AI Online Search — ask ChatGPT & Gemini, get answers + sources

Ask any question and get the answer from ChatGPT (web search) and Gemini (web grounding) side by side, with the source links each AI actually used, plus brand-mention and competitor share-of-voice tracking.

No account. No login. No API key. No cookies.

What you get per question, per AI

Field	Meaning
answerText / answerMarkdown / answerHtml	The full answer in three formats
citations[]	title, url, domain, snippet, position for every source the AI cited
citationCount, uniqueDomains	Quick citation stats
brandMentions[]	Was your brand mentioned, how many times, where in the answer, with context snippets
competitorMentions[]	Same, for competitors
shareOfVoice	Your mentions ÷ (yours + competitors') × 100
citedByBrand	Did the AI link to your own domain
responseTimeMs, costUsd, scrapedAt, errorCode	Run metadata

Real measured numbers

10-question benchmark (5 questions × 2 AIs, Apify 4 GB, residential proxy, Sept 2026):

Engine	Success	Answer length	Citations	Time per question	Cost per question
ChatGPT	4/5	130–360 words	4–8 sources	11–13 s (first try)	$0.006–0.0065
Gemini	5/5	118–600 words	0–6 sources	14–21 s	$0.008–0.012

Both English and Vietnamese questions return native-language answers with local sources (e.g. pnj.com.vn, vn.investing.com for Vietnamese finance questions).

Input

• queries — one or more questions, written exactly as a person would ask them.
• engines — chatgpt, gemini (both by default), googleAiMode (experimental, often blocked by Google's bot check).
• country / language — two-letter codes; answers and sources are localized.
• brandKeywords / competitorKeywords — what to watch for in the answers.
• includeScreenshot — save a PNG of each answer as proof (off by default, faster without).
• followUpQuestion — optional second question in the same ChatGPT conversation.
• maxConcurrency / maxRetries — throughput and resilience.
• proxyConfiguration — residential proxies are strongly recommended.

Error codes

Failed questions still produce a row, with an empty answer and errorCode:

• LOGIN_WALL — the AI refused the anonymous session on that attempt (retry usually fixes it).
• BLOCKED — bot check / captcha page.
• EMPTY_ANSWER, NO_ANSWER — the page rendered but produced no usable answer.

Nothing is ever reported as a success with an empty answer.

Typical uses

• Track whether AI assistants recommend your brand, and which sources they trust.
• Compare how two different AIs answer the same buyer question.
• Feed grounded, cited answers into your own agents and reports.

For AI agents (MCP-ready)

This actor is built to be called by AI agents. It works out of the box with the Apify MCP Server — add it to Claude Desktop, Cursor or any MCP client and the agent can ask ChatGPT and Gemini on its own:

{
  "mcpServers": {
    "ai-search": {
      "url": "https://mcp.apify.com/?actors=andrew_babo/ai-search",
      "headers": { "Authorization": "Bearer <YOUR_APIFY_TOKEN>" }
    }
  }
}

Agent skill (paste into your agent's instructions)

You can ask real consumer AI assistants (ChatGPT with web search, Gemini
with web grounding) using the "ai-search" tool. Use it when you need the
answer a real user would see, the source links the AI actually cited, or
brand/competitor mention tracking — NOT for general knowledge you already
have.

HOW TO CALL
- { "queries": ["best project management tools in 2026?"] }
  One row per engine per question, so you can compare ChatGPT vs Gemini.
- Localize with "country" (US, VN...) and "language" (en, vi...).
- Brand tracking: set "brandKeywords" and "competitorKeywords"; each row
  returns brandMentions, competitorMentions and shareOfVoice.

OUTPUT CONTRACT
- answerText / answerMarkdown: the full answer as shown to a user.
- citations[]: title, url, domain, snippet, position — the sources the AI
  actually used. Cite these, not your own guesses.
- errorCode on failed rows: LOGIN_WALL or BLOCKED (retry usually fixes),
  EMPTY_ANSWER. A row with errorCode is NOT an answer — retry or report
  the failure, never present an empty answer as success.
- noResults: true with errorCode NO_INPUT means the input was empty.

COST
- ~$0.006–0.012 per question per engine, 11–21 s each. Keep batch sizes
  modest and reuse results instead of re-asking.

Changelog

• 0.3 — Added the Gemini engine (anonymous, web-grounded, full citation extraction). Default engines are now ChatGPT + Gemini. Google AI Mode marked experimental.
• 0.2 — Anti-detection browser layer, reliable ChatGPT answer and source extraction.
• 0.1 — First release.