# ScrapingBee LLM Reference Notes

> Understand documented source discrepancies, LLM reliability rules, and the coverage of this reference.

## Source discrepancies and safe interpretation

The source documentation contains a few internal inconsistencies. This reference does not hide them:

| Topic | Safe interpretation used here |
| --- | --- |
| Google cost | The Google introduction says 15 credits, but its `light_request=true` default and pricing table say 10. Treat light requests as 10 and `light_request=false` as 15. |
| Google multi-page billing | The source defines aggregation for `pages` but does not state whether a multi-page call bills once or per fetched page. Confirm with the dashboard before relying on a projected cost. |
| Google `add_html` response | The standard example is a flat JSON object, while the `add_html` example wraps a body under `statusCode`. Handle either envelope and do not rely on the example comment, which names `full_html` rather than the documented `add_html` parameter. |
| Google shopping sort default | Prose calls `relevance` the default, while the parameter schema uses an empty default. Omit `sort_by` for default behavior unless a deterministic ordering is required. |
| Amazon Search light cost | One explanatory paragraph says 10 credits, while the quick-start and pricing table say 5. This reference follows the pricing table: 5 light / 15 regular per page. Verify the dashboard before billing-sensitive workloads. |
| `custom_google` | CLI documentation says 15 credits, while LangChain documentation says 20. The current HTML API parameter contract neither defines nor prices `custom_google`; proxy-mode Google traffic is separately documented as 20 credits per request. Do not infer that `custom_google` is supported by the direct HTML API or that it has a universal price; verify the dashboard. |
| Country targeting | `country_codes.md` says premium proxy is required; the HTML API page also documents stealth country targeting. Premium is the conservative choice. |
| `parse=false` | Mentioned only in Amazon/Walmart frontmatter, not described in their request sections. Do not rely on it without checking the live documentation. |
| Retry behavior | Google, Amazon, Walmart Search, YouTube Search, ChatGPT, and Gemini document up-to-30-second server retries. The Walmart Product and YouTube Metadata/Subtitles sections do not publish an endpoint-specific retry guarantee. |
| LangChain wrappers | The wrapper documents a subset or occasionally different shape from direct APIs (for example Google search types and Amazon Product parameter reuse). Use its per-tool contract and the direct endpoint contract; do not assume undocumented wrapper parameters work. |
| YouTube Search response | Source prose calls the response parsed, but the documented payload contains raw YouTube renderer structures. Treat it as an upstream-controlled, variable structure. |
| Full schemas | Hugo shortcodes provide detailed parameter tables and Remote MCP exposes live schemas through `tools/list`; this compact reference preserves operational controls but does not reproduce presentation-only sample payloads. |

## LLM reliability rules

When using this document as agent context:

- Treat endpoint paths, required field names, costs, and incompatibilities as hard facts.
- Prefer the `Authorization: Bearer` header. The legacy `api_key` query parameter remains supported for compatibility but is deprecated for new integrations.
- If asked for an unsupported or undocumented parameter, say it is not specified instead of inventing it.
- Do not use `GET /amazon`, `GET /walmart`, `GET /youtube`, `scrapingbee get`, a Bearer-header Remote MCP configuration, or `search` in place of a ChatGPT/Gemini `prompt`.
- For an action requiring live tool options, inspect the tool schema rather than constructing a call from memory.
- Country code support differs by proxy tier. Do not request a classic code absent from the classic lookup and expect a hard error: it silently falls back to `us`.
- In the CLI, keep URLs in double quotes but keep JSON and Smart Extract paths in single quotes; shell expansion can silently corrupt request data.
- Ensure the document is actually present in the model context or indexed by retrieval; an LLM cannot use a local file it was not given.

Representative routing checks:

| Question | Correct route |
| --- | --- |
| “Extract a product title and price from this URL.” | HTML API with `extract_rules`; use `ai_extract_rules` only if the page shape is unreliable. |
| “Get the current price for Amazon ASIN B08N5WRWNW.” | Amazon Pricing API with `asin=B08N5WRWNW`. |
| “Search current web results with ChatGPT.” | ChatGPT API with `prompt=...` and `search=true`. |
| “Configure this in Cursor as MCP.” | `mcp-remote` configuration with the `?api_key=` URL. |
| “Click a consent button after it appears.” | HTML API with `js_scenario` using `wait_for_and_click`. |
| “Search Google more cheaply and quickly.” | Fast Search API; use full Google Search only when detailed SERP data is needed. |
| “Find every product URL and title from this page.” | HTML API with `extract_rules`; consult Data extraction for a nested list schema. |
| “Run this from a proxy-only application.” | Proxy mode, then the HTML API capability references for parameters. |
| “Get a French premium proxy.” | Premium country-code lookup, then HTML API `premium_proxy=true&country_code=fr`. |

## Source coverage

This reference set incorporates:

- `_index.md` — HTML API, Auto-Mode, pricing, response behavior, usage, and proxy capabilities.
- `data-extraction.md` and `js-scenario.md` — extraction and interaction schemas.
- `proxy-mode.md`, `country_codes.md`, and `country_codes_classic.md` — proxy transport and complete country lookups.
- `google-api.md` (canonical) and `google.md` (duplicate/subset), plus `fast-search.md`.
- `amazon.md`, `walmart.md`, `youtube.md`, `chatgpt.md`, and `gemini.md`.
- `cli.md`, `remote-mcp.md`, `langchain.md`, `make.md`, `n8n.md`, and `zapier.md`.

The published files are directly maintained: `llms.txt` is the routing index, while individual endpoint, capability, lookup, and integration files are the operational source material. For version-sensitive pricing, changing APIs, or live MCP parameter shapes, verify the current ScrapingBee dashboard or live service before making a production decision.