AI Search 2026: Perplexity, Google AI, SearchGPT

The three AI search tools carry accuracy risks of different kinds, so source-tracing ability leads the choice for research workflows

Between 2025 and 2026 AI search tools grew from an emerging feature into a standard component of everyday research workflows. The three main tools — Perplexity, Google AI Overviews and SearchGPT — differ significantly in source transparency, hallucination rate and availability for Hong Kong users, and understanding these differences is essential to choosing the right tool.

Perplexity: The Leader in Citation Transparency

Perplexity's scale figures as of early 2026: 45 million monthly active users, 170 million global monthly visits, 780 million queries a month, and annualised ARR above US$450 million in March 2026. The Deep Research feature is available on the Pro and Max tiers and is currently powered by Claude Opus 4.5/4.6.

On accuracy, two important sets of data need to be assessed separately: a March 2025 study by the Tow Center at Columbia Journalism School (covering 1,600 queries and 8 AI tools) found Perplexity's standard search citation error rate at 37% — the lowest among the tools tested, but still significant in absolute terms. A 2026 Towards AI evaluation found Perplexity Deep Research mode accuracy at 92.3%, but this evaluation is directional data and not peer-reviewed research.

Perplexity cites 1,430 unique news sources (Google: 881), and its citation overlap with Google on the same query is only 10 to 25%, source diversity being a standout feature. It can be used freely in Hong Kong, and HKT also bundled Perplexity Pro for consumers in 2025.

Google AI Overviews: Market-Leading but With Trade-offs

Google AI Overviews holds a 61% share of the AI search market (Q1 2026) and has the fastest response (1 to 3 seconds, against 2 to 5 seconds for Perplexity). The underlying Gemini 3.1 Pro has a hallucination rate of 10.4% (Vectara leaderboard, March 2026), which is low among mainstream AI models.

Depth of indexing of Hong Kong local content is Google's greatest advantage — its coverage of Chinese-language media and local business information far exceeds competitors. The main drawback is that AI summaries provide source links below the summary in place of inline citations, so source transparency falls short of Perplexity; and the AI summary block coexists with ads, People Also Ask and other elements, giving lower information density.

SearchGPT: Limited Capability but Enterprise Channels Available

OpenAI's ChatGPT search feature (SearchGPT) has an 18% share of the AI search market (Q1 2026). Because OpenAI has restricted Hong Kong users since July 2024, direct access to ChatGPT is unavailable in Hong Kong. The Microsoft Azure OpenAI enterprise channel and Microsoft Copilot are available in Hong Kong, with Deep Research powered by GPT-5.2, the main path for Hong Kong enterprises to use OpenAI's search capability through compliant channels.

Gemini Deep Research

The Google AI Pro subscription ($19.99/month) includes the Gemini Deep Research feature, available directly in Hong Kong. Google Cloud Vertex AI provides an enterprise API in Hong Kong with data residency options, suitable for organisations with compliance requirements.

A Tool Selection Framework for Hong Kong Users

Summary

The accuracy risk differs in nature across the three tools: Perplexity's risk is citation error (the citation exists but the content does not support the claim); Google's risk is oversimplified summaries; and SearchGPT's direct availability in Hong Kong is already a barrier. In research workflows, a tool's source-tracing ability should be the main basis for selection, and "the smartest" alone falls short as a criterion.

For the division of labour between Perplexity and Claude, see Perplexity vs Claude: Fundamental Differences in Use Case Positioning; for concrete citation verification practice, see AI-Assisted Research: Hallucination Detection, Citation Verification and Source Tracking; for continuous news monitoring scenarios, see AI News Monitoring: The Data Source Question Vendors Skip in Quotes.