Keyword search finds results by matching the words you type against the words stored in an index. Most engines first normalize those words, and many add stemming, synonym lists, and typo tolerance, but every match still depends on the words themselves.
A page that describes the same thing in other words can be missed unless the engine has been told the words are equivalent. Google and AI answer engines use words too, but they also match meaning. We ran one query through a hardware store’s search box, Google, and AI Mode to show the difference.
What is keyword search?
Keyword search, also called lexical or full-text search, is a retrieval method that returns documents containing the words in a query, or forms of them the engine has been set up to treat as equal. The system looks up each term in a prebuilt index, finds the documents that contain it, and ranks those matches. Production engines such as Elasticsearch can be configured to reduce words to their root, apply synonym lists, and fuzzy-match misspellings; a basic implementation does none of this. Either way, keyword search answers one question: “which documents contain these words or their configured variants?” Whether a document means what the searcher meant lies outside that question.
Literal matching is precise but unforgiving. An e-discovery handbook (E-discovery: Creating and Managing an Enterprisewide Program, 2009) shows the basic case: a straight keyword search for e-discovery finds e-discovery and nothing else. It will not pick up electronic discovery or eDiscovery, and it will miss a typo like edisocvery.
Everyday examples of keyword search:
- Find on page (Ctrl+F or Cmd+F), the simplest keyword search there is
- Library and research databases, where a search like puppy barking training returns records containing those words
- Site search boxes on blogs, help centers, and online stores
- Email inboxes and the file search on your computer
- Recruiting software that searches résumés for skills named in a job ad
- Legal and compliance review, where teams search large document sets for specific terms
Each of these tools rewards the searcher who guesses the exact words a document uses. That dependency explains both the method’s precision and the failures covered below.
How keyword search works, step by step
A keyword search engine does most of its work before you type anything. It reads every document once, records which words appear in which documents, and stores that as an index. Each search then runs a fast lookup of your words in that index and ranks the documents that matched.
- 1Collect the documents. Pages, records, emails, or files that should be searchable.
- 2Split text into words (tokens). “Fixing a leaky faucet” becomes fixing, a, leaky, and faucet.
- 3Normalize the words. Lowercase everything and often reduce words to a base form, so fixing and fixed both index as fix.
- 4Build an inverted index. For every word, store the list of documents that contain it, like the index at the back of a book. The Stanford Introduction to Information Retrieval textbook calls this structure “essentially without rivals as the most efficient structure for supporting ad hoc text search.”
- 5Look up the query. Your search terms pass through the same normalization, then the engine pulls the matching document lists and combines them.
- 6Rank the matches. A common approach, tf-idf weighting, scores a document higher when a query word appears often in it but rarely across the whole collection. Rare words carry stronger evidence of relevance than common ones.
Operators: how you control a keyword search
Because keyword search is literal, it gives you precise controls. These are the most common:
| Operator | What it does | Example |
|---|---|---|
| AND | Every term must appear | email AND template |
| OR | Any of the terms can appear | faucet OR tap |
| NOT / minus | Excludes a term | jaguar speed -car |
| Quotes | Exact phrase, in order | "follow-up email" |
Wildcard * |
Matches word variations | negotiat* finds negotiate, negotiation |
| Field search | Limits matching to one field | title:pricing, or Google’s site: |
Quotes are what librarians call phrase searching: they stop the engine from splitting your words and matching them anywhere on the page. Library databases and e-discovery tools support most of these operators. Google’s own search operator guide supports quotes for exact matches, the minus sign to exclude words, and site:, filetype:, before:, and after: to narrow results.
For anyone publishing pages, the key lesson sits in steps 4 and 5: a keyword engine can return a page only for words that appear in its index entries for that page.
Where keyword search breaks: synonyms, typos, and ambiguity
Keyword search struggles when the searcher and the writer use different words for the same thing, or the same word for different things. Unless someone has added tap and faucet to a synonym list, a keyword engine has no way to know they name the same object, and it can’t tell jaguar the car from jaguar the animal.
| Problem | What happens | Example |
|---|---|---|
| Synonyms | Relevant pages are missed | In our test, a store search for dripping tap returned irrigation parts and no faucet repair kits |
| Typos and spelling variants | No match on a basic engine | On an engine without typo tolerance, pziza finds nothing unless a page contains that exact misspelling, which is how Google describes its own early search |
| Ambiguous words | Irrelevant pages are returned | jaguar speed mixes cars and animals unless you exclude one |
| Small words dropped | The meaning changes | Treating for someone as noise turns “can you get medicine for someone pharmacy” into a generic prescription search (Google’s example) |
| Over-broad matches | Too many false positives | In a harassment investigation, a wildcard search would also surface an email about a sex-education class that “gave students ample opportunity to harass the teacher” (e-discovery example) |
Stemming, synonym lists, wildcards, and fuzzy matching patch much of this, and well-tuned engines use all of them. Each patch has a cost: synonym lists cover only the pairs someone added, and wider matching brings in more irrelevant results. In practice, a keyword engine finds what its maintainers anticipated.
How Google matches meaning as well as words
Google started by matching words and now layers AI systems on top that match concepts and context. In Google’s own account, “in the early days of Search, before we had advanced AI, our systems simply looked for matching words.”
Google names these systems in its guide to Search ranking systems:
- RankBrain helps Google “return relevant content even if it doesn’t contain all the exact words used in a search.” In a Google blog example, Google reads the query “what’s the title of the consumer at the highest level of a food chain” as a search for an apex predator, a phrase the searcher never typed.
- Neural matching represents the concepts in queries and pages and matches them to each other. Google describes it as looking at “an entire query or page rather than just keywords.”
- BERT works out how combinations of words change meaning, so short words like for and to can change the result.
- Passage ranking identifies individual sections of a page to judge how relevant the page is to a search.
Google says its new and old systems work together, so word matching still runs alongside the AI systems. Exact words still matter, most obviously for names, product codes, and quoted phrases.
For AI Mode, Google describes a “query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.” One question becomes many searches, so a page can surface for a sub-question the user never typed.
We tested one query in a store search, Google, and AI Mode
We typed “dripping tap” into a hardware store’s site search, Google Search, and Google AI Mode on October 5, 2026. The store returned irrigation parts and keg coolers. Google’s AI Overview treated it as a repair job, but the links went to pages that say “tap”. AI Mode led with a written repair guide.
“Dripping tap” is the British way to describe what a US store files under faucet, so it gives a simple test of how each system handles a vocabulary mismatch. The three systems do different jobs, though: a store retrieves products, Google ranks web pages, and AI Mode writes an answer. This is a side-by-side look at what a searcher gets. It is not a controlled lexical-versus-semantic experiment, and it can’t reveal how each system works inside.
1. A store’s site search: word matches that missed the repair job
Ace Hardware’s site search returned 9 results for “dripping tap”. None of them fix a dripping tap: five were drip-irrigation parts, plus a hose tap timer, two beer keg coolers, and a chicken waterer. The results behaved like word matching: each product’s name or description involves drip or tap.
The same store returned 750 results for “leaky faucet”, led by a core puller and faucet cartridges. We didn’t check all 750, but the top results were faucet repair parts, so the products exist; the shopper’s words didn’t reach them. A shopper who says tap may never see those parts, which is a possible lost sale, though we can’t measure that from outside.
A second pair showed the same pattern: “adjustable spanner” returned 3 results (a spanner wrench and two pliers wrenches), while “adjustable wrench” returned 144. In both pairs, the British term returned a small fraction of the results the US term did.
2. Google Search: understood the job, returned pages in the searcher’s vocabulary
Google’s AI Overview read the query as a repair job and gave fix-it steps, even translating the vocabulary: it told us to finish by “reassembling the faucet.” The ambiguity showed too: the first link was a King Gizzard & the Lizard Wizard song called “The Dripping Tap”, which also took a knowledge panel and its Wikipedia page.
The repair links came almost entirely from British pages that use the word tap: B&Q, Wickes, Southern Water, Camden Council, AXA UK, and Reddit’s r/DIYUK.
As a control, we searched “leaky faucet” with the same settings. It returned an almost entirely different set of sites: The Home Depot, Delta Faucet, Kohler, Maui County, and Clackamas River Water, all American and all saying faucet.
Google may be matching the literal word, or reading “tap” as a sign of a British searcher; one test can’t tell which. Either way, pages written in the searcher’s own vocabulary were the ones that surfaced.
3. Google AI Mode: a written answer to the repair question
AI Mode led with a step-by-step repair guide that separates old-style taps (replace the rubber washer) from modern mixer taps (replace the ceramic cartridge), with a list of tools to have ready. It cited how-to videos alongside the answer and stayed on the repair job, leaving the song out. For a publisher, the searcher sees the written answer first, and the sources it credited were videos.
What the test shows
| Store site search | Google Search | Google AI Mode | |
|---|---|---|---|
| How it works | Behaved like word matching on its product index | Words plus concept matching, per Google’s documentation | Generated sub-queries, then one written answer, per Google |
| What it did with “dripping tap” | 9 results, none for fixing a leak | Repair steps in the AI Overview; links from UK pages that say “tap” | A full repair guide citing how-to videos |
Run once on October 5, 2026 from India, with Google’s country set to US (gl=us) while signed in. Results vary by location, account, and time, and three systems doing different jobs can’t isolate one cause.
On a basic keyword engine, such as many site search boxes, your exact words decide whether you are found. On Google, words still matter, as “tap” and “faucet” showed, but they form one signal among several: Google can return relevant pages that don’t contain every word in the query. There, the question is whether your page is the best match for what the searcher meant.
How keyword search differs from keywords, keyword research, and search queries
Keyword search is how a system retrieves documents; a keyword is the term a page targets; keyword research is how you choose which terms to target. Many top results for “what is keyword search” drift into keyword research, so it helps to keep four related terms apart:
| Term | What it is | Question it answers |
|---|---|---|
| Keyword search | A retrieval method that matches query words to indexed text | How does a system find documents containing my words? |
| Keyword | The word or phrase a page is built to be found for | What topic should this page target? |
| Keyword research | Finding and prioritizing the terms your audience searches | Which topics are worth creating pages for? |
| Search query | The exact words a person typed | What did this person ask? |
The distinction has a practical edge: keyword research picks the terms, and keyword search decides whether a system can connect a searcher’s query to the page that uses them.
What this means when you write a page
Write for meaning first, and make sure your audience’s exact words appear too. A page that ignores the searcher’s vocabulary loses on keyword-style matching and weakens a signal Google still uses; a page stuffed with one phrase reads badly and adds little for systems that already understand synonyms.
For any page you want found in Google and cited by AI tools:
- 1Use your audience’s exact term where it reads naturally: in the title, the first paragraph, and at least one heading. In our test, “tap” and “faucet” brought up different sets of sites for the same problem.
- 2Cover the language around the topic. Mention the synonyms and related terms a reader would expect (faucet and tap, keyword search and lexical search). This is the core of semantic SEO.
- 3Open every section with a direct answer. Google evaluates individual passages of a page (passage ranking), and AI answers often draw on specific passages. A section that answers its heading in the first two sentences is easier to lift on its own.
- 4Answer the follow-up questions. Query fan-out means an AI tool may search for the sub-questions behind your topic. A page that answers them can match more of those searches.
- 5Spell out names. Write “Google Search Console” each time, and avoid shorthand like “the tool”. A plain keyword engine can match only names that appear on the page.
- 6Repeat a phrase only where readers need it. Systems that match meaning gain little from repetition, and it makes the page worse for readers.
Together, these steps serve both kinds of matching: the exact words satisfy keyword-style lookup, and the surrounding coverage gives meaning-based systems more to work with.
If your own site search is a keyword engine
Many store, documentation, and help-center search boxes are plain keyword engines, so the synonym problem above may be happening on your own site. Test it: search your site for the words customers use (tap for faucet, cancel for unsubscribe) and note what returns nothing or, as in the hardware store above, returns the wrong products. Then add a synonym list in your search settings, or rename pages to match the words people type.
Our keyword-to-LLM-visibility workflow covers the steps from choosing a keyword to getting cited by AI tools.
J6 Venture helps brands rank on Google and get cited in AI answer engines. If you suspect your pages, or your own site search, miss the words your customers use, book a strategy session and we’ll check.