as the shopper typed it, such as “waterproof hiking boots wide”. It suggests one of your catalog categories, such as Outdoor › Hiking boots, alongside a confidence score.
The AI Classifier Model supports semantic similarity and multiple languages. Test slang, misspellings and the languages your shoppers use on held-out examples; improvement over your existing rules depends on your data.
How to set it up
1. Export searches that led somewhere
You need searches paired with where the shopper went. Use search analytics that connects a search event to a subsequent product click or purchase. GA4 site-search measurement records search terms, but linking those searches to products and purchases needs appropriate ecommerce event tracking and attribution; enabling site search alone does not provide these training pairs. Export one row per search, with the category of the product the shopper clicked or bought. That category is the answer the model learns, so you can start with candidate labels rather than label every row from scratch.
Review the pairs and keep searches whose clicks or purchases clearly point to one category. A click is not always a correct label. Check category coverage, remove personal information and split by distinct query so repeated searches do not appear in both training and testing.
2. Train
Upload the file, or use the Google Sheets add-on. Choose the search as the input and the category as the answer, then train an AI Classifier Model. The report shows how often it picked the right category for searches it never saw, category by category.
3. Classify every search, including the ones that found nothing
Export a month of searches, including zero-result searches and searches with no clicks, and fill the category column. Each filled row has a suggested known category and a confidence score.
4. Act on it
| What you see | What to do |
|---|---|
| Zero-result searches in a category you stock | Add a synonym or a redirect in your search tool, so "puffer" finds insulated jackets |
| Many confident searches for a category with few products | A buying or merchandising signal |
| Searches with low confidence in every category | Ambiguous searches, weak training examples, or possible unmet demand to investigate |
| Demand by category month on month | Merchandising, stock and paid-search budgets |
5. Repeat monthly
Review and add the new searches that led to clicks to the training file, and retrain every month or two. Retrain whenever your catalog's categories change.
What it doesn't do
- It doesn't replace your search engine or rank products. It tells you which category a search is about.
- It doesn't change your site. You add the synonyms, redirects and merchandising rules in your search tool.
- It doesn't know categories you've never sold. It can assign an unfamiliar search to a known category, sometimes confidently. Low confidence helps prioritize review but does not reliably identify missing products. Check a sample of confident answers too.
Common questions
What is site search query classification?
Sorting the searches people type into your website's search box into categories, usually your own catalog categories, so you can see demand, fix failed searches and improve results.
How do I fix zero-result searches?
Use suggested categories to prioritize failed searches, then check your catalog and search configuration. Missing synonyms, filters, spelling, unavailable stock or genuinely missing products can all cause zero results. A category prediction alone does not establish the cause.
Where does the training data come from?
Start from search history linked to product clicks or purchases, then review the candidate category labels. Your analytics must connect the search to the product; a search-term count alone is not enough.
Can it handle misspellings and other languages?
The AI Classifier Model supports multiple languages and semantic similarity. Test the actual misspellings, slang and languages used by your shoppers before relying on the model.
Is this the same as SEO keyword mapping?
It's the same idea applied to your own site's search box instead of Google. For Google searches, see AI keyword mapping with Google Search Console.