Category: Product updates

A shopper types “trainers”. Your store sells them – but the product is tagged “Sneakers” and the description never uses the word. Until this release, that product stayed invisible. AI Search 1.30.0 fixes the oldest known limitation in the plugin: products are now found through their categories and tags, not only through their own text.

What was broken

Since version 1.26 the plugin has understood taxonomy terms: every category, tag and attribute term gets its own meaning fingerprint, and a product whose term matches the query gets a scoring boost. The catch was that this only ever worked as a bonus on products the search had already picked up through their own description. A product whose only real signal was its category could never clear the bar – its content score was near zero, and a bonus on almost nothing is still almost nothing.

In practice that meant well-tagged catalogs were leaving matches on the table, precisely for the queries where customers use their own vocabulary instead of yours.

What changed

Search now runs a second retrieval pass. The query is compared once against every term that has an embedding; terms that genuinely match pull in the products attached to them, straight from the taxonomy tables. Those products join the results with a deliberately conservative score – a term-only match is real evidence, but weaker than a product whose own text matched, and it can never outrank an exact SKU or title match. Nothing that already worked ranks lower; the results simply include products that were previously missing.

How we picked the cutoff

A term only pulls in its products when it is close enough to the query – and picking “close enough” deserved measurement rather than a guess. Our first instinct was 0.60 similarity. Testing against a real “Sneakers” tag showed why that would have shipped a feature that does nothing: “trainers” scores 0.53 against it, “shoes” 0.61, “sneakers” itself 0.82. Terms are short strings, so their similarity scores sit lower than document-length text.

The same test run against unrelated queries gave the other half of the answer: “kitchen table” 0.19, “umbrella” 0.18, “winter coat” 0.26. Synonyms cluster above 0.5, unrelated queries below 0.3, and the cutoff belongs in the gap – so it sits at 0.45. If your taxonomy needs something different, the new ai_search_term_match_threshold filter takes a per-taxonomy value.

Do I need to do anything?

Update the plugin, and make sure term embeddings are switched on and generated – that is the Indexing tab, where “Term embeddings” shows how many of your terms are indexed. WooCommerce stores have this on by default. If your terms were indexed back in 1.26, there is nothing else to do: the new retrieval pass uses the embeddings you already have.

The ranking order, unchanged

Worth repeating, because it is the promise the whole plugin rests on: exact SKU and title matches first, then everything your traditional keyword search finds, then AI results – and now, among those AI results, products reachable through their categories and tags. You can watch all of it happen on the live demo, including what happens when you push the similarity threshold too high.

Get AI Search 1.30.0 on WordPress.org – free up to 10,000 embeddings, no API key required.

Published August 2026.

Two releases shipped this month, and together they change how AI Search ranks every result. Here is what is new in 1.27 and 1.28, and why the changes matter for your store.

Exact SKU and title matches always come first (1.27)

Random-looking codes like Amazon ASINs mean nothing to an AI model, so a shopper pasting a SKU used to get poor results – the exact product they named could score below the similarity threshold. Now the plugin checks the query against product SKUs (exact and prefix, with variation SKUs resolving to their parent product) and exact titles before any AI work, and pins matches to the top with a green “Exact Match” badge for editors. Always on, no setting. Developers can add their own rules – EAN, supplier references – through the new ai_search_exact_matches filter.

Bulk indexing tells you what failed, and why (1.27)

Embedding generation used to fail silently – connection fine, nothing indexed, no explanation. The Generate Embeddings screen now reports exactly what happened: how many items succeeded, how many failed, and the cause per group – connection errors, rejected token, exhausted quota, empty content. Provider-level errors stop the run early instead of burning your whole batch on a repeating failure.

Traditional results are never lost (1.28)

The biggest change of the year. Results now come in three strict tiers: exact matches first, then every keyword match your traditional search would find – reordered by relevance instead of publish date – and only then AI results, appended below. Installing AI Search can no longer make any search worse: if a product literally matches the query, it shows, in a better position than default WordPress would give it. We wrote up the full reasoning in AI Search vs the default WordPress search, and you can try the tiers yourself in the live demo.

Smaller fixes worth knowing

Site-wide searches (no post type specified) previously skipped AI ranking entirely due to a post-type resolution bug – fixed. WooCommerce catalog visibility is now respected everywhere, so products hidden from search stay hidden in every tier. Result counts shown by themes now match what the plugin actually returns. And 1.28.1 fixed tier treatment for secondary queries, so blocks and widgets that run their own search get the same ranking as the main one.

Update or install AI Search on WordPress.org – the free tier covers 10,000 embeddings, no API key required.

Published August 2026. Full changelogs on the plugin page.