Someone types „no VAT" into the search box on your website. The page that explains exactly that has been online for three years, it is called „small business VAT exemption", and the search box replies: no results. The visitor closes the tab and asks somewhere else. This piece is not about Google and not about whether your page appears in an AI answer; all of that is in Cited, not clicked. It is about the box in the top right corner of your own website, the internal search. In most places it is a character comparison, while the visitor is looking for a meaning. This page shows what the built-in search technically does, which queries fail on it, who you lose at zero results, what our own sample of ten websites produced, and what a search looks like that finds meaning instead of letters.
What technically happens when someone types into your search box?
On WordPress, the most widely used system for business websites, you can read it in the source code. The method WP_Query::parse_search(), in core since version 3.7, turns the query into a database statement. Every typed term gets a wildcard on the left and the right, the code literally reads $like = $n . $wpdb->esc_like( $term ) . $n;, and is then checked with LIKE against three columns: the post's title, excerpt and content. Nothing else happens. No word stem, no synonym, no typo tolerance, no weighting by meaning. „VAT" does not find „value added tax", because the string is a different one. „invoice" finds „invoices" and „invoicing", because the short form sits inside the long ones; the other way round, „invoices" does not find the page that only ever mentions an „invoice".
Shopify shows what the counter-design looks like within the same class of tools. Its documentation on search behaviour describes typo tolerance: results may differ from the search term „by one letter" or have „two letters in a different order", but only if „the first four letters of a search term" are typed correctly. Common words count as stopwords and are not indexed at all. Synonyms are not known to this search either unless you create them as groups by hand. That is better than a pure character comparison, but it is still a search for spellings. The difference between „VAT" and „sales tax" is, for both systems, a difference between two words, not a closeness between two meanings.
Which queries fail, and which get through?
The Baymard Institute has spent years testing how the search of large shops handles different query types. According to its benchmark of search query types, last updated on 29 April 2026 and based on more than 170 benchmarked shops and apps, 12 per cent of sites have issues with exact searches, meaning the precise product name. For a product type it is 20 per cent. For queries by feature 39 per cent, by use case 43 per cent, by compatibility 44 per cent, for abbreviations and symbols 54 per cent, and for queries that are not about a product at all, shipping or returns for instance, 66 per cent. Overall, says Baymard, „56% of sites fail to adequately support users' search needs".
The order is the story. As long as the visitor knows the wording that is on the page, search works almost everywhere. The moment they describe what they need rather than naming what it is called, the failure rate triples. That is the gap between string and meaning, measured on real shops. The numbers come from retail, but the mechanism behind them is the same on every website: a law firm has its „small business VAT exemption", a trade business its „heat pump", and visitors ask for „no VAT" and „heating without gas".
How people search while doing this was measured by Jakob Nielsen back in 2001, and the basics have changed little. In his analysis „Search: Visible and Simple" the mean query was 2.0 words long. The first search succeeded in 51 per cent of cases, the second in 32 per cent, the third in 18 per cent, and „almost half the users whose first search failed gave up immediately". Two words, one attempt: that is the budget a visitor gives your search. If those two words are not on the page letter for letter, the attempt is spent.
Who do you lose when the search says „no results"?
The most valuable visitor the site has. The search provider Constructor, in an analysis from May 2026, says it evaluated billions of shop interactions: visitors who use search „convert at 2.5x the rate of non-searchers", they make up 25 per cent of traffic and bring in 57 per cent of revenue. That is vendor data, a search provider measuring the value of its own category. But the direction matches what anyone sees in their own numbers: whoever searches knows what they want, and is therefore closer to a decision than anyone who merely scrolls.
What happens when that visitor finds nothing was surveyed for Google Cloud by Harris Poll in 2023: a survey of nearly 13,500 adults in 14 countries who had used the search on a retail website in the previous six months. 80 per cent globally (81 per cent in the US) say that after an unsuccessful search they are more likely to buy elsewhere. 77 per cent (US: 82 per cent) avoid websites where they have experienced search difficulties before. 53 per cent abandon their basket if there is even one item they cannot find. Google puts the global damage at more than 2 trillion dollars a year. This survey too comes from retail and was commissioned by a company that sells search technology; for a law firm or a trade business it does not apply one to one. The second number is still the one that stays: whoever has searched in vain once does not come back soon.
What does our own sample of ten websites show?
We measured it ourselves on 11 September 2026, with the example from the first paragraph. On 19 public websites from Germany and Austria that inform the self-employed about tax and bookkeeping, firms, software providers and founder portals, we typed the same question into the built-in search in three phrasings: the technical term „Kleinunternehmerregelung", the everyday phrasing „keine Mehrwertsteuer" (no VAT) and the transposition typo „Kleinunternhemerregelung". A hit counted when the result list contained a page with „kleinunternehmer" in its address that does not appear for a nonsense term. Nine sites dropped out because their search does not respond server-side or because even the technical term found no page. It is our own sample, not a representative measurement, and we do not name any site individually.
On the ten remaining sites the technical term found the page in ten out of ten cases. The everyday phrasing found it in four out of ten. The transposition typo in one out of ten, and that one site visibly runs a search with typo tolerance. So nine out of ten searches were finished by a single swapped pair of letters, and six out of ten failed because the visitor says „Mehrwertsteuer" where the page writes „Umsatzsteuer", two words for the same tax. The answer was on the site in all ten cases.
How does a search find meaning instead of letters?
With the same mechanism that sits behind AI answers, only without the answer. An embedding translates a text into a long sequence of numbers that describes not its letters but its meaning. „No VAT on the invoice" and „small business VAT exemption" produce two sequences that sit close together although not one word matches. A vector database is the store that, given a query, finds the nearest neighbours among thousands of such sequences within milliseconds. Together they are semantic search: the page is found because it means the same thing, not because it writes the same thing. For a website with a few hundred pages that is no major project; the sequences fit into the same database the pages already live in.
What matters is what is not needed. Retrieval-augmented generation takes the same passages and has a language model write an answer from them; that is the AI chatbot. For the search box you do not need that second step. A result list sorted by meaning solves the problem from the first paragraph completely, without anyone having to chat with your website, without hallucination and without the question of who answers for a wrong answer. The jump from string to meaning is the big one. The jump from result list to chat is small, and it is optional.
And for everyone who relied on Google for the search on their own website: according to its announcement of 21 October 2024, Google switched off the sitelinks search box, the search field under some results, globally from 21 November 2024, because „usage has dropped". The associated markup may stay on the page, it simply no longer does anything. What happens in your search box is now entirely your own affair.
What is in your search log that nobody reads?
Before you change anything about the technology, read what people type. Almost every system logs search queries, WordPress needs a small plugin for it, Shopify shows them in its reports. The interesting ones are the queries with no results. Each of them is a question a visitor asked in their own words, and the collection of those questions is the most honest list of search intents you can get, because it comes from people who were already on your site. Most of the time there is no missing content behind it, just a different word: the page exists, it is simply called something other than what your customers call it. The same list is the test for any new search. Take the twenty most frequent zero-result queries and check what the new search does with them before you install it.
Three levers for this week
One: search the way a customer searches. Take your most important topic and type it into your own search box in three phrasings: the technical term, the everyday word, a typo. If two out of three fail, you know what your visitors have been seeing for years.
Two: get the log of zero-result queries. The last three months will do. Sort by frequency and write next to each query the page it should have found. That table is the synonym list your search has needed for years, and at the same time the best topic list for your next piece of content.
Three: decide between a synonym list and meaning. For ten known word pairs a synonym list is enough, and every better search supports one. For everything visitors describe rather than name, you need a search that compares meaning. The effort for that is smaller today than the effort for a new contact form.
If you want to know what your search does today with the queries your visitors type: we will test it with you, using your log. 🔍