Grounding is the practice of connecting a generative model's output to retrieved source material, so that an answer rests on documents the system can point to rather than on the statistical residue of training. A grounded response carries provenance: the passages that informed it, the queries that fetched them, and the position in the answer each source supports. The technique exists because large language models compress knowledge into weights at training time, leaving it static, undated and impossible to audit.

Google describes it as the process of enabling large language models to leverage verifiable information sources, either from Google or customer data, in documentation accompanying its Maps grounding release. Microsoft goes further in scope. Jordi Ribas, corporate vice president of search and AI, called grounding "the connective tissue between generative models and the world's information" in a February 2026 post that positioned the technology as infrastructure beneath nearly every major AI assistant on the market.

How a grounded request runs

The Gemini API documents a five-step flow. An application sends a prompt with the google_search tool enabled. The model analyses whether a search would improve the answer. If so, it composes one or more queries and executes them, then synthesises the results into a response.

What comes back is structured. Current Gemini responses expose a sequence of steps: a thought summary, a google_search_call carrying the queries the model ran, a google_search_result containing an HTML snippet for rendering search suggestions, and a model_output block whose text carries inline url_citation annotations. Each annotation records a source URL, a title and a start_index and end_index marking the exact span of generated text it supports. Older Gemini responses used a groundingMetadata object with the same idea expressed differently: groundingChunks for the sources, groundingSupports mapping text segments to chunk indices, webSearchQueries for the executed searches, and searchEntryPoint for the Search Suggestions chip.

Not every prompt needs retrieval, and retrieval is neither free nor fast. Google's earlier dynamic retrieval configuration made that judgement explicit, scoring each prompt between 0 and 1 for how much it would benefit from fresh web data, then comparing that score against a developer-set threshold defaulted to 0.3. A cryptocurrency price scores high; the chemical formula of water scores low. Newer Gemini models fold the decision into the model itself.

Microsoft's Azure implementation splits the work differently. An agent decides whether to invoke Grounding with Bing Search, the service returns relevant chunks, and the model composes an answer from them. Developers and end users never see the raw content, only citations and a link to the Bing query. Result count defaults to five and caps at fifty, alongside freshness, market and language filters. Microsoft states that grounding queries leave the Azure compliance boundary.

Where the evidence comes from

Public web search is the visible case, but it is one option among several. Google's platform lists grounding against Google Search, Google Maps, its own enterprise search product, a configurable RAG engine, Elasticsearch indexes, third-party search APIs, and Web Grounding for Enterprise, a narrower index for regulated industries where fast-changing content refreshes every six hours and the whole index every twenty-four. That variant logs no customer data, trading breadth for compliance.

Private-data grounding is the enterprise workhorse: documents, product catalogues and support archives are indexed, then queried at inference. Google introduced a dedicated crawler, Google-CloudVertexBot, in August 2024 to fetch sites at their owners' request during agent development. Licensed datasets form a third category. On June 27, 2024 Google Cloud vice president Burak Gokturk announced grounding against third-party data with Moody's, MSCI, Thomson Reuters and ZoomInfo as launch providers.

Origins and evolution

The architecture predates the branding. Patrick Lewis and eleven co-authors at Facebook AI Research, University College London and New York University published "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" on May 22, 2020, pairing a parametric memory with a dense index over 21 million Wikipedia passages. The paper named the three problems grounding still addresses: knowledge cannot be updated without retraining, decisions lack provenance, and models hallucinate.

Commercial grounding arrived in 2024. Google made Grounding with Google Search generally available on Vertex AI that June, alongside an experimental high-fidelity mode aimed at further reducing hallucinations. On October 31, 2024 the capability reached the Gemini API and Google AI Studio at $35 per 1,000 grounded queries, through a google_search_retrieval tool on Gemini 1.5 models. European availability followed on December 5, 2024.

Microsoft integrated Grounding with Bing Search into its Azure AI Agent Service on January 31, 2025. Four months later it announced the retirement of the traditional Bing Search and Bing Custom Search APIs, effective August 11, 2025, naming grounding as the recommended replacement at prices PPC Land calculated as 40 to 483 percent higher depending on tier. The retirement compressed developer access to Microsoft's index into a single grounding product.

That consolidation created room for challengers. Brave launched AI Grounding on August 5, 2025, reporting 94.1 percent on the SimpleQA benchmark and positioning an independent index against the incumbents. Yahoo built its Scout assistant on Microsoft's Bing grounding API in January 2026. Google renamed Vertex AI as the Gemini Enterprise Agent Platform on April 22, 2026, folding grounding into a broader agent stack.

What it costs

Grounding is billed separately from tokens, which surprises teams budgeting on model pricing alone. Google's original structure charged per grounded prompt; Gemini 3 models bill per query, so a prompt answered with two searches counts twice. Brave prices its search API from nothing for 2,000 monthly queries up to tiers starting at $5 CPM.

Few organisations operate a general web index at scale, so grounding revenue accrues to Google, Microsoft and a small independent tier. France's competition authority examined that concentration when Qwant filed a complaint against Microsoft, rejecting it on November 27, 2025 on the grounds that dominance in search syndication had not been demonstrated.

Why marketers care

Grounding determines which pages get read when an assistant composes an answer, which makes it the retrieval layer beneath every AI visibility metric now being sold. Microsoft frames the resulting discipline as generative engine optimisation and has begun exposing the data: Bing Webmaster Tools gained an AI Performance dashboard on February 10, 2026, reporting citation frequency and the grounding queries that retrieved a page. Google's equivalent generative AI performance report reached every website on August 31, 2026, showing impressions inside AI Overviews and AI Mode while withholding clicks and query strings.

Exclusion is the other half. Google's robots.txt token Google-Extended keeps a page out of Gemini grounding, and Microsoft's data-nosnippet attribute, introduced in October 2025, excludes marked sections from snippets and AI answers while leaving indexing intact. Regulation has been more forceful. On June 3, 2026 the UK Competition and Markets Authority imposed its first Publisher Conduct Requirement on Google, obliging the company to let publishers withhold content from AI Overviews, AI Mode and Discover AI features either as links or as grounding material, without any loss of conventional ranking. Google launched a domain-level toggle the same day. Page-level grounding controls are due on March 3, 2027, after Google told the regulator the finer-grained version required substantial engineering work.

Limits and disputes

Grounding reduces fabrication; it does not eliminate it. Research published by the BBC on February 11, 2025 found that 51 percent of AI responses to news questions carried significant issues, that 19 percent of answers citing BBC material introduced factual errors, and that 13 percent of quotes attributed to BBC articles had been altered or invented. The BBC had lifted its access restrictions for the test, so the assistants were reading the source directly. A larger European Broadcasting Union and BBC study published in October 2025, covering roughly 3,000 responses across 14 languages, found 45 percent of answers contained at least one significant issue and 31 percent had sourcing problems. Press accounts of that study differ on the sourcing error rate attributed to Gemini, variously reported as 72 and 76 percent.

Three structural limits sit behind those numbers. A grounded answer inherits whatever the index contains, so retrieval errors and stale caches propagate silently. Attribution is generated rather than verified, so a citation can be attached to a claim the cited page does not make. And accuracy claims are largely vendor-supplied: PPC Land coverage has recorded Brave's SimpleQA score as both 94.1 and 94.9 percent.

Disambiguation

Retrieval-augmented generation is the architecture; grounding is the product category built on it. Vendors reserve "grounding" for managed services, including public web search, and use RAG for private document pipelines.

Fine-tuning alters model weights through additional training. Grounding leaves weights untouched, which is why grounded facts can change hourly while a fine-tuned model cannot.

Citation is a display convention; grounding is the retrieval that should precede it. A system can print sources it did not read, which is the failure mode the BBC research measured.

Crawling and indexing build the corpus. Grounding queries it. Blocking a crawler removes a page from grounding; blocking grounding does not stop the crawl.

Recent developments

Microsoft introduced Web IQ, a suite of AI-native grounding APIs, at BUILD 2026 in early June, announced by Knut Risvik on the Bing Search Blog. Coverage of the launch records a claimed 164 millisecond P95 latency against 406 to 2,090 milliseconds for competing services, figures supplied by Microsoft. The company reintroduced Web IQ on June 17, 2026 ahead of Cannes Lions, bundled with citation reporting in Microsoft Clarity and an advertising MCP server, naming Nasdaq as an early deployment. Google's surface widened in parallel, adding Maps as a source, external providers as grounding backends, and combinations of search grounding with URL-context and function-calling tools in a single request.

Timeline

  • May 22, 2020: Retrieval-augmented generation paper published by Lewis and co-authors
  • April 2024: Vertex AI Agent Builder introduced with grounding components
  • June 27, 2024: Grounding with Google Search reaches general availability; third-party datasets announced with Moody's, MSCI, Thomson Reuters and ZoomInfo
  • August 2024: Google-CloudVertexBot documented as a crawler for agent development
  • October 31, 2024: Grounding with Google Search launches in the Gemini API at $35 per 1,000 grounded queries
  • December 5, 2024: Gemini API grounding becomes available across Europe
  • January 31, 2025: Microsoft integrates Grounding with Bing Search into Azure AI Agent Service
  • February 11, 2025: BBC publishes error-rate research on AI assistants handling news
  • May 12, 2025: Microsoft announces retirement of the Bing Search APIs
  • August 5, 2025: Brave launches AI Grounding
  • August 11, 2025: Bing Search APIs shut down; grounding becomes the recommended path
  • September 26, 2025: Grounding with Google Maps reaches general availability
  • September 30, 2025: CMA designates Google with Strategic Market Status in UK general search
  • October 2025: European Broadcasting Union and BBC publish multi-language study on AI news accuracy
  • October 15, 2025: Microsoft introduces the data-nosnippet attribute for Bing and Copilot
  • November 27, 2025: France's competition authority rejects Qwant's complaint against Microsoft
  • January 27, 2026: Yahoo launches Scout on Microsoft's Bing grounding API
  • February 10, 2026: Bing Webmaster Tools opens its AI Performance dashboard in public preview
  • February 12, 2026: Microsoft publishes its first comprehensive account of grounding infrastructure
  • April 22, 2026: Vertex AI is renamed the Gemini Enterprise Agent Platform
  • June 2, 2026: Microsoft launches Web IQ grounding APIs at BUILD 2026
  • June 3, 2026: CMA imposes its Publisher Conduct Requirement; Google launches a Search Console AI opt-out toggle
  • August 31, 2026: Google completes the global rollout of generative AI performance reports
  • December 3, 2026: Substantive obligations of the CMA requirement take legal force
  • March 3, 2027: Page-level grounding controls scheduled to arrive

Summary

Who: Google, Microsoft and a smaller independent tier including Brave operate the grounding services that most AI assistants and agent platforms depend on. Publishers, SEO practitioners and enterprise development teams sit on the other side, controlling what is available to retrieve and measuring what gets cited.

What: A method of tying generative output to retrieved source documents at inference time, implemented as a tool call that returns cited passages, executed queries and character-level mappings between answer text and sources. Evidence can come from the public web, a private corpus, licensed datasets, geospatial data or a compliance-restricted index.

When: Founded on retrieval-augmented generation research published in May 2020, commercialised through Google Cloud in 2024 and the Gemini API on October 31, 2024, extended to Azure agents in January 2025, and consolidated across 2025 and 2026 as Microsoft retired its search APIs and launched Web IQ.

Where: Delivered through the Gemini API, the Gemini Enterprise Agent Platform, Azure AI Foundry and independent search APIs, with publisher-facing controls in robots.txt, HTML attributes, Search Console and Bing Webmaster Tools.

Why: Model weights hold static, unattributable knowledge, and no commercial deployment can rest on that alone. Grounding supplies freshness and provenance, and in doing so decides which pages participate in AI answers, which is now a regulated question in the United Kingdom and a contested one everywhere else.