Search is shifting from blue links to embedded infrastructure, with startups and giants racing toward agentic systems where retrieval organizes the work itself.
If you haven’t noticed, we are in the awkward middle of a search reset, where the old ritual of typing keywords into a box and sifting linkbait has given way to an interface that tries to tell you the answer first.
The front of the web is becoming a chat pane, and the back is a stack that blends crawling, indexing, retrieval, and synthesis with a growing bias toward doing the work on your behalf.
The next step is already visible in two places. First, the browser is turning into an agent that reads, compares, and acts for you. And second, the incumbents are doing way more than just ranking pages. In fact, we can look at what they’re doing now as more negotiating what gets summarized, what gets paid, and which models get to see which corpora at what latency.
The change will be remembered as far more than a UI tweak because what’s really happening is more a shift in power among index owners, model vendors, publishers, and the handful of companies that can make agents trustworthy at scale.
On the surface, the most obvious change is the rise of general “answer engines” that respond with sourced syntheses rather than a list of blue links.
Perplexity leaned into this earliest and built a reputation on fast, cited answers, then moved up the stack with Comet, an AI browser that runs whole browsing sessions for you and is now paired with a publisher revenue program that pays when content is visited, cited, or used.
Brave came from the opposite direction, starting with its own index and adding synthesized “Summarizer” and “Answer with AI” layers to the top of results. You.com merged search with Private RAG so an enterprise can bring its own corpus to the party. And Kagi built a paid, privacy-first model with a Universal Summarizer and an API.
And if you’re not overwhelmed yet I'd be remiss to not mention Andi and Komo, which have both been pushing the clean, ad-free chat aesthetic where you ask a question and get back an answer with sources. Taken together, this cohort is training users to expect citations by default, light follow-ups, and a working memory of the thread, the latter being the most crucial to many.
The developer corner tells a more specific story because the trend toward specificity in search is emerging.
For this crew, Phind built and then iterated its own tuned models for coding, positioning itself as a dev-first answer engine that searches, reasons, and writes runnable snippets with an eye toward speed and quality. Sourcegraph took a decade of code search and wrapped Cody and Deep Search around it, bringing agentic workflows into the IDE and across massive repos. And trusty old Stack Overflow, pressured by usage shifts, has pulled its institutional knowledge forward with OverflowAI, which synthesizes from Teams and public threads to give enterprises grounded answers and a way to keep that knowledge alive.
The pattern is consistent: constrain the corpus, preserve provenance, and wire the assistant to your tools so answers become diffs, PRs, and guardrailed changes.
The trend toward specificity in search is also cropping up in research and academia.
Consensus narrows the universe to peer-reviewed work and answers with paper-level grounding. Scite invented Smart Citations to show how a paper cited and whether the citation supports or disputes the claim. SciSpace stitched together search across roughly 280 million papers and wrapped it in an agent that can run a systematic review, summarize methods, and push you through the writing stack without losing the thread of what the sources actually say. In essence, these tools work because they treat “search” as literature triage under time pressure, not as a journey through ten blue links.
Inside companies, work search is converging with assistants and agents.
Glean calls the category Work AI and pairs permission-aware workplace search with an assistant and an agent framework so you can find, reason, and then act against the same index. Hebbia approaches the problem from finance and other research-heavy teams, turning dense filings, data rooms, and vendor content into agentic workflows that read, extract, cross-check, and draft. Platforms like Vectara sell the plumbing, combining semantic and keyword retrieval with reranking so builders can ship grounded generative answers. Algolia pushed hybrid search into retail-scale catalogs with NeuralSearch that fuses vector and keyword results on every keystroke.
The common denominator is hybrid retrieval, a permissions model, and answers that show their homework.
And speaking of specificity in this search transition, retail is where this flips from Q&A to action.
Amazon’s Rufus lives inside the store and explains, compares, and narrows choices, then Lens Live lets you point your phone at an object and shop visually, with real-time matches that drop directly into a cart or a list. Shopify has been pushing semantic search across its Search & Discovery app so smaller merchants inherit intent-aware results without building their own retrieval stack.
Shopping queries are becoming conversational and multimodal by default, which is how you compress the last mile between curiosity and checkout.
But back to the broader view.
The browser is literally the most important battleground because it is where agentic search feels natural.
Arc Search’s “Browse for me” reads multiple sources and builds an instant, sourced page that answers your question. As mentioned before, Perplexity’s Comet goes further, turning the browser into a personal assistant that can navigate web apps, compare pages you have read, and even manage your inbox and calendar as part of a research thread. And back to Brave, remember it too is hardening an agent that can browse and complete transactions, which telegraphs a future where you do not ask for links, you ask for outcomes and review the receipts.
This is search as orchestration with a UI that looks a lot like your current browser, except it works while you think and it’s going to be difficult for the old blue link world to compete.
And by the way, beneath all of this sits the search infra tier, which explains why there are so many “new search engines”.
Pinecone, Weaviate, and Qdrant turned hybrid retrieval into a commodity. You can blend dense vectors for intent with sparse signals for precision, filter by metadata, and rerank for relevance, all from a few API calls. That makes it trivial to slap an answer layer on any domain corpus and call it search. The hard part has moved to curation, grounding, and policy, not distance metrics.
And by the way, in case you’re wondering (and probably expecting), VAST is not a search startup in the same sense but it does belong in this layer because it collapses storage, indexing, and retrieval and this is a big deal for how all of this gets put together. The VAST DataBase and DataEngine make hybrid search native so think dense vectors, sparse terms, and metadata filters all running against the same exabyte-scale corpus without exporting data to an external service.
In other words, where Pinecone, Weaviate, and Qdrant package vector search as a service, VAST bakes the same retrieval logic into the fabric of its AI Operating System, making search a native function of the data itself.
Before we go let’s get back to the changing face of search and spend just a minute on vertical search because this is where AI starts to feel indispensable fastest.
Companies like AlphaSense stand out here. They’re sitting on a premium universe of filings, transcripts, and analyst notes, and now ships Smart Summaries and a Deep Research agent that derive “why it matters” across thousands of documents with citations back to exact snippets. vLex’s Vincent AI reads across jurisdictions and drafts with linked authorities.
In medicine, OpenEvidence is becoming a point-of-care habit, answering clinical questions against peer-reviewed literature with referenced guidance, and it now claims daily use by a very large share of U.S. physicians. This is search as decision support in high-stakes domains, and the bar for provenance is set accordingly.
And while it seems wrong to not mention the incumbents until now, it’s not like they’re standing still. Google put AI Overviews at the front of Search, then introduced AI Mode, a deeper chatbot-style experience that encourages follow-ups and more complex reasoning inside the results page. The company has since expanded Overviews well beyond the U.S., and is actively iterating the agentic layer. This is both defense and offense, because it keeps the query inside Google’s UI and gives the model license to synthesize the open web before you click.
Microsoft reframed Bing as Copilot Search, which returns summarized answers with citations and suggests where to go next. The move mirrors Copilot’s broader push inside Microsoft 365, where search collapses into an assistant with your permissions and context. The lesson is that “search” becomes far more valuable when it already knows your files, meetings, and teams, and when it can produce a working artifact rather than a list of links.
OpenAI folded browsing into ChatGPT Search so that a general model can answer timely questions with sources, which has the side effect of turning ChatGPT into a competitor for both the engines and the publishers that rely on them for traffic. DuckDuckGo chose a privacy posture, standing up Duck.ai for anonymous conversations across multiple third-party models and extending AI-assisted answers in its classic search. xAI’s Grok advertises native tool use and real-time search integrated with X. Baidu is weaving ERNIE back through its product surface, including Baidu Search. In Korea, Naver retired its Cue: experiment and is shipping an AI Briefing layer directly in search. Each path pushes more of the work into a chat surface, with slightly different tradeoffs on privacy, index control, and what counts as a citation.
And let’s not forget about Reddit. They’re something of a special case because they own a corpus that answer engines crave. It launched Reddit Answers, an AI layer that synthesizes from posts and comments with links back to the conversations, and it is openly repositioning itself as a search destination in its own right.
And by the way, this is all happening against a legal and economic backdrop that will matter more than any single UI. A U.S. judge just issued limited remedies in the Google antitrust case that, among other things, require portions of Google’s search data to be shared with qualified competitors. That does not break the monopoly on distribution, but it kinda does nudge the market toward more credible entrants by reducing one of the hardest costs in search, which is bootstrapping an index that models can trust.
At the same time, Perplexity’s Comet Plus is a test balloon for a new détente with publishers, where access and summarization are paired with revenue and reporting. Expect a messy period of blocklists, robots.txt brinkmanship, and experiments in pay-per-crawl and pay-per-cite as the winners figure out how to fund the open web they summarize.
But no matter which of the incumbents or upstarts we talk about, where this goes next is agentic.
Answer engines will feel less like a chatty search box and more like a colleague that reads what you read, keeps your place, and acts on instructions that span days or weeks.
Browsers will host that colleague so you can delegate a path through pages, comparisons, and forms without manually stitching context.
Enterprise search will blur into workflow, where retrieval is only the first step before a model drafts a doc, files a ticket, or opens a pull request under guardrails.
And the giants will keep pulling the answer to the top of the page because every extra click is a chance to lose the session to a rival.
Still, the open questions are the hard ones, like how we attribute and pay for the sources, how we separate high-confidence synthesis from confident error, and how we write the rules so that “search” remains a window onto the web rather than a wall around it.
But be sure of one thing. The balance of power will follow whoever solves those three with the least friction for users and the most oxygen for the ecosystem that makes answers possible in the first place.