Bright Data - Search

One name for everything that starts from a search query: Google, Bing and the other engines, the Google verticals, and the answer engines. The user never has to name a product. Pick one, do the work, and state the choice in one line the user can override with one word.

The line that decides: a query that returns a results list is this skill. Pulling records from a known place, like Google Maps reviews or a LinkedIn profile, is scrape. One known URL is fetch.

A SERP page asked for as raw HTML with no URL is still this skill - that is the REST path with brd_json left off the search URL (on a zone whose default format is HTML, see references/serp.md), because the CLI always returns parsed results for Google. With a URL, it is fetch.

The fork - now versus volume

The ask	Path	What it is
No volume word. The default.	SERP API	Structured JSON in one synchronous call. About 10 results per request.
"Top 100", "all the results", a rank report past page one	Web Scraper API - top 100 Google results	A Web Scraper API dataset job. Trigger, poll, download. Not instant. One request for 100 results, against ten SERP requests for the same depth.
"daily", "weekly", "track this keyword"	SERP or the top-100 job, on a timer	Recurring and only the top ten matters -> SERP plus a cron the agent writes. Recurring and the user wants all results and their ranks -> the top-100 job on a timer. Answer-engine tracking is a dataset job plus a cron, see references/answer-engines.md.
"top 5", "the first 3", any number under ten	SERP API	The normal fast path, then truncate the list to what was asked.

Only Google comes back parsed by default. On the CLI, --engine bing returns the raw results page unless you add --json, and Yandex has no parser at all, so its page comes back raw. The MCP search_engine tool returns Bing as page markdown whose links are bing.com/ck/a redirects; for clean Bing results use REST with "data_format": "parsed_bing_api". Details in references/serp.md.

A volume word always means more than the default 10, never fewer. "Top 5" is not a smaller or cheaper job, it is this same one call with a shorter answer, so never route it anywhere else and never treat it as a reason to change product.

Route silently, defaulting to SERP.

A single agentic query can also run through the MCP server's search_engine tool - the brightdata-mcp skill covers that surface.

Rank tracking - pick depth first

A top-ten snapshot on a timer is SERP plus a cron the agent writes. A full rank report every run is the top-100 dataset job on a timer, one request per run instead of ten.

Example - the fast path

The ask: "what does Google show in the US for best crm for startups?"

1. Starts from a query, not a URL. Stay in this skill.
2. No volume word, no "100". Fast path.

bdata search "best crm for startups" --country us --pretty

If it fails with zone "<name>" not found, a stale default_zone_serp is in the way: pass --zone with an unlocker zone from bdata zones --json, which beats every other source, and fix the config as references/serp.md describes.

Then state the choice:

Ten live Google results through SERP API. Say top 100 for the deeper set, which runs as a job and takes longer.

Answer engines belong here

What ChatGPT Search, Gemini, Perplexity or Google AI Mode answer about a brand is still a query in and an answer out. Each is its own pre-built scraper, so each runs as a dataset job, same shape as the top-100 path. Their ids are in references/answer-engines.md and in the live scrapers catalogue; confirm one with find-scraper.mjs <gd_ id> before the first trigger.

KYC

None for normal SERP use. Send the user to brightdata.com/cp/kyc only if a call is actually refused for that reason.

Read next

• Read references/serp.md before the first call, when the ask is a live results list from Google, Bing or Yandex, when the ask is one fast page of Google Images (a dataset scraper also exists, see google-scrapers.md), when a zone error blocks the call, or when the same query has to repeat on a timer.
• Read references/google-scrapers.md when the ask names a Google vertical or wants depth past page one: top 100 results, Shopping, Maps, Hotels or Flights.
• Read references/answer-engines.md when the query targets ChatGPT, Perplexity, Gemini or Google AI Mode.
• The general trigger, poll, download mechanics: the scrape skill's references/web-scraper-api.md. Not repeated here.
• A refused call: the agent-onboarding skill for "No API key found", 401, 407, and an account with no unblocker or serp zone at all (its check-auth.mjs --for search). The zone name does not matter: cli_unlocker, agent_unlocker and mcp_unlocker all work. A stale or missing serp zone is different: logging in again does not clear it, so fix it in references/serp.md's zone section. The billing skill for cost and credits.

Red flags - stop if you catch yourself doing one of these

• Promising 100 results at SERP speed, or calling the top-100 job instant
• Reaching for the top-100 job when the user never asked for volume
• Sending a results-list ask to scrape, or handling a records-from-a-known-place ask here
• Switching product because the job repeats. Depth picks the product, "daily" is only the scheduler
• Guessing a dataset_id instead of looking it up live
• Sending the user to KYC before a call is actually refused