We wanted to know a simple thing. If a visitor arrives on this site and searches for something, will they find the right page, and would an AI search tool do it better than the search WordPress already includes? So we ran the same five questions through three different routes on this site, on October 27, 2026, and wrote down what actually happened.
The baseline: built-in WordPress search
The baseline was the search WordPress ships with. We used the front-end search and ran five questions, exactly as a visitor would type them. For “how much did the chatbot cost,” it returned 1 result, and the top result was right: Experiment 001. For “chatbot answers,” it returned 2 results, and the top result was right again. For “site design,” it returned 1 result, the day-one post, which was right. For “privacy,” it returned 1 result, the Privacy Policy page, which was right. For “who runs this site,” it returned 3 results, but the top result was Experiment 001, not the About page, so that one was wrong. That is 4 right out of 5 at the top, which is better than we expected from plain keyword search on a small site.
The dedicated AI plugin branch stopped early
Next we tried one dedicated free plugin, AI Vector Search (Semantic), version 1.0.2. Its listing describes WooCommerce search, and that turned out to be the problem. Its free Lite mode runs locally, but it is described as TF-IDF search with synonym expansion, not semantic search, and it is built around products. While it was active, a front-end search for “privacy” returned 0 results, even though the Privacy Policy page exists. The site has no WooCommerce products, so a product index had nothing useful to return for posts and pages. The settings also pointed to a Supabase setup, including a master key and SQL to run in Supabase. We did not enter any key, create any Supabase account, or pay anything. Because the plugin only worked for products in this setup, we stopped that branch, deactivated the plugin, and deleted it. After deletion, we checked again: a search for “privacy” returned 1 result, the Privacy Policy page, just as it had before.
Chatvox as site search
Finally, we asked the Chatvox chat widget in the bottom-right corner of the live site the same five questions. It answered in sentences rather than returning a list of pages, so we graded each answer as right, hedged, or wrong. On “how much did the chatbot cost,” it was right: it said the chatbot test had no extra cost and ran on the existing Chatvox team plan, with the plugins used costing $0. On “chatbot answers,” it was also right: it said the chatbot answered 2 out of 3 questions correctly before tuning and 3 out of 3 after a targeted Q&A entry was added, and that its main weakness was hedging instead of giving a number plainly stated on the site.
On “site design,” it was wrong in an interesting way. It said the site was still using the Twenty Twenty-Five theme, which is true, but it also said the site was still in maintenance mode, which did not match what visitors saw: the front end loaded normally during the test. On “privacy,” it hedged. It said it did not have verified details and pointed to the Privacy Policy page, even though that page was open behind the widget and the answer could have been more direct. On “who runs this site,” it also hedged. It said WPLLMs is run by the site owner, but that the available information does not identify them by name, and it pointed to the About page. That is 2 right, 1 wrong, and 2 hedged out of 5.
Verdict
For this small site, the built-in search did the plain job better than the chat widget did. It put the right page first in 4 out of 5 cases. The chat widget was helpful when the question was about a number or a result from an earlier experiment, but it hedged on privacy and ownership and gave a stale maintenance-mode answer on design. The dedicated AI plugin never got to compete, because its free mode was product search, not post and page search, and its full setup needed Supabase. The whole experiment cost $0 extra. The Chatvox update applied before testing, from version 1.2.3 to 1.4.0, was free, and the test plugin was free to install and then deleted. Anything that would have cost money here would have been the paid side of a semantic search setup, such as a Supabase plan or a paid plugin tier, and we did not buy, sign up for, or enter a key for any of it. If a future plugin offers real semantic search over posts and pages with no key and no account, it is worth testing the same five questions again. Until then, built-in search stays.
