Semantic search finds things by what they mean. You describe what you’re after in your own words, and the right message comes back, even when the message is written in words of its own.
The name makes sense against its opposite. Ordinary search compares the letters you typed with the letters in the document. Semantic search stops doing that, and almost everything it’s good at, along with every way it lets you down, follows from that one change.
- The gap it closes was measured before the web existed. Ask two people to name the same familiar thing and they pick the same word less than a fifth of the time. In a mailbox, the two people are you and yourself six months ago.
- The comparison happens between positions rather than words. Text becomes coordinates in a space arranged so that similar meanings land near each other. Searching becomes a question of what sits nearest.
- That mechanism fails in its own way. A keyword search that comes up empty tells you so. A semantic search hands back the closest thing it holds, and the closest thing looks exactly like an answer.
- It’s weakest where ordinary search is strongest: an invoice number, a client reference, a phrase you can quote. Serious products run both kinds of matching and merge the results, and it’s worth knowing whether yours does.
- The same two words name three other things. What a web search engine does to your query, an older idea about hand-written structured data, and the product a shop buys for its own site.
What the word semantic is doing there
Semantic is a borrowing from linguistics, and it was fairly new there too. Etymonline dates the English adjective to 1894, from French sémantique, “applied by Michel Bréal (1883) to the psychology of language”, from Greek sēmantikos, “significant”, and behind that sēma, “sign, mark, token”.
The contrast the word carries in its home discipline is the one that matters here. Semantics is about what an expression means, as against the shape it happens to take. The marks on the page are one thing. What they signify is another. Applied to search, that splits the job cleanly in two. A document’s form is the string of characters it’s made of. Its meaning is what it’s about. Every search engine ever built matches on the form and hopes it stands in for the meaning.
Mostly it does. If you and the sender both wrote “invoice”, the letters and the meaning point the same way, and the difference stays invisible. The gap opens when you reach for a message in your words and the sender wrote it in theirs. In a mailbox, that’s most of the time.
Wikipedia’s article on the term (checked September 5, 2026) puts the distinction in the form you’ll meet it in product copy. Semantic search is “search with meaning, as distinguished from lexical search where the search engine looks for literal matches of the query words or variants of them, without understanding the overall meaning of the query”. Lexical is the useful word in that sentence, because it names the other half properly. Lexical search works at the level of the word, and at that level it’s exact.
The phrase commits to one thing. It tells you how a match is decided. What gets searched, how far back, how much of it gets read, whether you’re handed a list or a sentence, whether the thing is any good: those are all separate questions. They’re where two products carrying the same label come apart.
The gap it exists to close
The problem has a name in the research literature and a number attached to it. The number is older than most of the software arguing about it.
In 1987, four researchers at Bell Communications Research published “The vocabulary problem in human-system communication” in Communications of the ACM. They asked people to supply names for the same objects across five different domains. The result is reported plainly: “In every case two people favored the same term with probability <0.20.” Building a system around one designer’s preferred word, they concluded, produces “80-90 percent failure rates in many common situations”.
That study was about command names and index terms, and it holds anywhere people name things. Two people rarely reach for the same word. In an inbox, the second person is you, six months back. You remember a conversation about the leak in the flat roof. The surveyor wrote “schedule of dilapidations”. You remember the client who was upset about being charged twice. The mail says “duplicate line item on 4419”. A lexical search compares letters, and your phrase and the surveyor’s are made of different ones.
Two things make this worse in mail than almost anywhere else.
Your memory of a message is a gist. You remember what a message was about, roughly when it arrived and how it made you feel. A keyword search asks you for an exact sender address, a subject line, a date. Those are the parts that fade first.
The failure is silent, and it lies to you. You searched, the screen came back blank, and the natural conclusion is that the message is gone. So you stop looking and write to the client to ask again. An empty result looks the same whether the mail is missing or your word was wrong. That’s the specific harm the vocabulary problem does in an inbox, and it’s why a search that fails softly, by handing back near-misses, has something to recommend it.
There’s a second route around the vocabulary gap, and it’s the older one. Most mail clients give you a language of operators for sender, date, label and attachment. Gmail’s own help page for searching opens by describing them as “words or symbols called search operators to filter your Gmail search results” (checked September 5, 2026). That’s a genuinely powerful instrument, and it solves a different problem. Operators need you to know a fact about the message, and to have learned the syntax. Semantic search asks you for an impression instead, which is what you’re more likely to have.
How a machine compares two meanings
One picture is enough here, and you can skip the engineering. The picture explains both what the feature is unexpectedly good at and the way it fails.
Every piece of text is turned into a long list of numbers. That list is a position in a space with a great many directions. The positions are arranged so that texts about similar things end up near each other, whatever words they used. “Schedule of dilapidations” and “the roof needs fixing” sit close together, because the model that placed them learned from a great deal of writing where such phrases turn up. Those lists of numbers are called embeddings. The space is what people mean when they say vectors.
Searching is then the same operation in reverse. Your question is turned into a position too, and the system returns whatever sits nearest. That’s the whole idea, and four consequences fall straight out of it.
Matching becomes a matter of degree. Every message in your archive sits at some distance from your question, so there’s always a nearest one. A semantic search always comes back full, and it comes back ranked. The ranking is happy to put a stranger at the top. Lexical search fails loudly and this one fails politely, which puts the checking back on you.
The comparison is over a whole passage at once. That’s why it copes with paraphrase, and why small words that reverse a meaning slip past it. “We can do Thursday” and “we cannot do Thursday” occupy nearly the same position, because almost everything about them is the same. Negations, exclusions and “anyone except” are where these systems are shakiest, and people ask for them often.
The index has to be built in advance. Something reads every message you own and stores the derived form. That’s where the cost lives. It’s also where the questions come from: privacy, how far back the index reaches, how quickly this morning’s mail becomes findable. All of that sits behind the interface.
Exactness is the first thing the mechanism throws away. A number, a reference and a quoted clause are meaningful to you precisely because they’re exact. The space stores meanings, and “this exact string and nothing else” is a request it has no position for.
Where it is beaten by ordinary search
Worth being direct about this, because the marketing rarely is. Some searches belong to ordinary keyword matching, and they’re common ones.
Anything you can quote exactly. An invoice number, a case reference, a booking code, an address, a clause you can recall word for word. If you have the string, matching the string is faster and exact, and it hands you the message itself. Gmail’s search over a large archive is superb at this. That’s the strongest thing about Gmail, and one of the reasons a great many firms should stay where they are. The full comparison with Gmail sets out what is and is not worth moving for.
Anything that is really a filter. “Everything from Anna in March with an attachment” is three conditions. A system that scores them by similarity approximates an answer it could have gotten exactly right. Good products route this sort of request into ordinary filters, even when you typed it as a sentence.
Anything that has to be complete. Semantic search ranks. Ask for every message that mentions a matter, and what you get back is a list of the best candidates, cut off wherever the product cuts it off. For a question you’re curious about, that’s fine. For an insurance disclosure, a records request, or working out whether you ever agreed to something, ranking is the wrong shape of instrument, and better quality leaves it the wrong shape.
The answer the field settled on is to run both. A hybrid search does lexical and semantic matching together and merges the two rankings, so exact terms stay exact and paraphrases still land. Wikipedia’s account notes those “hybrid search models combining lexical retrieval with semantic ranking” as standard practice (checked September 5, 2026). If a product describes its search only as AI-powered, it’s fair to ask what happens when you paste in a reference number.
Questions that separate two products selling it
Every product in this category says it understands what you mean. Five things actually differ, and you’ll usually have to ask about all five.
What is in the index. Mail only, or sent mail too, attachments, calendar entries, files, the chat that answered an email? A search that reads inside a PDF can find the figure in the PDF, and the figure in the PDF is often the thing you wanted. This matters more than any amount of matching quality.
How far back it reaches. Depth is frequently a purchase decision rather than a technical one. Shortwave’s pricing page lists “5 years AI search history” on its Business tier and “Unlimited AI search history” above it (checked September 5, 2026). An archive that goes back eleven years and a search that goes back five are a bad pair, and you find out at the worst moment.
How much it reads to answer one question. This is the question almost everybody skips, and it decides more than it sounds like. The same Shortwave page states the limit outright: “Max 50 threads per AI search” on Business, 100 on the tier above, 150 above that (checked September 5, 2026). Every product of this kind has such a budget, published or otherwise, because reading is what costs. In practice, an answer is drawn from a shortlist, and something assembled that shortlist before any thinking happened. Publishing the number, as Shortwave does, is more honest than most.
Whether it returns messages or an answer. These are different products. Retrieval hands you a ranked list and leaves the reading to you. An assistant reads the list and writes you a sentence. The second is more pleasant, and it inherits every weakness of the first. It summarizes what was retrieved, so a shortlist that skipped the relevant thread produces a confident answer about the wrong mail. Seeing the messages behind the sentence in one move is what keeps it safe, and whether you can trust AI judgments about your inbox is the fuller version of where that trust should stop.
Where the index lives and what else it is used for. Semantic search reads your mail and keeps a derived copy of it. For anyone holding client confidences, that has a concrete answer worth asking for, and it sits apart from the marketing: what’s stored, where, for how long, who can reach it, and whether any of it trains a model. What happens to your email and whether it trains anything is the version of that question worth putting to any vendor, and what an AI email client can actually see is the plainer walkthrough.
The same two words, three other jobs
If you arrived from a different corner of the industry, the phrase may be pointing at something else.
What a web search engine does to your query. In search marketing, semantic search describes the engines themselves getting better at meaning, and it’s a thing you optimize for rather than a feature you switch on. Google’s own announcement of BERT, published October 25, 2019, by Pandu Nayak, describes the shift. The models “consider the full context of a word by looking at the words that come before and after it”. That helps “particularly for longer, more conversational queries, or searches where prepositions like ‘for’ and ‘to’ matter a lot to the meaning”. If somebody in your business talks about writing for semantic search, this is the sense they mean, and it’s about how pages rank on the web.
The structured-data sense, which runs the other way. Before embeddings, semantic search usually meant retrieving knowledge from richly structured sources such as ontologies and RDF, the machinery of the semantic web. There, meaning was written down deliberately, by people, in advance. Today’s version infers meaning statistically, and what it stores is numbers. Both get called semantic search, and papers from around 2008 almost always mean the first.
Site search, for shops and support sites. A visitor types “something warm for a toddler in November” and the catalog is expected to cope. It’s the same mechanism against a small, shared, deliberately described body of content. That’s a much easier problem than a private archive of eleven years of correspondence, written by hundreds of people who were thinking about something else entirely.
Vector search, in a database. The plumbing underneath the first sense, sold to engineers by the people who build storage. If the page you’re reading talks about indexes and dimensions rather than about finding your mail, you’ve landed one layer down.
The names it is sold under
The vocabulary around this is wide, and each term means its own thing.
Vector search, embedding search, dense retrieval and neural search all name the mechanism rather than the promise. They name what’s happening. Semantic search names what it’s for.
AI search and smart search stretch to cover hybrid systems, pure keyword systems with a language model in front, and assistants that read your mail and answer in prose. The label fits all three, so ask which one you’re getting.
Natural-language search is about the shape of your question, and people read it as a claim about matching. A product can accept a plain sentence and then quietly convert it into ordinary filters on sender and date. That’s a good design, and it’s a different mechanism. For “everything from Anna in March” it beats semantic search outright.
Fuzzy search and typo tolerance work at the level of the letters, forgiving a misspelling. Useful, and a separate thing.
Ask your inbox, chat with your mail and retrieval-augmented generation all describe search with an answer written on top. The retrieval underneath is usually semantic, and the sentence you read is one step further from the evidence.
Two questions sort the whole list. What does the system compare, the characters or the meaning? What does it hand back, the messages or a sentence about them? Those two answers describe the product. Every other word on the page is a name for one of them.
How Point uses the word
In Point the search is the plain-language kind. You look something up by what you remember about it rather than by the words that were used. A half-recalled description of a conversation brings back the thread it belongs to, and mail written in the sender’s language is reachable in yours.
The same box takes a request rather than a query. Ask for John’s last email and you get it. Ask for the reply drafted or the meeting moved and that happens instead, at whatever level you have set for that kind of work. Point reaches the files as well as the messages, so a question about what a document says is answered from the document rather than from the note it arrived attached to.
The archive Point searches is the one you already have. Point runs on top of Gmail or Microsoft 365. Your mail stays where it is, your address stays yours, and your existing mail is the corpus from the first day rather than whatever accumulates after you switch.
The honest limit is the one this whole page has been circling. Search answers a question you thought to ask, so it reaches the mail you already know about. Most of what Point does happens earlier than that. Point weighs mail as it lands, summarizes it before you arrive, and lifts the asks inside it out as dated tasks. So the message you’d have walked past is in front of you anyway. How email triage works is the account of that pass, and it’s the half of the product that matters most on the days you have no time to go looking.
How far Point acts on what it finds is a dial you set, and each kind of task carries its own. The settings run suggest only, hold it back for approval, then carry it out. Every one of them starts at approval. There’s a plain record of what has been done, and undo covers Point’s own actions. The dial itself is set out in how much your inbox does without asking. The rest of what Point takes off your desk is listed under benefits.
Common questions
What is semantic search in simple terms?
It’s search that matches on meaning instead of on the exact words you typed. You describe what you’re after in your own language, and the system returns things that are about that, even when your words and the document’s words are completely different. The everyday test is whether you can find a message by describing it rather than by quoting it.
What is the difference between semantic search and keyword search?
Keyword search compares the characters you typed with the characters in the document. It’s exact and fast, and it stops the moment you use a different word from the one that was written. Semantic search compares meanings, so paraphrase works and exactness goes. They fail in opposite directions. Keyword search comes back empty when you’re wrong. Semantic search returns the nearest thing it has, which can look like a correct answer. Most good systems run both and merge the results.
Is semantic search the same as vector search or AI search?
Vector search is the mechanism and semantic search is the purpose, so a product doing one is usually doing the other. The two words still point at different things. AI search stretches to cover hybrid systems, keyword systems with a language model in front, and assistants that read your mail and reply in prose. Natural-language search is different again, and it describes the form your question takes rather than how matching is decided.
Can semantic search find an invoice number or a client reference?
On its own it’s shaky, and this is its clearest weakness. A reference means something to you because it’s exact, and exactness is the property the mechanism discards when it converts text into positions. A system that also runs ordinary keyword matching will handle it, which is why hybrid search became the standard approach. If a product describes its search only in terms of understanding your meaning, ask what happens when you paste in a reference number.
Does semantic search mean my email has been read and stored somewhere?
Yes, and that’s true of any implementation. Something has to read every message and keep a derived form of it, or there’s nothing to search. So the question is what: what’s stored, where it sits, how long it’s kept, who can reach it, and whether it’s used to train anything. Those all have concrete answers, and a vendor who can’t give them plainly has told you something.
What is semantic search in SEO?
A different sense of the phrase. There it describes web search engines understanding queries in context rather than matching keywords. That changed how pages are written and ranked, and it’s a behavior to write for rather than a feature anyone installs. The mail-client sense on this page is a capability inside a product. The two share a name and a family resemblance, and that’s the whole of the overlap.
Why does semantic search sometimes return confident nonsense?
Because it ranks by distance and something is always nearest. Every message in an archive has some measurable closeness to your question, so a result list always comes back full. When the thing you wanted is absent, the list fills with the least distant alternatives. The output reads the same either way, whether it’s a close match or the best of a bad set. That’s why being able to open the underlying messages quickly is worth more than any claim about accuracy.
The short version
- Semantic search matches on meaning rather than on the words you typed, so you can find a message by describing it. Its opposite is lexical search, which compares the characters and stops there.
- It exists because two people rarely pick the same word for the same thing, less than a fifth of the time in the classic study, and in an inbox the second person is you six months ago.
- Text is turned into positions in a space where similar meanings sit close together, so matching is a matter of degree. There’s always a nearest result, which is why this fails by returning plausible near-misses.
- It’s at its worst on exact strings, on requests that are really filters, and on anything that has to be complete. Serious products run keyword and semantic matching together, and it’s fair to ask whether yours does.
- Five things separate two products selling it. What got indexed, how far back it reaches, how much mail one question is allowed to read, whether you’re handed messages or a written answer, and what happens to the copy of your mail that had to be made.
- The same phrase names what a web search engine does to your query, an older idea about hand-written structured data, and the search box on a shop’s website. Only one of them is a feature you switch on.