Yandex AI Overview Scraper

Collects web-grounded answers from Yandex AI search products without representing ordinary search wizards as AI overviews.

Sources

• aliceWeb opens a fresh https://yandex.ru/alice/ chat for each query. Yandex requires a signed-in Yandex ID, so supply an authorized Cookie header through the secret cookieHeader input or YANDEX_COOKIE_HEADER environment variable.
• searchApi calls the official Yandex Search API v2 gen/search endpoint. Supply an API key with yc.search-api.execute access and its folder ID through secret inputs or YANDEX_SEARCH_API_KEY and YANDEX_FOLDER_ID.

Every record is explicitly labeled with its provider and retrieval method. The two products are not silently substituted for one another.

Modes

• single: one question.
• batch: up to 20 unique questions, with one record per web-grounded answer.

Alice uses a fresh one-use browser session per query so chat context cannot leak between questions. The Actor prefers validated JSON responses and uses current rendered DOM as a fallback. It extracts answer text/HTML, lengths, public source URLs, citation counts, follow-up questions when shown, locale, provider, and run metadata.

Safety and failure behavior

• Cookies, API keys, authorization headers, and raw response payloads are never logged or stored.
• CAPTCHA/robot-verification and sign-in walls fail immediately; the Actor does not solve or repeatedly retry them.
• An answer is stored only when it contains non-empty answer text and at least one public web source. Ungrounded answers complete without a record only when another query succeeds; a wholly empty run fails clearly.
• Empty values are omitted recursively, public source URLs are deduplicated, and the OUTPUT key-value-store record contains a dataset audit.

The legacy v2 Actor incorrectly pushed hasAiOverview: false placeholder records and matched broad Yandex “wizard” elements. Version 3 removes that behavior.