AI Search Engines 2026 Head-to-Head — Perplexity · You.com · Phind · Exa · SearchGPT · Gemini AI Mode · Kagi · Tavily, and the Deep Research Category

Prologue — the "ten blue links" era is ending
1. The new layer between query and answer
2. Consumer AI search apps — what humans type into
3. Developer search APIs — hands and feet for agents
4. The Deep Research category — when a five-minute answer becomes a thirty-minute report
5. The citation reliability problem — AI search's largest weakness
6. When AI search wins vs when it doesn't
7. The AI-native browser thesis — Comet, and beyond
8. Honest decision tree — what to use when
Epilogue — checklist, anti-patterns, what's next
References

Prologue — the "ten blue links" era is ending

Through 2024, the act of "searching" was unambiguous. You typed a few words into Google or Bing, you got a SERP with ten blue links, you clicked some of them, you read the pages, and you synthesized the answer yourself. The engine searched. The human composed.

By spring 2026 that split had collapsed. The engines themselves answer now. Perplexity was designed that way from day one; OpenAI's ChatGPT Search and SearchGPT followed; Google began integrating AI Mode into its own crown jewel at google.com. On top of all that, a new category emerged — Deep Research, the multi-minute autonomous browse-and-synthesize tools that fetch dozens of pages, cross-reference, and hand back a cited report.

Below that, a separate market quietly exploded — search infrastructure. Exa, Tavily, Serper, You.com's Search API. These are not consumer products. They sell search APIs for AI agents to call. RAG pipelines, agent workflows, the toolboxes underneath the LLMs we already use. Many of the queries you fire at Perplexity are themselves powered by infrastructure of this shape.

This piece looks at both markets together. Consumer AI search (humans typing) versus developer AI search APIs (agents calling) on one page. We pull Deep Research out as its own animal. And we ask the honest question — when does AI search actually beat classic search, and when does it not? Plus the thesis Perplexity's Comet browser keeps making — that search will eventually be the browser — and how plausible that really is.

Prices and features move fast. Every number in this post is as of May 2026 and the focus is on the decision frame. Six months from now the numbers will shift, but the axes — consumer vs infra, single-query vs Deep Research, citation reliability — will still apply.

1. The new layer between query and answer

Classic search in one line.

user query → search engine (index match + ranking) → SERP (ten links)
                                                          ↓
                                                human reads + synthesizes

AI search in one line.

user query → intent parsing (LLM) → multiple sub-queries generated
           → search index called (own index or external API)
           → result pages fetched + main content extracted
           → model synthesizes → cited answer

The crucial point is that a layer has been added. The user's query no longer goes straight to an index. An LLM first parses intent, decomposes the question into sub-queries, fetches results, reads the body text, and feeds all of that back into the model to synthesize an answer. Two things flipped.

First, the definition of a "search engine" blurred. Perplexity runs its own crawler but also leans on Bing and Google APIs. ChatGPT Search sits on the Bing index with OpenAI's curation layer over it. You.com has its own index but mixes in external sources. Owning an index and producing an answer are now decoupled.

Second, results are the answer. Users click out less. Perplexity's own data shows roughly 1.2 to 1.5 citation clicks per answer on average — meaning most users read and stop. For publishers, that is traffic vaporizing, and it is the heart of the 2025–2026 publisher-versus-Perplexity licensing fights.

Three architectures are worth naming.

Own index + own synthesis — Google AI Mode is the cleanest example. Google's index plus Gemini on top. Kagi's AI features are similar.
External index + own synthesis — Perplexity in some modes, ChatGPT Search (Bing-backed). They rent the index and own the synthesis and UX.
Own index + API exposure — Exa, Tavily, the API side of You.com. They do not answer for the user. They expose results (or extracted body text) for some other LLM to synthesize on.

These are not three players fighting for the same crown. They are three markets. Consumers see (1) and (2). Developers see (3). And several companies — You.com being the clearest — sit in both at once.

2. Consumer AI search apps — what humans type into

2.1 Perplexity — the company that defined the category

Perplexity launched late 2022 and basically named the AI-search category. As of May 2026 it has four major surfaces.

Pro Search. The default. The user's question goes to an LLM (the user picks among GPT-4, Claude, or Perplexity's own Sonar line), which fans out sub-queries, searches the web, and synthesizes a cited answer with inline references and a sidebar of follow-up questions. Free tier has a cap; Pro at about $20/month removes the cap and unlocks model choice.

Deep Research. Showed up late 2024 and became the headline differentiator through 2025–2026. Spends five to ten minutes on one question, fetching dozens of pages, cross-referencing, and producing a structured report. Where Pro Search reads five to ten sources in under a minute, Deep Research reads thirty to eighty sources in five to ten minutes. Pro plan has a daily cap (in the order of five to ten runs as of May 2026); the Max plan (around $200/month) raises it dramatically.

Spaces. A workspace concept added in 2024. You pin a topic — say "my PhD reading list" or "Korean instant noodle market" — accumulate context inside it, upload your own PDFs and notes, and let them flow into search. Collaborative. Roughly Notion-shaped on top of Perplexity.

Comet browser. Beta in 2025, generally available in early 2026. Not just a browser — it pitches itself as an AI-native browser. Every page has a sidebar assistant that holds the current page as context. Multi-tab summarization. Agentic mode that can drive pages on your behalf (for example, "compare the pricing on these companies and put it in a table"). The thesis is direct — search's future is not a box, it's the browser itself.

Perplexity's strengths are UX consistency and readable citation rendering — every claim is traceable. The weaknesses are freshness of its own index and fidelity of cited facts — in multi-hop reasoning, citations and underlying text often diverge. Going into 2026 the Sonar line got faster and cheaper, so Pro Search frequently feels snappier than ChatGPT Search.

2.2 You.com — the early mover that ate its own lunch

You.com was actually first to ship AI search. Richard Socher founded it in 2020 and shipped AI answers inside it by 2022. By 2026, though, You.com had clearly lost the consumer race to Perplexity.

The reason is plain. You.com tried to do too many things at once — search, AI chat, image generation, code agent, ads bolted on — and the UX fractured. None of the surfaces ever became number one in their lane. By late 2025 the company very visibly pivoted weight onto the API business — You.com Search API became the headline revenue product, and consumer you.com leaned more toward demo and marketing. Their enterprise B2B angle leans hard on the same API.

So there are two ways to score You.com.

As a consumer search engine? Perplexity or ChatGPT Search is better.
As a developer search API? You.com Search API is genuinely worth a look — pricing is reasonable and non-English (Korean, Japanese) results are surprisingly good.

The real story lives in the next chapter, in the infra market.

2.3 Phind — the one only developers use

Phind chose a totally different lane — developer search. It started as a Stack Overflow alternative and weights code blocks, library docs, and GitHub issues heavily. Answers are code-dense, and citations skew to official docs, GitHub, SO, MDN.

By 2026, Phind ran along two tracks.

Phind Search — the default, an AI search tuned for code. Free tier with daily caps, Phind Pro around $20/month.
Phind 70B / Phind Models — its own code-tuned model family, some of which were released as open weights and got integrated into other tools (Cursor and similar).

A specific quirk — Phind also ships a CLI. Running the phind command in your terminal does search and code generation right there. Some developers treat it as a faster-first-response than Stack Overflow.

The weaknesses are obvious. Outside coding Perplexity wins easily; the user base is narrow so data effects compound slowly; and coding itself is being eaten from inside the IDE — GitHub Copilot Chat, Cursor's inline AI, Claude Code. The "search then answer" workflow keeps getting subsumed into IDE tooling.

2.4 SearchGPT / ChatGPT Search — OpenAI's late entry

OpenAI announced SearchGPT as a prototype in July 2024 and folded full web search into ChatGPT in October 2024. As of May 2026, how it works:

Inside ChatGPT the user can hit a "Search" toggle, or the model auto-decides search is needed.
The index is Bing-backed plus OpenAI's own crawler topping up gaps.
Answers carry inline citations; a side panel lists sources.
Free users have search — that is the biggest distribution wedge.

OpenAI's Deep Research is a separate mode, launched early 2025, built on the o-series reasoning models with multi-hop browsing layered on. Averages ten to thirty minutes per question. Limited runs on Plus ($20), much higher caps on Pro ($200). OpenAI Deep Research is often deeper than Perplexity Deep Research but also slower. It shines on academic and heavy market-research jobs.

ChatGPT Search's biggest strength is that search is free inside a product that already has hundreds of millions of users. That is a brutal pressure on dedicated apps like Perplexity. For mainstream users the default seat for "ask AI and have it search" is already ChatGPT.

The weakness — the UX is chat-first, not search-first, so citation clickthrough is lower and source rendering is less crisp than Perplexity. Citation-vs-claim drift in multi-hop answers is similar or a hair worse.

2.5 Gemini AI Mode / Grounded Search — Google rewires its core

The most important shift gets the least airtime — Google itself. Google started showing "AI Overviews" above the SERP in 2024 (US). In 2025 it added an "AI Mode" tab. By early 2026 it was running broad experiments to make AI Mode the default search UX for many query classes.

How it works:

You type into google.com → if AI Mode is on, a Gemini-synthesized answer with inline citations appears at the top, with the conventional SERP underneath.
Click the "AI Mode" tab and you get something very close to Perplexity — chat with Gemini, citations, follow-ups.
For developers, the Gemini API's Grounding with Google Search automatically attaches cited search results to your LLM calls.

Google's overwhelming asset is the index. Nobody matches Google's freshness and coverage. The constraint is that Google's ad revenue lives on clicks, so pushing AI Mode aggressively cannibalizes its own P&L. Google goes deliberately slow. Through 2026 the publisher compensation and ad model experiments are still in flight.

For mainstream search users, Gemini AI Mode is honestly "already close enough" — quality is comparable to Perplexity, the index is fresher, and it is free. The reason to keep going to Perplexity is shrinking.

2.6 Bing / Copilot — Microsoft's two tracks

Bing got the AI search story started in early 2023 by bolting GPT-4 chat onto search. Once the hype settled, Microsoft cleanly split the work in two.

Bing Search itself — became wholesale infrastructure, powering ChatGPT Search and parts of Perplexity. Effectively an index for hire.
Copilot (formerly Bing Chat) — evolved into a Microsoft 365 assistant. Baked into Windows, in the Edge sidebar, threaded through Office apps. It does more than search — documents, code, email.

As a standalone consumer AI search tool, Copilot is not the leader. But inside the Microsoft ecosystem it is the default, and the user count is not small.

2.7 Kagi — paid privacy-first search with optional AI

Kagi is a different kind of company. Paid search engine, around $10/month. No ads. No user data collection. You can block or boost domains yourself. Loyal, small user base.

Its AI surfaces:

Quick Answer — a short AI summary above the result list, with citations.
The Assistant — separate chat UI; the user picks the model (Claude, GPT, Gemini, others).
Universal Summarizer — a side tool for summarizing URLs or YouTube videos.

The value prop is sharp — search without ads, without tracking, with AI as an honest option. You can turn AI off and just search. The opposite philosophy of Perplexity or Gemini AI Mode, both of which force an answer by default.

Weaknesses: the price tag ($10/month is friction in a market trained on free search), and index coverage that lags Google/Bing in some niches (it blends its own crawler with external indexes).

For serious knowledge workers, academics, journalists — Kagi is a genuinely good pick. Not a mass-market tool.

3. Developer search APIs — hands and feet for agents

In the same window, a market most users never see has exploded — search APIs for agents to call directly. Infrastructure layer.

3.1 Exa — embeddings-first developer search

Exa (née Metaphor) was designed from the start assuming an LLM would be the caller. Its signature is semantic search, not keyword search. "Find me pages similar to this one" or "find blog posts that talk about this topic" works well. Keyword search is also supported.

Core surface:

/search — semantic or keyword. Returns URLs, titles, publication dates, snippets.
/contents — fetches clean body text for a page, stripped of ads and navigation. Drop-in ready for LLM context.
/findSimilar — give it one URL, get back similar pages. Not something traditional engines expose.
/answer — convenience endpoint that stacks the above into a short synthesized answer.

Exa's strength is LLM-friendly API design. The body extraction (/contents) is clean enough that RAG pipelines hand it to the LLM with almost no post-processing. findSimilar is genuinely useful for research workflows and has no real equivalent in classic engines.

Pricing is usage-based (a few dollars per thousand queries; content extraction extra). For consumer-scale products that gets expensive fast.

Startups building Perplexity-shaped answer engines call Exa more often than any other infra. Perplexity has its own index; smaller teams plug in Exa.

3.2 Tavily — the de facto agent search API

Tavily targeted one narrow market from day one — search APIs for LLM agents. It got integrated into LangChain and LlamaIndex early and became the default web search tool for agent workflows.

The API is simple — tavily.search(query, depth=...) parses intent, runs multiple queries, scrapes and cleans results. depth='basic' is fast and cheap; depth='advanced' is heavier and more thorough.

A wrinkle — answer synthesis is optional (include_answer=True). Tavily will synthesize a short answer for you. Quality is okay but generally weaker than letting your own LLM compose. Most teams use Tavily for retrieval only.

Tavily's pitch is narrow and obvious — "want to fill in LangChain's WebSearchTool quickly? Tavily." That is the whole pitch. Free tier exists; usage-based above it.

3.3 Serper / SerpAPI — Google results, plain

Serper, SerpAPI, ScaleSerp and friends sell exactly one thing — Google results as an API. No index of their own. They scrape (or pull through approved channels) and hand back structured JSON.

Why is there a market? Google's official Programmable Search API is expensive and capped, with limited result processing. So unofficial proxies became the standard.

Characteristics:

Cheapest pricing (a dollar or two per thousand queries).
Same results Google would show — index freshness and coverage matched.
No synthesis. Just a result list.

When Tavily or Exa start to feel expensive in a RAG pipeline, dropping down to Serper is a common cost lever. Body extraction has to be added separately (Reader API, Trafilatura, similar libraries).

3.4 You.com Search API — infrastructure with its own index

In the consumer chapter we said You.com stumbled. On the infra side they actually built a real seat. The You.com Search API:

Own web index plus external blends. Less dependent on Bing or Google than competitors.
Content extraction included, similar to Exa's /contents.
Pricing comparable to or slightly under Tavily and Exa.
Multilingual results (Korean, Japanese) are surprisingly strong.

They actively target enterprise (combining first-party data with web search) and have been adopted as the search backend of several large B2B SaaS products.

3.5 Brave Search API — another own-index option

Brave is best known as the browser, but it runs its own search index (Brave Search) and exposes it via the Brave Search API. Pricing is reasonable, and the data policy is clean — queries and results are not used for training.

Because it is a separate index it disagrees with Google on some queries, and in some domains the quality lags. But its privacy and licensing posture is unambiguous, and several AI companies have picked it up as a backend.

3.6 Consumer vs API matrix

Product	Consumer app	Developer API	Own index	Deep Research	Starting price	Headline differentiator
Perplexity	Strong (Pro Search, Spaces, Comet)	Weak (Sonar API)	Partial (own + Bing etc.)	Strong (Deep Research)	Free / Pro $20 / Max ~$200	UX, citations, Comet browser
You.com	Weak	Strong (Search API)	Strong (own)	Partial	Free / Pro ~$20 / API usage	Multilingual, enterprise
Phind	Medium (devs only)	Weak	Partial	None	Free / Pro $20	Code and docs focus
Exa	None	Strong	Strong	Partial (Research API)	Usage (a few dollars per 1k queries)	Semantic search, content extraction, `findSimilar`
OpenAI Search / Deep Research	Strong (ChatGPT Search)	Medium (web_search tool)	Partial (Bing-backed)	Strong (Deep Research)	ChatGPT Plus $20 / Pro $200	User base, model integration
Gemini AI Mode / Grounding	Strong (google.com AI Mode)	Strong (Grounding API)	Very strong (Google index)	Strong (Gemini Deep Research)	Free / Google One AI ~$20 / Vertex usage	Index freshness, free
Bing / Copilot	Medium (Copilot)	Medium (Bing API)	Very strong (Bing index)	Partial (Copilot Pages)	Free / Copilot Pro $20 / API usage	M365 integration
Kagi	Medium (Search + Assistant)	Weak (small API)	Partial (own+external)	Partial	From $10/month	No ads or tracking, user control
Tavily	None	Strong	None (external curation)	Partial (Research API)	Free tier / usage	LangChain and LlamaIndex default
Serper / SerpAPI	None	Strong (Google results)	None	None	Usage ($1–2/1k queries)	Cheapest, Google results as-is
Brave Search API	Weak (Brave Search)	Medium	Strong (own)	None	Free tier / usage	Own index, no training use

Pin this matrix to memory. We move on.

4. The Deep Research category — when a five-minute answer becomes a thirty-minute report

The single most interesting event of 2025 was the rise of Deep Research as a category. OpenAI, Perplexity, and Google all shipped a product with that name within months of each other. All three do the same shape of work — spend five to thirty minutes per question, autonomously browse dozens of pages, and return a cited report.

The mechanism is similar.

user query → model lays out a research plan (what subtopics to cover)
           → auto-generates multiple search queries
           → fetches result pages, accumulates body text into context
           → if gaps remain → more search → fill in
           → reconciles conflicting facts → cross-check
           → synthesizes a structured report with citations on every claim

The three products differ on style.

OpenAI Deep Research — the deepest. Often runs over thirty minutes. Built on o-series reasoning models with tool calls bolted on. Excels at academic research, market reports, due diligence. Downsides: slow, expensive (Pro plan or API usage), and occasionally so deep it overshoots the actual question.

Perplexity Deep Research — the fastest and most often-used. Usually five to ten minutes. Lighter output than OpenAI's, but enough depth for everyday information work. Pro users hit a daily cap; entry barrier is lower than the other two.

Gemini Deep Research — bundled into Google One AI. Inherits Google's index freshness, which matters. Synthesis quality has steadily climbed, and Gemini's 1M+ token context window holds more accumulated material in working memory at once. That qualitative difference helps when you are stitching long, scattered information together.

When is Deep Research worth it.

Worth it:

Market research ("major players and funding flows in 2026 SE Asia fintech")
Academic synthesis ("recent papers comparing Mamba and Transformer architectures")
Company DD ("XYZ Inc — product, team, funding, competition, risks")
Policy and legal tracking ("EU AI Act rulemaking changes in 2026")

Not worth it:

Simple fact checks ("React 19 release date") — a basic search answers in five seconds.
Coding debugging — reading and running the code yourself is faster.
Real-time info (stock prices, breaking news) — index freshness matters more.
Anything where the answer lives in a single authoritative page (one official doc and you are done).

Deep Research wins when the work is cross-source synthesis — when no one page has the answer. Anything else is overkill and a waste of clock.

One more thing — hallucinations are more dangerous in Deep Research output. In a short answer, when a citation does not match the text, the user can verify in a click. In a thirty-page report with thirty citations, nobody clicks all thirty. The longer the output, the higher the verification cost and the more likely a bad citation slips through. For serious deliverables every load-bearing claim must be human-verified against the source. Always.

5. The citation reliability problem — AI search's largest weakness

The headline promise of AI search is "citations attached, so trustworthy." Be honest — that promise is half-true.

Multiple independent evaluations across 2024–2025 converged on the same number. Sample random citations from AI search answers and 20 to 40 percent of them diverge from the underlying text. The divergence comes in three flavors.

The fact is not in the source. (Most dangerous.) The model picked up the fact elsewhere and pinned it to that citation. Click through and the claim is not on that page.
A similar fact is in the source but subtly different. Numbers off, conditions stripped, dates shifted. The most common pattern.
The fact is correctly in the source but pinned to the wrong sentence. The model swapped citation slots.

All three are synthesis-stage failures, not retrieval-stage. Search did its job and pulled relevant pages. But moving from "what the page says" to "what the answer says" is where the model slips.

Honest impressions across products (averaging across independent evaluations):

Gemini AI Mode (Google index-backed) — citation accuracy ranks highest on average, especially on quick answers (two to three pages).
Perplexity Pro Search — best citation rendering, short answers usually hold up. Multi-hop answers see more type 1 and 2 drift.
ChatGPT Search — comparable level, but the citation rendering is more buried, so users verify less.
Deep Research products — three to six bad citations among thirty-plus on average. Absolute accuracy per citation may be higher than a quick answer, but a long report almost always carries some bad cites somewhere.

Practical conclusion is clean.

Low-stakes decisions (where to eat, how useEffect changed in React 19) — trust AI search as-is.
Mid-stakes decisions (drafting a market-entry analysis, choosing a stack) — use AI search for the first pass, manually verify two or three load-bearing facts.
High-stakes decisions (legal, medical, financial, policy) — AI search is a starting point only. Every load-bearing fact verified against primary sources. If the report has thirty citations, you click thirty citations.

This is not "do not use AI search." It is do not assume the answer is 100% correct; grade your answers by stakes. The same grading was always needed for classic search — Wikipedia sentences also require verification. The difference is that AI search reads so fluently that users lose the verification reflex.

6. When AI search wins vs when it doesn't

A lot of writing announces the death of classic search prematurely. The honest 2026 read is that the two markets coexist. Default-tool changes per case.

Where AI search clearly wins.

Explanatory questions — "what is X, how does X work". Concept questions, not fact questions. Example: "how does the Mamba architecture differ from Transformers". The synthesis is dramatically faster than reading the SERP yourself.
Cross-source synthesis — comparisons, market scans, trend analysis. Work that would have meant opening five to ten tabs and stitching it together.
Vague natural-language queries — "that company, the one that raised a Series B last year, what was its name". When you do not know the keywords, intent parsing carries you home.
Cross-language searches — ask in Korean, synthesize from English sources. AI search just does this.
First-step coding — "how do I do X in this library" — though IDE-integrated tools (Cursor, Claude Code, Copilot Chat) are eating this seat fast.

Where classic search still wins.

Clean single-fact lookups. "Samsung closing price yesterday." It is one number. AI search is slower and pricier; the SERP knowledge box answers in a second.
Single official documentation page. React docs, Python docs — going straight there is faster. AI summaries lose subtle details.
Exploratory browsing. When you want to be surprised. Image search, design references, shopping. AI answers are too decisive to leave room for browsing.
Breaking and real-time information. Index freshness rules. Google still wins by a margin.
When the source itself is the information. Which outlet broke the story matters. AI answers smear the "who said this" question.
When the act of searching itself is sensitive. AI search gives away more context. The privacy posture is worse. That is exactly Kagi's pitch.

The serious 2026 information worker uses both. Default search box is still Google for fast facts. Deeper questions go to Perplexity or ChatGPT Search. Heavy research jobs go to Deep Research and you go make coffee. Then a human verifies.

Do not try to use only one tool. That is the anti-pattern.

7. The AI-native browser thesis — Comet, and beyond

Perplexity's Comet, Arc's Max (the AI-browser push the Browser Company once made), Brave's Leo integration, Opera's Aria — they all reduce to one claim. The future is not a search box; the browser itself becomes AI-native.

The argument:

A user's context does not fit in a search box. The user is already on a page. That page is context. Going back to a search box to start over breaks the thread. A sidebar assistant that already sees the current page is the natural shape.
Multi-tab is one task. Comparison shopping, market research — the user already has tabs open. The AI has to see all of them to be genuinely useful.
AI is becoming agentic. Not just answering, but operating pages on the user's behalf. That requires living inside the browser.

Comet is the cleanest implementation of this thesis. Sidebar assistant, multi-tab summarization, agentic mode. As of 2026 its installed base is small — hundreds of thousands. Against Chrome and Safari's market share it is a rounding error.

The headwinds are sharp.

Browser-switching cost is huge. Bookmarks, extensions, sessions. People do not move.
Chrome is integrating Gemini into itself. When Google puts the same feature into the browser most people already use, Comet's thesis gets absorbed.
Safari is integrating ChatGPT (the iOS/macOS 26 motion). Apple is going for the same seat.

The real value of Comet is that it serves as a lock-in mechanism so Perplexity does not lose the users it has. Consumer AI search is squeezed between ChatGPT and Google; owning the browser owns the user channel. The business logic is cleaner than the UX logic.

On the other hand, an AI-native browser is probably the final form of consumer AI search. Project forward five years and we likely do not perform "search" as a discrete act. An assistant is always on inside the browser; the pages we view, the tabs we open, the text we write — all of it is context. Perplexity Comet, Chrome+Gemini, Safari+ChatGPT — which one claims that seat is the next round of the fight.

The thesis is probably right. Who implements it is undecided.

8. Honest decision tree — what to use when

General users — knowledge workers, academics, journalists.

Fast fact checks — Google (if the search bar is already a habit) or Gemini AI Mode (same box, with an AI answer on top). Answer-in-five-seconds is what matters.
Explanatory or synthesis questions — Perplexity Pro or ChatGPT Search. Both are fine. If ChatGPT is already in your day, lean there.
Heavy research (market scans, academic synthesis, DD) — Deep Research. Pick among OpenAI (deepest), Perplexity (fastest), Gemini (freshest index) by job character. Always human-verify the output.
Privacy and tracking matter — Kagi. Accept the monthly cost.
Code-related — IDE-internal tools first (Claude Code, Cursor, Copilot Chat). If you still need a search tool, Phind.

Developers — RAG/agent builders.

Need to ship fast, standard integrations — Tavily. LangChain and LlamaIndex default.
Need semantic search and findSimilar — Exa. Content extraction is clean.
Need Google results, minimum cost — Serper / SerpAPI. Body extraction is your problem.
Multilingual results, enterprise governance — You.com Search API.
Index freshness plus citations attached to your own LLM calls — Gemini API Grounding with Google Search.
Own index plus no-training guarantee — Brave Search API.
Heavy usage, cost pressure — mix backends; trade quality vs cost. Avoid single-vendor dependency.

Teams and organizations.

Building a small consumer SaaS — answer engine builder. Look at Exa plus your own LLM, or Perplexity's Sonar API.
Internal knowledge search — Glean, Mem, and similar enterprise search live in a different market. Out of scope here.
Agent workflows (e.g. a Slack bot that fetches research) — Tavily ships fastest. Diversify to Exa or Serper once costs rise.
Copyright and licensing sensitivity required (media, legal) — Brave Search API (no training use), or your own crawl with explicit licensing.

Price sensitivity.

Can you stay free? — For consumers, Gemini AI Mode is already good enough free. For developers, start inside the Tavily / Exa / Brave free tiers.
Around $20/month — Perplexity Pro, ChatGPT Plus, Kagi Pro. Pick one based on workflow.
$100–$200/month — Perplexity Max or ChatGPT Pro. Heavy Deep Research users.
Variable usage — API side is inherently usage-based. If heavy use is likely, set a monthly budget and dashboard it.

Most common mistake: trying to cram every workflow into one consumer tool. Do not try to do everything in Perplexity. Fast facts are faster on Google. Code is better in IDE tools. Heavy research is a separate tool. Splitting two or three across workflows is the 2026 standard.

Epilogue — checklist, anti-patterns, what's next

A one-week check after adopting a tool

Replayed five real searches from the past week across two or three tools and compared.
Picked three random citations and clicked through to verify the fact actually appears in the source.
Ran Deep Research once and honestly assessed whether the time was worth it.
Checked mobile patterns (browser, desktop app, mobile app — UX differs).
Checked privacy settings (whether searches train models, retention policy).
For developer work — wired a small RAG demo against one API (Tavily or Exa).
Estimated monthly cost (subscriptions + API usage + Deep Research calls).

Anti-patterns — common mistakes

Lifting answers without verifying citations. The most common, most dangerous failure. Smooth ≠ true. Verify two or three load-bearing claims at the primary source.
Using Deep Research for trivia. Don't burn a thirty-minute report on a five-second fact. The tools have different purposes.
Forcing every workflow into one consumer tool. Fast facts, explanatory questions, heavy research, code — different tools win each lane.
Assuming AI search answers are 100% correct. 20–40% citation drift is a real number across products. Grade your answers by stakes.
Putting a RAG pipeline on a single API. Cost volatility and availability risk. Mix two or three with a fallback structure.
Ignoring publisher and copyright posture. For media or legal, prefer indexes with no-training guarantees (Brave) or explicit licensing.
Declaring search dead too early. In 2026 Google still handles the vast majority of queries. Both markets coexist.
Switching entire browsers to a new one like Comet overnight. Run it as a side browser for a month first. The cost of losing your daily browser is real.

Coming next

The next post is RAG search backend bake-off — Exa vs Tavily vs You.com vs Serper on the same queries. We throw the same one hundred queries at four APIs and quantify result relevance, body-extraction quality, latency, and cost. After that, inside an AI-native browser — how Comet reads pages, fuses multi-tab context, and runs agentic actions; and how to imitate the parts that matter inside your own product.

And after that, building a Deep Research system yourself — getting an OpenAI Deep Research-shaped multi-hop research agent to work in your own domain. The structure of search API plus LLM plus synthesis loop.