AI Search Agents: Three Platforms, Three Different Jobs

AI search agents crossed a threshold in May 2026. This post anchors to Google's information-agents announcement — made yesterday at I/O 2026 Day 1, May 19 — compared head-to-head against Perplexity Pro and ChatGPT Search / Deep Research as of today, May 20, 2026. The framing that no vendor-produced comparison will give you: these are three genuinely different tools optimized for three different jobs, and optimizing your presence for “AI search” as a single channel is the wrong frame.

The evidence is in the citation data. According to a Profound analysis of 680 million citations, only about 11% of domains are cited by both ChatGPT and Perplexity for similar prompts. ChatGPT leans on Wikipedia (47.9% of top-10 citations); Perplexity leans on Reddit (46.7%); Google AI Overviews leans on YouTube (23.3%). These are not quirks — they are architectural and design choices baked into each product. Building one “AI SEO” strategy across all three is like running one ad creative for three different audiences.

This guide covers the job-framing that separates the three agents, the verified pricing-tier decision matrix (Free → $20 → $100 → $200 across all three), the most rigorous public citation-error benchmark available, and the critical distinction between Google's information agents and Gemini Spark — two different products that much of the May 19 coverage conflated. Internal links to our agentic SEO service and our GEO citation guide expand on the operational response for brands navigating these three stacks.

The comparison wars between AI search products tend to miss the more useful question: what job is the user actually trying to do? When you frame it that way, the choice becomes much clearer. Perplexity was designed around the “I need a cited answer to a specific question right now” job. ChatGPT Search and Deep Research were built for “I need to synthesize a complex topic across many sources.” Google information agents — announced yesterday, not yet live — are built for an entirely new job: “monitor this topic for me continuously and tell me when something relevant happens.”

That third job is new. It reframes search from a pull activity (user initiates, agent responds) to a push activity (agent monitors, surfaces proactively). That is why the information-agents announcement deserves standalone analysis rather than being folded into a generic “Google vs Perplexity” comparison. It is not the same product category.

For AI search market context, Google's AI Mode already has 1 billion monthly active users — the distribution advantage is unmatched. But user count is not the same as job fit. A 1-billion-user surface that does not yet have live information agents is still losing to Perplexity Pro for a user who needs cited, real-time facts today.

At I/O 2026 on May 19, Liz Reid — Google's VP and Head of Search — announced information agents as a new capability baked directly into Google Search. The verbatim framing from Google's Search I/O 2026 blog: “Operating in the background, 24/7, these agents intelligently reason across information to find exactly what you need at exactly the right moment.”

The mechanics: a user enters a natural-language monitoring prompt — “keep me updated on nearby movie tickets for The Mandalorian and Grogu” or “track market movements in the EV sector with my parameters” — and the agent maps out a monitoring plan including the tools and data sources it needs. It then runs autonomously in the background and pushes a synthesized update with links via the Google app when conditions are met. Per Google, use cases include flight prices, live sports results, housing and job market trends, quarterly earnings, and breaking news.

These agents run on Gemini 3.5 Flash, which reached general availability on May 19, and they represent what one might describe as an evolution of the original Google Alerts (launched 2003) — though that is an observer's framing, not Google's literal claim. The distinguishing feature is synthesis and reasoning, not just keyword matching.

Two pricing constraints define who gets access. Information agents launch this summer in the US, exclusively for Google AI Pro ($19.99/mo) and Google AI Ultra ($99.99/mo) subscribers. They are not available in the free AI Mode tier, and the $99.99 AI Ultra entry-point is a repricing from the prior $249.99 — announced at the same I/O event per Engadget's coverage. The old $200 tier was renamed “AI Ultra Premium.”

The practical implication for today (May 20): information agents do not yet exist in a form you can subscribe to. The$19.99 AI Pro plan is live and covers other features, but the agents themselves are a summer 2026 launch. Teams evaluating whether to build these into their research or monitoring workflows should begin watching the TechCrunch how-to coverage for the rollout cadence.

Perplexity Pro costs $20/month ($16.67 on annual) and, per vendor architecture analysis, is reported to run a real-time web search for every single query — meaning it does not rely on a fixed training cutoff for freshness. This reported behavior distinguishes it from ChatGPT, which uses a hybrid of training data plus selective browsing depending on query type. Note: Perplexity has not published a formal whitepaper confirming this architecture, so frame it as “reported” rather than verified engineering fact.

Pro's standout data point is from the Columbia Journalism Review / Tow Center benchmark (March 2025): Perplexity had the lowest citation-error rate of the eight tools tested — 37%. For context, Grok-3 was at 94% and the collective failure rate across all 8 tools was over 60%. That is a meaningful signal. It is also a 2025 snapshot taken before GPT-5.5 (launched April 23, 2026) and before Perplexity's own Sonar Pro and Sonar Reasoning Pro models matured — so treat it as directional, not a locked current ranking.

The Pro tier includes unlimited Pro Search, 20 Deep Research queries per day (using the Sonar Deep Research model), and a multi-model picker giving access to GPT-5.4, Claude Opus 4.7, and Gemini 3.1 Pro as alternatives. For teams building on top of Perplexity's infrastructure, our Perplexity developer API guide covers the Sonar family (Sonar, Sonar Pro, Sonar Reasoning Pro, Sonar Deep Research) and their pricing tiers. Perplexity's Comet browser extends the Pro subscription with native browser integration and Comet Plus ($5/mo, bundled with Pro/Max) for access to premium publisher content from CNN, Washington Post, and Fortune.

Perplexity Max — launched in 2026 at $200/month — is the power-user ceiling with higher quotas and priority access, mirroring ChatGPT's Pro $200 structure. Perplexity reached approximately 45 million monthly active users by early 2026 and is valued at around $21B, making it a well-funded challenger, not a side project.

ChatGPT Search runs on GPT-5.5, which launched April 23, 2026 and is now the top model across Plus, Pro, Business, and Enterprise tiers. GPT-5.5 specifically powers the Search browse mode — the feature that pulls live web results into ChatGPT responses. This is a different model from Deep Research: ChatGPT Deep Research runs on a GPT-5.2-based model (updated February 2026), which was optimized for multi-step agentic research rather than single-query retrieval. Do not collapse them — the distinction matters for evaluating output quality and cost by task type.

ChatGPT Atlas — OpenAI's agentic browser introduced in 2026 — extends the platform further into web-based autonomous task completion. Our ChatGPT Atlas guide covers the browser strategy in depth. Atlas added agent mode for Plus/Pro/Business users.

The Tow Center benchmark put ChatGPT Search at 67% citation error in March 2025 — 30 percentage points higher than Perplexity's 37%. That gap was measured on a news-attribution task (identify article title, publication date, publisher, URL from a quote) rather than general synthesis, and the test predates GPT-5.5 — so the 67% figure is a historical ceiling, not a current rating. ChatGPT's strength has always been synthesis and reasoning, not raw citation accuracy on lookup tasks.

On the pricing side: ChatGPT Plus is $20/month (Deep Research capped at 10 queries/month). ChatGPT Pro split into two tiers in April 2026 — Pro $100/month (launched April 9) and Pro $200/month, the power ceiling with 250 Deep Research queries per month. ChatGPT also offers a Go tier at $8/month (ad-supported in the US) for users who want more than the free offering without the full Plus commitment.

The 11% domain-overlap figure between ChatGPT and Perplexity is the headline, but the citation-source bias data is more operationally useful for teams thinking about content strategy. According to the Averi.ai B2B SaaS Citation Benchmarks Report — which aggregated Profound (680M citations, Aug 2024–Jun 2025), SE Ranking (129K domains across 216,524 pages), and Surfer (36M AI Overviews, Mar–Aug 2025) — each platform has a distinct citation profile.

ChatGPT cites Wikipedia for 47.9% of its top-10 citations — nearly half. Reddit accounts for 12.9%, YouTube for 8.6%. Perplexity's pattern is strikingly different: Reddit leads at 46.7%, Wikipedia drops to 19.8%, YouTube at 13.4%. Google AI Overviews prioritizes YouTube at 23.3%, with Reddit at 21% and Wikipedia at 18.4%.

What this means analytically: the three systems are not converging on the same definition of “authoritative source.” They are trained, fine-tuned, and RLHF'd on different implicit quality signals. ChatGPT's Wikipedia dominance suggests its training data overweighted encyclopedic sources. Perplexity's Reddit dominance reflects its real-time index pulling heavily from community discussion. Google's YouTube bias is the least surprising — YouTube is a Google property and its content has been deeply indexed.

The forward-looking implication for content teams: if you want citation visibility across all three agents, you need fundamentally different content types. Wikipedia presence and Wikipedia-adjacent factual writing targets ChatGPT. Community-format content (long forum posts, Q&A threads, Reddit-style commentary) targets Perplexity. Video content and YouTube optimization targets Google AI Overviews. Our GEO citation guide covers the operational playbook for each stack in detail.

The cleanest way to frame the three-platform purchase decision is tier-by-tier. All three have something near a $20/month entry paid tier — but what that $20 buys diverges sharply. Above $20, the divergence accelerates. The matrix below is sourced from the vendors' own pricing pages (Perplexity pricing) plus the Google AI subscriptions blog and ChatGPT's pricing page, all as of May 20, 2026. Verify on vendor sites before subscribing — these tiers are in active flux.

One observation worth making explicit: the three-platform pricing structure has converged at the $20 tier (Google $19.99, Perplexity $20, ChatGPT $20) and again at the $200 tier (all three have an offering). The strategic divergence is in the middle — Google is racing to fill the $99.99 slot (repriced AI Ultra), while Perplexity jumps directly from $20 to $200. If autonomous monitoring is your use case, the $99.99 AI Ultra is today the most important new pricing tier in the market. For brands tracking their AI search visibility across all three platforms, the tooling costs now need to account for subscriptions across multiple tiers.

The Columbia Journalism Review / Tow Center benchmark is the most rigorous independent comparison of AI search tools available as of May 2026. Understanding its methodology matters for calibrating how much weight to put on its numbers.

The study ran in March 2025. Researchers fed each of eight AI search tools — ChatGPT Search, Perplexity, Perplexity Pro, Gemini, DeepSeek Search, Grok-2 Search, Grok-3 Search, and Copilot — a quote from one of 200 articles (20 publishers × 10 articles) and asked the tool to identify the article title, publication date, publisher name, and URL. This is a news-attribution task, not a general-purpose benchmark. The collective failure rate was over 60% of 1,600 queries. Grok-3 was worst at 94%. Perplexity was best at 37%.

What the benchmark does confirm — regardless of the specific percentages, which will shift with each model update — is the directional ranking at the time: Perplexity's citation architecture outperformed ChatGPT Search on this specific task by a substantial margin. The operative lesson is not “Perplexity is always more accurate” — that is not what the data says. The operative lesson is “even the best AI search tool at the time was wrong more than a third of the time on news attribution, and all of them presented wrong answers with confidence.”

Third-party blog roundups citing “92% Perplexity accuracy” and “87% ChatGPT accuracy” circulate frequently — these numbers come from LMSYS-adjacent sources and are not independently verified at the methodological level of the Tow Center work. The 37%/67% figures from Tow Center are the most rigorous public benchmark available and should be treated as the reference floor, not the third-party roundup numbers.

The zero-click and AI answers context is relevant here too: the broader trend toward AI answers reducing click-through is documented in our zero-click crisis analysis. If agents are surfacing wrong answers with confidence, the reputational stakes for cited brands are higher, not lower.

This distinction got lost in nearly all May 19 coverage. Google announced two distinct agentic products at the same I/O keynote, and they are different products at different price points for different use cases. Conflating them leads to wrong purchase decisions and wrong competitive framing.

Information agents are a Search-surface feature. They run on Gemini 3.5 Flash. They monitor specific topics or conditions and push synthesized updates via the Google app. They are accessible at the AI Pro tier ($19.99/mo). They are search-adjacent — an evolution of the query interface, not a replacement for your operating system.

Gemini Spark is a full 24/7 personal agent. Per TechCrunch's coverage, Spark runs on dedicated Cloud VMs and continues working when your device is off. It has MCP integrations with Canva, OpenTable, and Instacart. It has Gmail integration for autonomous task management. It is exclusively available at the AI Ultra tier ($99.99/mo) — not at AI Pro.

The operational framing: if you want autonomous search monitoring, you need information agents (AI Pro, summer 2026). If you want a broad personal agent that can take actions across apps and services, you need Gemini Spark (AI Ultra). If you want both, AI Ultra at $99.99 includes information agents as well. If you want neither and just need current AI search results with good citation quality, Perplexity Pro at $20 is available today and does not require a Google subscription.

For teams building SEO strategies around the agentic search landscape, the distinction matters because the audience for each product is different. Our agentic SEO service addresses how brands should prepare their content and citation profile for both the current real-time search (Perplexity, ChatGPT) and the upcoming autonomous monitoring wave (Google information agents). The Q2 2026 citation analysis in our AI search citation domains ranked post provides the data layer for that planning.

For the broader operational response to I/O 2026 and what it means for SEO teams right now, our companion post SEO after Google I/O 2026: the 90-day playbook covers the concrete action priorities.