scanned May 12, 2026

Parallel

parallel.ai

ai-mlVisit site →

Parallel provides a web search API purpose-built for AI applications, offering high accuracy and predictable costs

44/100

Tier 3 · Agent-Accessible

Content answers33/100
Protocol plumbing69/10011 of 16 checks pass
Re-audit2867h ago· 1 credit

Scored by asking 15 questions a buyer of a ai-ml product asks, then grading this site’s own pages: answered, hedged (partial or vague), or silent (no page answers it). How scoring works

15 buyer questions · graded against the site’s own pages

1 of 15 come back answered. 8 get a hedge. 6 meet silence.

1 answerable8 half-answers6 silent

Goes silent · integration · high importance

I saw you have an MCP server for competitive intelligence. Does it support all MCP 2024-11-05 spec features like sampling, roots, and prompts, or is it a basic tools-only implementation?

Goes silent · limits · medium importance

Our workload involves launching thousands of parallel research tasks. Is there a hard ceiling on concurrent tasks per account, and does it vary by processor type or pricing tier?

This report is public. Own parallel.ai? Claiming is free: crawl every page, re-audit as you fix, and track your score over time.

Sign in to claim

The fix queue

56 points sit between parallel.ai and 100: 14 open questions and 5 missing protocol checks, ordered by estimated payoff.

Point estimates are per fix under scoring v2. They are not additive to a promised total.

01integration · importance highGoes silent+7 content pts est.

I saw you have an MCP server for competitive intelligence. Does it support all MCP 2024-11-05 spec features like sampling, roots, and prompts, or is it a basic tools-only implementation?

What the pages say

No page on the site addresses this.

The fix

Add documentation page or blog post detailing MCP protocol version support, specifically which spec features are implemented (sampling, roots, prompts, tools) and which MCP revision (2024-11-05 or 2025-03-26) the servers target.

confidence high · grounding world-knowledge · weight 0.00 · Absent

02limits · importance mediumGoes silent+7 content pts est.

Our workload involves launching thousands of parallel research tasks. Is there a hard ceiling on concurrent tasks per account, and does it vary by processor type or pricing tier?

What the pages say

No page on the site addresses this.

The fix

Add documentation on concurrency limits, rate limits, or account-level quotas for parallel task execution, including whether limits vary by plan tier or processor type.

confidence high · grounding synthesized · weight 0.00 · Absent

03technical · importance mediumGoes silent+7 content pts est.

The Chat API is intriguing for our conversational search use case. What's the maximum context window for conversation history, and are search results injected as system messages or user messages?

What the pages say

No page on the site addresses this.

The fix

Add a dedicated Chat API reference page documenting: (1) maximum context window/token limits for conversation history, and (2) the message structure showing how search results are injected (system message, user message, or separate field).

confidence high · grounding synthesized · weight 0.00 · Absent

04security · importance lowGoes silent+7 content pts est.

Our use case involves analyzing publicly leaked databases for security research. Does your Acceptable Use Policy explicitly prohibit accessing leaked credentials, or only distributing them? I need the exact prohibited activity wording.

What the pages say

No page on the site addresses this.

The fix

Add explicit language to the Acceptable Use Policy addressing whether accessing, analyzing, or distributing leaked credentials or data from public breach databases is permitted or prohibited, with specific wording for security research use cases.

confidence high · grounding synthesized · weight 0.00 · Absent

05support · importance lowGoes silent+7 content pts est.

We're evaluating Parallel for a government contract that requires guaranteed support response times. Do you publish specific SLA tiers with P1/P2/P3 response commitments, or is it best-effort with account management only?

What the pages say

No page on the site addresses this.

The fix

Create a dedicated Support or SLA page that documents guaranteed response times by severity tier (P1/P2/P3), support channels, and escalation procedures for enterprise and government customers.

confidence high · grounding world-knowledge · weight 0.00 · Absent

06technical · importance lowGoes silent+7 content pts est.

Your semantic search mentions 'web index'—is this the same underlying index as your keyword search, or a separate vector index? And what embedding model generates the vectors (your own, OpenAI, open-source)?

What the pages say

No page on the site addresses this.

The fix

Add technical documentation or architecture page explaining: (1) whether semantic and keyword search share an index or use separate vector/inverted indexes, and (2) which specific embedding model(s) Parallel uses (OpenAI, open-source, or proprietary).

confidence high · grounding world-knowledge · weight 0.00 · Absent

Hedged · 8 of 15 questions

A buyer gets something, then has to guess the rest. Expand any row for the evidence and the fix.

07The Core2x processor mentions 'improved compute control'—what specific parameters can I tune (timeout, memory, CPU) and what are the min/max values for each?technicalHedges+7 pts est.

The Core2x processor mentions 'improved compute control'—what specific parameters can I tune (timeout, memory, CPU) and what are the min/max values for each?

What the pages say

No page on the site addresses this.

The fix

Add a dedicated documentation page or expand the Core2x announcement with specific tunable parameters and their valid ranges. The current content uses 'improved compute control' as marketing language without technical substance.

Where we looked: /blog/core2x-processor, /blog/core2x-processor, /blog/core2x-processor

confidence high · grounding synthesized · weight 0.00 · Vague

08For our academic compliance workflow, I need to know: when Deep Research generates citations, does it provide raw URLs only, or does it include access dates, page titles, and archive links like Wayback Machine references?operationsHedges+7 pts est.

For our academic compliance workflow, I need to know: when Deep Research generates citations, does it provide raw URLs only, or does it include access dates, page titles, and archive links like Wayback Machine references?

What the pages say

No page on the site addresses this.

The fix

Add a documentation page or FAQ entry showing an example Deep Research report with the actual citation format, including whether access dates, page titles, and archive links like Wayback Machine references are included.

Where we looked: /blog/deep-research-reports.md, /docs/resources/changelog.md, /task-api/examples/task-deep-research.md

confidence high · grounding world-knowledge · weight 0.00 · Vague

09We're building a real-time agent that occasionally spikes to 500 requests/second during user sessions. What's Parallel's burst capacity versus sustained rate limit, and is there a queue or do requests get rejected?limitsHedges+2 pts est.

We're building a real-time agent that occasionally spikes to 500 requests/second during user sessions. What's Parallel's burst capacity versus sustained rate limit, and is there a queue or do requests get rejected?

What the pages say

Parallel's default rate limits vary by product: Chat API is 300 requests per minute, while Tasks/TaskGroups API is 2,000 per minute. The site does not specify burst capacity separate from sustained rates, nor does it document whether excess requests are queued or rejected. For production capacity above default limits, Parallel directs users to contact support. Your 500 req/s (30,000 req/min) spike would exceed all default limits shown.

The fix

Add explicit documentation on burst capacity (if any), whether rate limits are enforced as hard caps with immediate rejection or soft limits with queuing/backoff, and recommended client-side handling patterns for 429 responses.

Where we looked: /ai/products/chat, /getting-started/rate-limits.md

confidence medium · grounding synthesized · weight 0.65 · Vague

10I'm comparing search APIs for my AI product and need predictable monthly costs. What specific mechanisms does Parallel use to prevent bill shock—do they offer hard caps, prepaid credits, or something else?pricingHedges+2 pts est.

I'm comparing search APIs for my AI product and need predictable monthly costs. What specific mechanisms does Parallel use to prevent bill shock—do they offer hard caps, prepaid credits, or something else?

What the pages say

Parallel uses per-request pricing rather than per-token pricing as their primary mechanism for cost predictability. Their Search API costs $5 per 1,000 requests with excerpts included, and they emphasize that 'you always know the exact cost of a query before you run it.' For the Task API, they offer 'fixed and predictable costs' with budget-aware execution that tracks cumulative cost across iterations and injects budget warnings when remaining spend drops below a threshold. However, the chunks do not explicitly mention hard caps, prepaid credits, or other specific bill-shock prevention mechanisms like spending limits or automatic cutoffs.

The fix

Add explicit documentation about whether customers can set hard spending caps, purchase prepaid credits, or configure automatic service suspension at billing thresholds. The current 'predictable cost' messaging is marketing language without concrete guardrails.

Where we looked: /pricing, /blog/deepsearchqa-taskapi-harness, /articles/openai-to-parallel-search-api

confidence medium · grounding synthesized · weight 0.65 · Vague

11We're migrating from Bing API v7.0. Does Parallel support all the same response fields—specifically computation, relatedSearches, and rankingResponse—or will we need to restructure our prompt engineering?migrationHedges+2 pts est.

We're migrating from Bing API v7.0. Does Parallel support all the same response fields—specifically computation, relatedSearches, and rankingResponse—or will we need to restructure our prompt engineering?

What the pages say

Parallel does not support the same response fields as Bing API v7.0. The chunks explicitly state that "The most significant difference lies in the output format. Bing returned brief snippets optimized for human readers. Parallel returns dense, token-efficient excerpts with enough context for AI agents to reason effectively." The Parallel Search API response schema shown in the OpenAPI docs includes: search_id, results (with url, title, publish_date, excerpts), and session_id. There is no mention of computation, relatedSearches, or rankingResponse fields from Bing's API. A migration guide is referenced at https://docs.parallel.ai/search/search-migration-guide but its contents are not provided in the chunks.

Source: /articles/bing-api-comparison.md, /articles/bing-api-comparison, /api-reference/search/search.md

confidence high · grounding synthesized · weight 0.70 · Answered

12For compliance reasons, I need to know how long Parallel retains task results and logs. Is there a fixed retention period, can I configure it per-task, or do I need to export everything immediately?operationsHedges+2 pts est.

For compliance reasons, I need to know how long Parallel retains task results and logs. Is there a fixed retention period, can I configure it per-task, or do I need to export everything immediately?

What the pages say

Parallel operates with **zero data retention** for request payloads and response data — they do not store the URLs you send or the data you extract after serving the response. However, the chunks do not specify: (1) whether there are any retention periods for task metadata, logs, or execution records; (2) whether retention is configurable per-task; or (3) whether users need to export anything immediately. The zero data retention claim applies to the actual request/response content, not necessarily to operational logs or task history.

The fix

Add explicit documentation on task result/log retention periods, whether retention is configurable, and any export requirements. The current 'zero data retention' messaging addresses payload data but leaves operational data retention ambiguous for compliance buyers.

Where we looked: /articles/ai-data-extraction-how-to-extract-structured-data-from-websites-at-scale

confidence medium · grounding stated · weight 0.75 · Vague

13Your authenticated page access feature sounds useful for our internal dashboards. What auth methods are supported—cookies, bearer tokens, OAuth 2.0, or custom headers—and can we rotate credentials without breaking active tasks?securityHedges+2 pts est.

Your authenticated page access feature sounds useful for our internal dashboards. What auth methods are supported—cookies, bearer tokens, OAuth 2.0, or custom headers—and can we rotate credentials without breaking active tasks?

What the pages say

For authenticated page access with the Parallel Task API, the chunks show support for Bearer tokens (via `Authorization: Bearer` header) and OAuth 2.0 (with PKCE). Custom headers are also supported for MCP server authentication (e.g., `x-api-key` or other arbitrary headers). Cookies are not mentioned as a supported method. However, there is no information about credential rotation or whether rotating credentials would break active tasks.

The fix

Add documentation on credential rotation procedures and whether active tasks maintain session continuity when API keys or OAuth tokens are rotated.

Where we looked: /integrations/mcp/programmatic-use, /integrations/mcp/programmatic-use.md, /task-api/mcp-tool-call.md, /docs/task-api/mcp-tool-call.md, /blog/mcp-tool-calling.md, /docs/integrations/browseruse.md, /integrations/browseruse.md, /docs/integrations/mcp/programmatic-use.md

confidence medium · grounding synthesized · weight 0.65 · Vague

14We're implementing the SSE streaming task manager from your cookbook. If the connection drops mid-stream, what's the reconnection behavior—does it resume from last event ID, restart the task, or require manual polling fallback?technicalContradicts itself+2 pts est.

We're implementing the SSE streaming task manager from your cookbook. If the connection drops mid-stream, what's the reconnection behavior—does it resume from last event ID, restart the task, or require manual polling fallback?

What the pages say

The chunks contain contradictory information about reconnection behavior. The Task SSE documentation (chunks [7] and [8]) states that "Event streams are not resumable - there are no sequence numbers or cursors to resume from a specific point" and that reconnecting gives you "the complete reasoning trace again plus current progress stats" for running tasks or "the same complete sequence as the first connection" for completed tasks. However, the API reference for Fetch Task Group Runs (chunks [5] and [9]) explicitly states "The stream is resumable using the `event_id` as the cursor" with a `last_event_id` parameter to resume from a specific point. The cookbook implementation (chunks [11], [12], [20]) describes "Stateless recovery: Each reconnection gets the complete current state" with exponential backoff and attempt limiting, but does not use event IDs for resumption. The FindAll API (chunk [19]) does document resumability with `last_event_id`. Given these contradictions between the Task SSE documentation and the Task Group Runs API reference, a buyer cannot determine whether the SSE streaming task manager supports resumable streams with `last_event_id` or requires stateless reconnection that restarts from the beginning.

The fix

Reconcile the Task SSE documentation with the API reference. Either update the Task SSE docs to document the last_event_id resumption capability, or remove the resumable stream claims from the API reference if they don't apply to task SSE streams. Clarify which endpoint(s) support resumable streams and which use stateless reconnection.

Where we looked: /task-api/task-sse.md, /api-reference/tasks/fetch-task-group-runs.md, /docs/api-reference/tasks-beta/fetch-task-group-runs.md, /blog/cookbook-sse-task-manager-with-parallel.md, /blog/cookbook-sse-task-manager-with-parallel, /blog/cookbook-sse-task-manager-with-parallel.md, /docs/findall-api/features/findall-sse.md

confidence medium · grounding synthesized · weight 0.65 · Contradictory

What the pages already answer cleanly

1 of 15. Answered with enough specificity for a buyer to act on. This is the bar the queue above is chasing.

technicalAnswered

You cite SealQA benchmarks for price-performance, but I need to verify this for my procurement team. What exact metric is 'price-performance'—cost per query, cost per accurate answer, or something else—and what's the comparison baseline?

What the pages say

Parallel's 'price-performance' metric on SealQA is defined as accuracy (%) per cost per thousand queries (CPM in USD). Specifically, they measure 'cost per thousand queries (CPM) as measured on the benchmark' against 'factual accuracy against verified ground truth' evaluated via LLM-as-a-judge. The comparison baseline includes Perplexity DR, Exa Research Pro, and GPT-5. For example, on SEAL-0: Parallel Pro achieves 52.3% accuracy at $100 CPM vs. Perplexity DR at 38.7% accuracy for $1,258.20 CPM; on SEAL-HARD: Parallel Pro achieves 66.9% accuracy at $100 CPM vs. Exa Research Pro at 59.1% accuracy for $2,192.40 CPM. The methodology explicitly states: 'For token-based APIs, we normalized to cost per thousand queries (CPM) as measured on the benchmark.'

Source: /blog/benchmarks-task-api-sealqa.md, /blog/benchmarks-task-api-sealqa.md, /blog/benchmarks-task-api-sealqa.md, /blog/benchmarks-task-api-sealqa, /blog/benchmarks-task-api-sealqa, /blog/benchmarks-task-api-sealqa

confidence high · grounding synthesized · weight 0.95 · Answered

Protocol plumbing · 69/10011 of 16 checks pass · each fix +6 protocol pts est.

The other half of the score: 16 checks for the files and headers agents look for. The 5 below are installs, not judgment calls, and most are an afternoon. Expand any for the snippet and the standard it follows. They sit after the queue because none of them changes what your pages say.

Content signalAccess+6 pts est.
Install snippet
User-agent: *
Content-Signal: search=yes, ai-input=yes, ai-train=no
Allow: /

StandardCloudflare proposalVendor proposal

Clean crawlAccess+6 pts est.

StandardSitedex metricSitedex metric

MCP cardInteraction+6 pts est.

Sitedex generates this file from your crawl. Grab it in Files from this audit below.

StandardModel Context ProtocolCommunity spec

WebMCP widgetInteraction+6 pts est.

Sitedex generates this file from your crawl. Grab it in Files from this audit below.

StandardW3C WebMCP draftW3C / WHATWG

Canonical URLsHygiene+6 pts est.
Install snippet
<link rel="canonical" href="https://parallel.ai/" />

StandardRFC 6596IETF RFC

Already passing 11 of 16: robots.txt, sitemap.xml, llms.txt, AI crawler access, Markdown negotiation, Server-rendered content, OpenAPI spec, Meta descriptions, HTML lang attribute, Organization schema, Sitemap lastmod.

Ask this site’s index

Sitedex already serves parallel.ai as an MCP endpoint. Ask parallel.ai anything an AI agent might ask, and see what its index returns. (To score your own site, use the form below.)

Snippets & configs

For developers and the engineer-on-call: copy these into your tools or your site.

Files from this audit

Built from this crawl. Download or copy each, then install it at the path noted.

llms.txt

Built from this crawl. Install at /llms.txt so agents start here.

organization.json

Organization JSON-LD, pre-filled from this crawl. Wrap in a ld+json script.

server-card.json

MCP server card built from this crawl. Host at /.well-known/mcp/server-card.json.

webmcp.json

WebMCP discovery manifest built from this crawl. Host at /.well-known/webmcp.json.

MCP endpoint

https://mcp.sitedex.dev/s/parallel-ai/mcp

The URL anyone's agent points at. Read-only; safe to share.

Claude Code

claude mcp add parallel --transport http https://mcp.sitedex.dev/s/parallel-ai/mcp

One command, then the agent has it.

Cursor / Continue

{
  "mcpServers": {
    "parallel": {
      "url": "https://mcp.sitedex.dev/s/parallel-ai/mcp"
    }
  }
}

Drop into mcp.json.

WebMCP: two parts

WebMCP-capable browsers run the widget at runtime. Crawlers without JS rendering need the discovery manifest to find your tool surface. Install both.

1 · Widget script

<script async src="https://sitedex.dev/widget.js"></script>

Drop in <head>. WebMCP-capable browsers (Chrome 146+ Origin Trial) call navigator.modelContext.provideContext() via this script.

2 · Discovery manifest

{
  "$schema": "https://wellknownmcp.org/schemas/webmcp.json",
  "name": "parallel.ai",
  "tools": [
    { "name": "search", "description": "Search parallel.ai's indexed content." },
    { "name": "get_page", "description": "Fetch a page from parallel.ai as markdown." }
  ]
}

Host alongside the script at /.well-known/webmcp.json. Crawlers that don't render JS rely on this.

Your turn

See which of these questions your site goes silent on.

Free, about 5 minutes. We crawl your site, test it against the buyer questions your category asks, and name what’s vague, contradictory, or missing, plus the files AI agents look for.

ComingEmbeddable grade badgeScore history and deltasOpt-in public board