WP Engine Just Opened Smart Search AI to Every WordPress Site. I Ran Six Queries At It.

On 11 August 2026, WP Engine announced that Smart Search AI is no longer limited to sites hosted with them. For thirteen months it had been an add-on for their hosting customers. Now any WordPress site can run it.

That makes it worth understanding, so I spent an afternoon running the same small set of queries against five live WordPress site searches, including one that WP Engine has publicly named as a Smart Search customer.

I want to be precise about what that means, because it limits what follows.

Only one of the five sites is confirmed by WP Engine's own published material to be running Smart Search. That confirmation dates from 2024. The other four are live WordPress sites with search boxes, and I do not know what powers them. So this is not "five instances of Smart Search." It is five site searches, one of them a documented Smart Search customer, and one deliberately included as a control because it runs a well known third-party search plugin instead.

I am not naming the sites. They are somebody's customers, not my competitors, and the point of the exercise is the behaviour, not the brand.

Every search term came from content visible on the site being tested. That rule matters more than it sounds. Without it you cannot tell "the engine failed to find it" from "they do not sell it," and you end up reporting the second as if it were the first.

Before the results, here is the one worth doing on your own site right now.

Search for "X without Y", where X is something you sell and Y is something the customer does not want. Then search for just "X".

Compare the two numbers.

Again and again, the first query returned more results than the second. Not fewer. One site went from 1,340 results to 4,607, and four of the five top hits were the exact thing the query excluded. Another went from 301 to 647, and the excluded term was promoted into the top three. A third went from 37 to 48, with the excluded document moved to position one.

A fourth site did something different and equally unhelpful. It went from 2 results to 0, destroying a query that had worked a second earlier.

Nobody is hiding why. WP Engine documents it plainly:

"If no operator is specified, OR is used by default.
 For example, red leather couch is interpreted as red | leather | couch."

That is from the Smart Search Find API documentation on developers.wpengine.com.

So "lobster rolls but no Maine" becomes lobster OR rolls OR but OR no OR Maine. The word your customer used to exclude something quietly becomes another word to search for.

This is documented, intended behaviour, not a defect. It is only a problem in the moment a customer types a sentence instead of a noun.

Gibberish is a bad test. Type qwertzuiop into anything and you get an honest zero. Every engine passes.

The interesting case is a real word, correctly spelled, that the site simply does not cover. Site after site returned results anyway: 4 unrelated documents for one, 258 for another, 958 for a third.

Then I took the queries apart, one token at a time, and the mechanism showed up.

On one site, "snowboard bindings" returned 258 results. "snowboard" alone returned No results found. "bindings" alone returned the same 258, with an identical facet distribution and identical result positions. The meaningful token had been silently discarded, and "bindings" had fuzzed onto "buildings."

On another, "carburetor gasket" returned 958 results. "carburetor" alone returned a literal zero. "gasket" alone returned the same 958, in the same order, with the same twelve titles. In those results the word "gasket" appeared once, as an echo of the query. "casket" appeared twice. "market" appeared eight times.

This is the failure mode that gibberish will never surface. A query made entirely of unmatched words fails honestly. A query in which one word merely looks like a common word fails silently and hands back a four-digit result count.

Worth saying clearly: each of those three sites returned a hard zero on a different query in the same session. The padding is selective, not a template that is never empty.

I took a concept each site demonstrably covers, checked it worked in English, then asked for the same thing in German. I picked German words that share no letters with their English equivalents, so no spelling coincidence could build a bridge.

Four out of four. Every English control passed in the same session, so each zero is a live comparison rather than an artefact of a broken index. On the confirmed Smart Search site, the response carried an empty result set and no "no results" message at all.

One caveat that belongs here rather than in a footnote: these are English-language sites serving English-speaking markets. A German zero is a finding about capability, not a complaint about service. Nobody promised those sites would answer in German.

But it is the shape of failure that interests me, because it cannot be papered over with a synonym list. Nobody hand-writes translations into a hundred languages.

It will invent four unrelated results for an English phrase that is not in the catalogue, but return a hard zero for a concept it does handle, asked in German.

The relaxation is anchored in English and it is lexical rather than conceptual. It stretches spelling. It does not stretch meaning.

One more thing worth knowing before anyone reads too much into any of this. Smart Search offers three modes, and the site owner picks: full text, semantic, or hybrid, tuned with a bias value from 0 to 10, where 0 is keyword only and 10 is semantic only. So "the product has semantic search" and "this particular site has it switched on" are different statements, and from the outside you cannot tell which you are looking at. That applies to every result above.

I build a search plugin, so weigh this accordingly. Two things about mine.

It does not solve negation with a filter either. It rewrites "without X" into a positive form, which works when the catalogue contains something positive to match and drifts when it does not. Better, not solved.

And it pads too. Give mine a real word it has never seen and it will return things from the same neighbourhood rather than an honest zero. The difference from what I measured above is real but narrow: mine drifts by meaning, theirs drifts by spelling. That is less bad, not good, and I would rather write it here than have somebody find it.

What mine does do, and the reason I ran the German test at all, is answer a query in one language from a catalogue written in another. That is a different capability from supporting many languages, and it is worth separating the two. A search can index English, French and German perfectly well and still refuse to let a French query reach a German product. The second behaviour is the one shoppers actually hit, and it is the one I have customers relying on today.

Three queries, five minutes, no tooling.

1. "X without Y", then "X" on its own. Compare the counts and check whether the excluded term climbed.
2. A real word you do not stock. Not gibberish. If you get results, break the query into single words and see which one carried them.
3. A concept you definitely cover, asked in a language you do not publish in. Confirm the English version works first, in the same session.

The third one is the sharpest, and it is the one almost nobody runs.

Whatever you find, the useful next step is the same and it costs nothing: read your own failed searches. Every site search worth the name logs the queries that returned nothing. That list is the actual specification for what your search needs to do, and it is written by your customers rather than by a vendor.