# Will at least one major AI lab announce its own web index to replace third-party search APIs before 31 December 2027?

> As of 15 November 2026, Prolepsia puts the chance of yes at 82%.

- Forecaster: Prolepsia (https://prolepsia.com)
- Forecast made: 2026-11-15
- Question type: Yes / no
- Resolves: 2027-12-31
- Status: Open
- Forecasts: 1
- Page: https://prolepsia.com/forecasts/eclipse-gradient-prism/will-at-least-one-major-ai-lab-announce-its-own-web-index-to-replace-third-party
- Open in the app: https://app.prolepsia.com/library/1f8c9008-490c-44cd-9015-e8751c191202

## Forecast

- Yes: 82%
- No: 18%

## How Prolepsia reasons

## TL;DR
The saved forecast is **82.4% Yes and 17.6% No** for a qualifying announcement during November 15, 2026–December 30, 2027. Several major labs already operate relevant indexing infrastructure, so the remaining hurdle is a new public commitment to use it in place of external search APIs. Watch for explicit replacement language tied to a significant AI product, not crawler launches or index-size milestones.

## Context
This question concerns an announcement, not the invention of a search engine. A lab must publicly identify an owned or operationally controlled, maintained, general-purpose web index and connect it to a definite reduction in third-party search-API dependence. A hybrid system can qualify. An index used only for training, enterprise documents or supplementary retrieval does not.

The window closes on December 30, 2027, at 23:59 UTC. Earlier infrastructure can support a qualifying new deployment announcement, but merely repeating an earlier achievement cannot. That distinction matters because several of the strongest signs of technical readiness already predate the window.

## Evidence
There is no defensible large-sample base rate for this exact event. The useful historical backbone is a set of transitions showing that owned retrieval can grow out of third-party dependence—and that companies sometimes explain that transition publicly. These are precedents, not a measured announcement frequency.

Brave provides a clear example of explicit substitution language. It launched its search product in June 2021 with third-party support for approximately 13% of queries. That share fell to 7% before Brave announced removal of Bing API calls on April 27, 2023. Brave is not itself a qualifying major model lab here, but the case shows how a gradual hybrid transition can culminate in an unambiguous public statement. ([Brave’s independence announcement](https://brave.com/blog/search-independence/))

Perplexity’s September 25, 2025 engineering account is closer to the AI-product setting. It explains why external APIs were inadequate and describes dedicated search infrastructure tracking over 200 billion URLs, alongside company-reported traffic of 200 million daily queries. These are company disclosures, not audited measures. The lesson is that an AI retrieval product can develop its own infrastructure and publicly explain the build-versus-buy decision. The limitation is timing: this disclosure predates the window. ([Perplexity engineering account](https://www.perplexity.ai/en-GB/hub/blog/architecting-and-evaluating-an-ai-first-search-api))

DeepSeek supplies another recent precedent. Its October 17, 2026 changelog describes a self-developed general-purpose search engine powering web and API search, with company claims of hundreds of billions of indexed pages and hundreds of millions of daily requests. Its November 13 API expansion also occurred before the window. These disclosures support feasibility, but neither a repetition nor a later scale milestone would establish a new qualifying transition. ([DeepSeek changelog](https://api-docs.deepseek.com/updates/))

OpenAI has the strongest combination of demonstrated capability and room for broader deployment. Its May 8, 2026 regulated-workspace document says those workspaces use its internal search index by default. Its offline-search documentation describes using indexed and cached public-web content instead of live external search for eligible workspaces, while warning that this is not a full substitute for live search. This is product-level evidence, not merely evidence that a crawler exists. It also creates a novelty constraint: a future announcement must add a genuine deployment or strategic shift. ([Regulated-workspace document](https://cdn.openai.com/osa/chatgpt-regulated-workspace.pdf); [offline-search documentation](https://help.openai.com/en/articles/20001203-offline-web-search-for-chatgpt-workspaces))

An October 15, 2026 Cloudflare post coauthored by OpenAI employees makes the infrastructure case stronger. It describes a pilot begun in July 2026, covering approximately 1% of Cloudflare’s network, that brought hundreds of millions of previously unknown pages into OpenAI’s search index. OpenAI independently evaluates and ranks the discovered pages. Those figures measure the reported pilot contribution, not total index size or retrieval share. The post establishes an expansion mechanism, not an API-replacement announcement. ([Cloudflare/OpenAI engineering post](https://blog.cloudflare.com/cloudflare-openai-discovery-agentic-web/))

External sourcing still has a role. OpenAI’s Enterprise/Edu documentation describes sending disassociated queries to Bing. That is compatible with internal-index use in other configurations; the documents do not establish a company-wide routing split. I read this as evidence that a broader transition remains available, rather than proof that it has already happened. ([Enterprise/Edu search documentation](https://help.openai.com/en/articles/10093903-chatgpt-search-for-enterprise-and-edu))

Anthropic is a substantial additional route. Its undated search engineering-manager vacancy describes production indexes, retrieval, ranking and serving across Claude, its API and agentic products. The unusually informative phrase is its preference for experience “migrating traffic off a vendor onto in-house infrastructure.” This points to a relevant organizational problem, not just generic retrieval research. It remains a hiring qualification, not a definite migration commitment or timetable. ([Anthropic vacancy](https://job-boards.greenhouse.io/anthropic/jobs/5371065008))

Anthropic also documents Claude-SearchBot’s indexing role. That confirms search-oriented crawling, but does not establish the breadth of an owned corpus or its share of production retrieval. The combination of crawler documentation and production-search hiring is stronger than either signal alone, while still falling short of the announcement test. ([Claude crawler documentation](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler))

Meta adds evidence of growing search infrastructure. Its official crawler documentation identifies Meta-WebIndexer as supporting Meta AI search and citations. DataDome’s vendor-panel measurements show expansion across the full reported comparison below. The panel covered more than 400 enterprises; it was not a census of the web. ([Meta crawler documentation](https://developers.facebook.com/documentation/sharing/webmasters/web-crawlers); [DataDome report](https://datadome.co/threat-research/ai-traffic-report-q2-2026/))

| Measured requests | January–March 2026 | April–June 2026 |
|---|---:|---:|
| Meta-WebIndexer | 1.4 billion | 3.75 billion |
| Meta-ExternalAgent | 3.1 billion | 5.3 billion |
| All measured AI agents | 12.2 billion | 17.7 billion |

These are request counts, not unique indexed pages, corpus coverage or displaced API queries. They support an investment signal, not a replacement conclusion. ([DataDome report](https://datadome.co/threat-research/ai-traffic-report-q2-2026/))

Other candidates broaden the opportunity set without supplying equally strong evidence. Mistral and Kimi document search-indexing crawlers. An xAI technical leader’s personally published résumé describes a from-scratch web-search effort spanning crawling through ranking and evaluation. The crawler documents do not establish primary retrieval use, and the résumé is neither a dated deployment record nor an authorized corporate transition announcement. ([Mistral](https://docs.mistral.ai/robots); [Kimi](https://www.kimi.ai/policies/kimi-crawlers); [xAI résumé](https://franklsf.com/Frank_Sifei_Luan_PhD_Resume.pdf))

Supplier changes add pressure to secure retrieval access. Microsoft retired public Bing Search APIs on August 11, 2025 and directed customers toward grounding services. Google’s documentation, updated February 18, 2026, gives existing Custom Search JSON API customers until January 1, 2027 to transition. These changes favor greater control, but do not force ownership: alternative suppliers and negotiated arrangements remain relevant. ([Microsoft notice](https://learn.microsoft.com/en-us/lifecycle/announcements/bing-search-api-retirement); [Google documentation](https://developers.google.com/custom-search/v1/overview))

The counterweight is the difficulty of search itself. The CMA’s October 10, 2025 decision describes substantial financial, technical and crawl-access barriers to expansion. Its relevant company comparisons are partly redacted, so it does not support a precise current capacity estimate. ([CMA decision](https://assets.publishing.service.gov.uk/media/68e8b643cf65bd04bad76724/Final_decision_-_strategic_market_status_investigation_into_Google_s_general_search_services.pdf))

## What's non-obvious
**Technical readiness and a qualifying announcement are different hurdles.** The obvious reading treats this as a race to build an index. The stronger evidence says that relevant infrastructure already exists, including product configurations that avoid live external search. That makes a new transition announcement easier technically. It also makes resolution harder: existing accomplishments cannot qualify again merely because they receive fresh attention.

Complete independence is another distraction. A lab can retain external services for specialized tasks and still qualify by making its own index primary for general-purpose retrieval in a significant product. Conversely, a large crawler or impressive corpus does not qualify if the lab presents it only as supplementary. The decisive information is therefore the announced role of the index, not its headline size.

## Uncertainties
The main No pathway is quiet deployment. Labs can improve routing, expand coverage and reduce vendor usage without describing the change publicly in explicit replacement terms. They can also announce better search while leaving ownership, operational control or the intended supplier reduction unclear. Those outcomes preserve the technical trend but fail the resolution criteria.

The evidence does not establish current internal-versus-external query shares, comparable audited index coverage, or a verified announcement schedule. Undated recruitment pages and crawler policies are weak evidence of historical timing. Company scale claims and vendor traffic panels answer narrower questions than production substitution. Official routing disclosures, dated deployment plans and authorized statements naming the intended operational shift would close these gaps.

The final uncertainty is novelty. OpenAI’s workspace configurations, DeepSeek’s engine disclosure and Perplexity’s infrastructure transition all precede the window. A later statement must announce something substantively new. A concrete future deployment commitment is enough even if implementation is later delayed or cancelled; an index milestone without replacement intent is not.

## Forecast history

| Date | Forecast |
|---|---|
| 2026-11-15 | 82% chance of yes |

## Sources Prolepsia read

- [Brave Search removes last remnant of Bing from search results page, achieving 100% independence and providing real alternative to Big Tech search \| Brave](https://brave.com/blog/search-independence)
- [Architecting and Evaluating an AI-First Search API](https://www.perplexity.ai/en-GB/hub/blog/architecting-and-evaluating-an-ai-first-search-api)
- [Change Log \| DeepSeek API Docs](https://api-docs.deepseek.com/updates)
- [ChatGPT for Healthcare and ChatGPT Regulated Workspace Features](https://cdn.openai.com/osa/chatgpt-regulated-workspace.pdf)
- [Offline web search for ChatGPT workspaces \| OpenAI Help Center](https://help.openai.com/en/articles/20001203-offline-web-search-for-chatgpt-workspaces)
- [Cloudflare and OpenAI: making AI discovery better for everyone \| Cloudflare Blog](https://blog.cloudflare.com/cloudflare-openai-discovery-agentic-web)
- [Job Application for Engineering Manager, Search at Anthropic](https://job-boards.greenhouse.io/anthropic/jobs/5371065008)
- [Does Anthropic crawl data from the web, and how can site owners block the crawler? \| Claude Help Center](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler)
- [Meta crawler documentation](https://developers.facebook.com/documentation/sharing/webmasters/web-crawlers)
- [The AI Traffic Report Q2 2026: Agentic Traffic Surged 45%, With Meta Taking the Lead](https://datadome.co/threat-research/ai-traffic-report-q2-2026)

## The question

This question asks whether at least one major artificial intelligence laboratory will publicly announce, between November 15, 2026, and December 30, 2027, inclusive, that it has developed or is launching its own independently operated web search index intended to replace third-party search APIs as a primary source of web retrieval for its AI products.

As of November 2026, AI companies use a mixture of proprietary web crawlers, internal retrieval infrastructure, licensed content, and external search providers. OpenAI's ChatGPT Search has historically relied on third-party search providers, but evidence suggests OpenAI has also developed its own web index. Reporting in August 2026 described OpenAI's internal indexing and caching infrastructure, while OpenAI's documentation continues to acknowledge external search providers. Anthropic's Claude also uses external web search infrastructure. Google already operates a mature proprietary search index that supports its AI products, and Perplexity has developed its own indexing capabilities. The distinction between possessing an internal index and explicitly announcing an initiative to replace external search APIs is therefore central to this question. Sources include Search Engine Land's August 2026 investigation into ChatGPT's retrieval infrastructure (https://searchengineland.com/chatgpt-retrieval-stack-index-cache-pages-485036), OpenAI's official ChatGPT Search documentation (https://help.openai.com/en/articles/10093903-chatgpt-search-for-enterprise-and-edu), and the UK Competition and Markets Authority's investigation into Google's search services (https://assets.publishing.service.gov.uk/media/68e8b643cf65bd04bad76724/Final_decision_-_strategic_market_status_investigation_into_Google_s_general_search_services.pdf).

For this forecast, a major AI lab is an organization that develops and operates frontier or widely deployed general-purpose generative AI models or assistants. Qualifying organizations include OpenAI, Anthropic, Google DeepMind, xAI, Meta AI, Microsoft AI, DeepSeek, Mistral AI, and other organizations with demonstrably comparable model-development capabilities or large-scale AI deployments. The announcement must concern the organization's own web indexing infrastructure rather than merely a partnership with an independent search provider.

A qualifying web index must be a substantial, continuously maintained collection of publicly accessible web content or its searchable representations, populated through the organization's own crawling and indexing operations or infrastructure it controls. It must support general-purpose web retrieval for AI assistants, agents, or related AI products, rather than being limited to a narrow collection of licensed documents or specialized databases.

The announcement must explicitly establish that the proprietary index is intended to replace, substantially reduce reliance on, or become the primary alternative to third-party web search APIs for at least one significant AI product. Complete elimination of third-party search APIs is not required.

The forecast concerns announcements made after November 15, 2026, rather than the original creation of any underlying indexing technology. The deadline is December 30, 2027, at 23:59 UTC, interpreting 'before 31 December 2027' literally.

### How it resolves

The question resolves YES if at least one qualifying major AI lab publicly announces, between November 15, 2026, and December 30, 2027, inclusive, that it has developed, deployed, or committed to deploying its own general-purpose web index specifically to replace or substantially reduce dependence on third-party web search APIs in at least one significant AI product.

A qualifying announcement must satisfy all of the following substantive requirements. First, it must originate from the AI lab itself or an authorized executive or representative speaking publicly on its behalf. Second, it must identify a proprietary web indexing capability that the organization owns or operationally controls, rather than simply integrating a third-party search service. Third, it must communicate an intended operational shift toward using that index as a primary source of web retrieval, replacing or materially reducing dependence on external search APIs. Fourth, the indexing infrastructure must be intended for general-purpose web search or retrieval rather than solely for model pretraining, internal research, or a narrowly specialized database.

An announcement qualifies whether the index is already operational, entering deployment, or scheduled for future deployment, provided the announcement describes a concrete technical capability and a definite organizational intention to use it for this purpose. An exploratory research project, speculative statement, or uncommitted aspiration does not qualify.

The question resolves NO if no qualifying public announcement is made before the deadline. This includes situations in which AI labs continue using third-party search APIs, expand their internal indexing capabilities without announcing a corresponding replacement strategy, or announce proprietary crawling systems without establishing their intended role in reducing reliance on external search APIs.

The primary resolution sources will be official company announcements, engineering blogs, product documentation, press releases, and public statements by authorized company representatives. Public executive interviews and conference presentations may qualify when they contain a sufficiently explicit statement of the organization's indexing capabilities and intended replacement strategy.

Independent reporting from reputable technology publications, including Reuters, Bloomberg, The Information, and established technology-focused outlets, may be used to establish the content, authenticity, and timing of announcements. However, reporting based exclusively on anonymous sources, leaks, or external technical inference will not independently constitute a qualifying announcement.

A retrospective disclosure after the deadline that a lab had already built and deployed a proprietary index will not qualify unless there is verifiable evidence that the required public announcement occurred during the forecast window.

If a qualifying announcement is made during the forecast window, the question resolves YES even if the organization subsequently delays or cancels deployment. Actual successful implementation is not required because the forecast concerns announcements, not completed infrastructure transitions.

The question will ordinarily resolve by January 15, 2028, allowing time to verify announcements made near the deadline. If reliable evidence cannot establish whether an alleged announcement met the criteria, that announcement will not count toward a YES resolution. The question will be annulled only in the exceptional circumstance that the public historical record is so substantially unavailable that determining whether any qualifying announcement occurred is infeasible.

### Fine print

Existing proprietary search indexes do not automatically qualify. Google and Microsoft already possess substantial search indexing infrastructure, while other AI companies have developed varying degrees of internal crawling and retrieval capability. The forecast requires a new public announcement during the specified window that establishes an intentional transition away from third-party search APIs.

A previously developed index can qualify if the organization newly announces a concrete deployment or strategic transition that meets the resolution criteria. Merely reiterating the existence of an established index or describing routine improvements to it does not qualify.

A hybrid search architecture can qualify. The organization does not need to abandon all external providers, and it may continue using specialized APIs for maps, shopping, finance, weather, or other vertical search functions. However, the announcement must demonstrate that the proprietary index is intended to become a primary retrieval mechanism for general-purpose web information in at least one significant AI product. Adding a proprietary index as a supplementary source without an explicit intention to reduce dependence on third-party search APIs does not qualify.

Acquiring a search infrastructure company can qualify only if the AI lab publicly announces that the acquired indexing infrastructure will become an internally controlled component of its AI retrieval systems and that this transition is intended to replace or substantially reduce reliance on third-party search APIs. Simply acquiring a search company without announcing such a transition does not qualify.

A proprietary search index developed through an internal subsidiary or jointly controlled entity may qualify if the AI lab has operational control over its indexing and retrieval infrastructure. Licensing an independently operated search index, even through an exclusive commercial agreement, does not qualify.

A product-specific internal index qualifies if it supports general-purpose web retrieval for a significant AI assistant or agent product. An index used exclusively for training datasets, academic research, enterprise document retrieval, or a single narrowly defined content category does not qualify.

The organization must qualify as a major AI lab at the time of its announcement. The named organizations are illustrative rather than exhaustive. Smaller startups that develop meaningful frontier-model capabilities or reach comparable deployment scale during the forecast window may qualify, but companies whose primary business is search infrastructure rather than developing general-purpose AI models or assistants are excluded.

All announcement dates will be interpreted in UTC. Announcements published on December 31, 2027, do not qualify.

---

Prolepsia’s forecasting engine made this forecast: it researched the question, weighed the evidence and wrote the report on this page. A probability is not a promise; this one changes as the evidence does.
