Working with Results

Once a task reaches succeeded, the answer and its sources are in the task object. This page covers how to read them and where the common pitfalls are. For the full field definitions, see The Task Object.

One result, three shapes

Field	Contents	When to use it
answer	The answer text (Markdown), identical to raw_result.rawContent	When you only need the text
raw_result	The structured result: answer text, sources, execution steps, cards, timing, and more	Recommended for most use cases
result_json	The original result as a JSON string	For full archiving, or to read new fields raw_result doesn’t expose yet

The three overlap, and each field can be up to 8 MiB, so task responses can be large. raw_result is only present once the task has a result (it succeeded, or the platform run failed). Queued, canceled, or dispatch-failed tasks don’t have this field, so check for it before reading.

How raw_result is encoded

raw_result uses protobuf’s JSON encoding, which differs from the outer fields:

• Field names are camelCase: rawContent, shareLink, agentSteps — whereas outer fields use snake_case like share_link.
• Zero-value fields are omitted: success: false, thinking: false, empty strings, empty arrays, and a rank of 0 don’t appear. Fill in defaults when reading — in Python, for example, raw.get("references", []).
• elapsedMs (execution time in milliseconds, queueing excluded) is a string, such as "48213"; convert it to a number yourself.
• completedAt is a UTC timestamp in ISO 8601 format, such as 2026-10-26T12:41:37.512Z.

Answer text

rawContent (and answer) is the platform’s original Markdown. The platform’s own citation markers and card placeholders are kept as-is, not converted.

On Global platforms (ChatGPT, Gemini, Claude, Perplexity, Copilot, Grok, Google AI Mode, Google AI Overview), the answer text is returned as the platform wrote it; no platform-specific marker conventions have been catalogued for them yet.

For China platforms, these are the markers we’ve observed:

Platform	Markers in the answer text (examples)
DeepSeek	[citation:3]; [reference:3] with thinking mode on
Tencent Yuanbao	[citation:2]; card placeholders [](@replace=media_…)
Kimi	Citations wrapped in <REF>…</REF>
Zhipu ChatGLM	【turn0search2】
Metaso	[[2]]; images left in the text as Markdown images
ERNIE Bot	^2^ (uncommon); may also appear as plain Markdown links
Qwen	Card placeholders such as [(video_note_list_1)]
Doubao, iFlytek Spark	Citation positions aren’t marked in the text
Xiaohongshu Diandian	Platform-native <u> underline tags, zero-width spaces, etc.

Marker formats are up to each platform

The observations on this page are from October 2026, and a platform redesign can change them at any time. Clean up these markers before showing the text to users if you need to, but don’t rely on their format for strict validation — use references and cited to determine what was cited.

Sources: references

references lists every source the platform found or read for this answer, deduplicated by URL. Sources cited in the answer text come first, followed by those found but not cited. Each item:

Field	Description
rank	Display order on the platform, starting at 1
title	Title
url	Link; platform redirect links are resolved to the original URL where possible
snippet	Excerpt; may be empty
source	Source name (site name); may be empty
domain	Domain; may be empty
contentPlatform	The content platform detected (such as a particular content community); empty if not detected
cited	Whether the answer text cites it — three states, see below
publishTime	Publish time: ISO 8601, YYYY-MM-DD, or the platform’s relative wording (e.g. “1 day ago”); may be empty
author	Author (post author, video creator, etc.); may be empty

The three states of cited

Value	Meaning
true	The answer text cites this source
false	The platform found or read it, but the answer text doesn’t cite it
Field absent	The platform doesn’t distinguish cited vs. uncited — don’t treat this as false

By platform (observed October 2026):

• Global platforms: cited is not set. Every source arrives without the field.
• Doubao, Kimi, Tencent Yuanbao, Zhipu ChatGLM, Metaso, ERNIE Bot: distinguish true / false.
• DeepSeek: sources it can tie to a specific web page are marked true; sources it can only tie to a batch of search results are left unset.
• iFlytek Spark: the platform currently doesn’t return search sources, so references is built from links the model wrote into its answer. cited is always true, and there’s no information beyond the title.
• Qwen and Xiaohongshu Diandian: cited is not set.

Things to watch when reading

• url isn’t always a valid link. A few platforms have sources without a link (map plugins, for example), leaving url empty; some return an internal platform ID instead. Check that it starts with http:// or https:// before using it.
• domain can be empty. If you’re aggregating by domain and domain is empty, parse it from url.
• An empty references list is a normal result. Some questions don’t trigger a web search, so the answer has no sources.

Example: was your site cited?

The code below finds sources from a given domain in the answer and splits them into “cited in the answer”, “found but not cited”, and “platform doesn’t distinguish”:

Python

from urllib.parse import urlparse


def match_domain(ref, domain):
    host = ref.get("domain") or urlparse(ref.get("url", "")).hostname or ""
    return host == domain or host.endswith("." + domain)


def citation_report(task, domain):
    refs = (task.get("raw_result") or {}).get("references", [])
    hits = [r for r in refs if match_domain(r, domain)]
    return {
        "cited": [r for r in hits if r.get("cited") is True],
        "not_cited": [r for r in hits if r.get("cited") is False],
        "unknown": [r for r in hits if "cited" not in r],
    }


report = citation_report(task, "example.com")
print({k: len(v) for k, v in report.items()})

Node.js

function matchDomain(ref, domain) {
  let host = ref.domain ?? "";
  if (!host) {
    try {
      host = new URL(ref.url ?? "").hostname;
    } catch {
      host = "";
    }
  }
  return host === domain || host.endsWith("." + domain);
}

function citationReport(task, domain) {
  const hits = (task.raw_result?.references ?? []).filter((r) => matchDomain(r, domain));
  return {
    cited: hits.filter((r) => r.cited === true),
    notCited: hits.filter((r) => r.cited === false),
    unknown: hits.filter((r) => r.cited === undefined),
  };
}

const report = citationReport(task, "example.com");
console.log(Object.fromEntries(Object.entries(report).map(([k, v]) => [k, v.length])));

On Global platforms every hit lands in unknown, because they don’t set cited. There, “the platform used a page from your domain as a source” is the strongest signal available.

Execution steps: agentSteps

Some platforms expose how they arrived at the answer, in order. Read the fields that match each step’s type:

type	Fields to read	Contents
thinking	content, summary	The reasoning text and the platform’s summary of it
search	name, queries, results	The search queries the AI actually ran, and the results of each search (title, url, siteName, snippet) — not all of which are cited
tool	name, args	Other tool calls and their raw arguments

The queries of search steps are especially useful: they show what the AI actually searched for to answer the question, which can guide the keywords in your content. When a platform doesn’t expose its steps for a tier, agentSteps is absent. On Global platforms, agentSteps is currently usually absent.

thinking steps mostly appear only with thinking mode on; with it off, some China platforms such as Qwen, Zhipu ChatGLM, and Metaso may still include a short thinking or progress note. To tell whether an answer was produced with thinking on, check raw_result.thinking, not whether there are thinking steps.

Media cards: mediaCards

Video, image, product, and ranking cards in the answer, in the order they appear:

• type: image, video, product, ranking, or other.
• anchor: the card’s placeholder text in rawContent, which tells you where the card sits. It’s empty when the platform doesn’t put placeholders in the text; then you can only go by order.
• items[]: each entry in the card, with title, url, imageUrl, source, author, duration, price, rating, publishTime, and description. Fields that weren’t captured are omitted.

On Global platforms, mediaCards is currently usually absent.

Image links expire

imageUrl is usually a signed, temporary link from the platform and stops working after a while. For long-term storage, keep the url (the video, post, or product page), or copy the image to your own storage soon after you get the result.

Share link

share_link (raw_result.shareLink) is the platform’s share page for this conversation, letting someone review the original answer on the platform. It’s an empty string when the platform doesn’t provide one or creating the share failed; availability varies on Global platforms. Note that Xiaohongshu Diandian share links only open inside the Xiaohongshu app — they load as a blank page in a browser.

raw_result.conversationId and messageId are the platform’s own conversation and message IDs, for reference only. The API doesn’t support continuing a conversation from them.

Results of failed tasks

Tasks whose platform run failed (error_code is crawl_failed) also carry a raw_result. In it, success is absent (that is, false), rawContent is empty, and error holds the failure message (the same {"zh","en"} JSON string as error_message). To detect failure, check the outer status and error_code — see Error Handling.

Related pages

The Task ObjectDefinitions of the outer fields and every raw_result sub-structure.

Platforms, Priority & CountryEach platform's region and status.