The grounding clause

What Google's terms say

Grounding with Google Search is the Gemini API feature that answers from live web results and returns the sources it used. It is the obvious instrument for measuring what Gemini cites. The Gemini API Additional Terms of Service, effective 23 March 2026, close that door twice.

First, on who is allowed to see the results:

You will only use Grounding with Google Search in an application that is owned and operated by you and will only display the Grounded Results with the associated Search Suggestion(s) to the end user who submitted the prompt.

A public evidence base is, by construction, read by people who did not submit the prompt. Then, on what may be done with the links themselves:

You will not, and will not allow your end user or any third party to, cache, frame, syndicate, resell, analyze, train on, or otherwise learn from Grounded Results or Search Suggestions. For clarity, Grounded Results, Search Suggestions, and Links are intended to be used in combination to respond to a given end user prompt and it is a violation of these terms to use Grounding with Google Search to extract or collect one or more of these components for another purpose (for example, using programmatic or automated means to collect Links, using Links to build an index, or using Links to identify destination pages for crawling or scraping).

Collecting links by automated means. Building an index from them. Analysing them. Each is named separately. That is not an accidental overlap with what a citation dataset does. It is a description of one.

One caveat, stated because the clause is not absolute and we would rather give you the whole thing than the convenient half of it. Google's terms carry limited storage exceptions: a developer may retain Grounded Result text for up to two years for display optimisation, for chat history, or temporarily in order to resubmit it for refined results. None of those exceptions permits publishing an index built from the Links, which is the specific act named above and the specific act a public citation dataset requires.

Source: ai.google.dev/gemini-api/terms, effective 23 March 2026, read 10 July 2026.

What Microsoft's terms say

Microsoft's equivalent is Grounding with Bing Search, the route to Copilot-style answers with citations. Its terms, last updated November 2025, are shorter and blunter.

Copy, store, cache, archive, or create a database of Output.

That is Section 4(b)(o), in the list of things you will not do. Section 4(b)(g) adds:

Use the Services to create a database or service that competes with Output received when using the Services.

There is a display obligation too, which matters to anyone planning to redraw citations as a chart: "You must retain the links and attributions to References in the Output without modification, including without limitation retaining the location, description, and linking of the Resources."

Source: microsoft.com/en-us/bing/apis/grounding-legal, last updated November 2025, read 10 July 2026.

Which engines are left

"No prohibition found" describes the public terms we read on 10 July 2026. It is not legal advice and not a warranty that no such clause exists elsewhere in a provider's agreements.

The asymmetry this creates

The engines a lawful public dataset may measure are close to the inverse of the engines that matter most by reach. Anyone may publish what Perplexity cites. Nobody may publish what Gemini cites by using Gemini's own grounding API. Google AI Overviews, larger than the three permitted engines combined, offers no interface at all.

Published tables of "most cited domains in Gemini" and "in Copilot" nevertheless exist. They were therefore produced some other way. We are not accusing anyone of anything, and we have not audited how any specific vendor sources its data. We are saying the question has an answer, the answer changes what the numbers mean, and almost nobody states it.

So when you read an AI citation study, the useful question is not how large the sample was. It is which interface produced this, and were you permitted to keep it?

What it means for this site

We measure Perplexity, OpenAI and Anthropic. We do not query Gemini's or Bing's grounding APIs. We do not scrape Google AI Overviews. Our answer-volatility harness calls Gemini's plain generateContent endpoint with no search tool attached, so it returns the model's own output rather than Grounded Results, and the clause above does not reach it.

The cost is real and we would rather state it in advance than have it found later: our citation coverage understates Google, the largest AI answer surface in the world. Every citation figure we publish carries that limitation. So should everyone else's.

How the grounding clause shapes pricing and measurement options

The legal restriction itself does not prevent measurement, but it creates a cost asymmetry that explains why tool pricing differs across engines. Measuring Perplexity, OpenAI, and Anthropic through their official APIs is straightforward: the terms permit building indices, the APIs return real publisher URLs, and vendors can publish reproducible data. These vendors can charge relatively stable, transparent pricing (Nightwatch, Gumshoe, Peec AI, Rankscale all measure these three engines freely and publish their prices on public pages). Measuring Google Gemini and Microsoft Copilot requires workarounds. A vendor cannot use the official grounding APIs without violating the terms, so they must either scrape rendered results from a browser interface (labor-intensive, fragile to UI changes), purchase or license an existing citation dataset, or negotiate a private data partnership with Google or Microsoft. Each workaround costs more than using the public API. Browser-based scraping requires infrastructure to run thousands of concurrent sessions and parse unstructured HTML—a cost baked into every measurement. Proprietary datasets or partnerships require either cash payments, revenue share deals, or exclusive restrictions that limit what the vendor can offer clients. That cost asymmetry is why Profound (which claims to measure all major engines including Gemini) does not publish pricing: it has negotiated private terms that are not price-efficient at entry level. Nightwatch, which measures only the three legal engines, publishes £299 per month upfront. The pricing difference reflects the measurement cost difference, not measurement quality. A brand wanting to measure all five major AI engines (Gemini, Copilot, Perplexity, ChatGPT, Claude) faces a choice: commit to an expensive vendor like Profound that has negotiated workarounds, or accept a cheaper vendor that measures only the three engines with legal APIs. That choice is not about tool quality. It is forced by Google and Microsoft's contractual position. For a bootstrapped founder or growth-stage firm, this asymmetry makes the three-engine tools (Nightwatch at £299/month, SE Ranking at £89 to £200/month, Gumshoe at £349/month) more cost-efficient than chasing comprehensive coverage through Profound or similarly enterprise-focused solutions. For an enterprise team that has already allocated budget for comprehensive tracking and the legal clause is not a binding constraint, the premium vendor may deliver the breadth they need. Understanding the clause's effect on pricing helps you distinguish between "this tool is more expensive because it is better" and "this tool is more expensive because it faces legal restrictions that force it to use costly measurement methods."

---

Questions

Can you legally build an index of the sources Gemini cites?

Not through Google’s Grounding with Google Search API. Its Additional Terms of Service, effective 23 March 2026, state that it is a violation "to use Grounding with Google Search to extract or collect one or more of these components for another purpose (for example, using programmatic or automated means to collect Links, using Links to build an index, or using Links to identify destination pages for crawling or scraping)". The same clause separately forbids caching, analysing or syndicating Grounded Results, subject to limited exceptions that let a developer store Grounded Result text for up to two years for display optimisation, chat history, or temporary resubmission. None of those exceptions permits publishing an index built from the Links.

Do Microsoft’s Bing grounding terms allow a citation database?

No. Section 4(b)(o) of the Grounding with Bing Search terms, last updated November 2025, prohibits you from "Copy, store, cache, archive, or create a database of Output." Section 4(b)(g) separately prohibits using the service "to create a database or service that competes with Output received when using the Services".

Which AI engines can lawfully be used to measure citations?

Perplexity, OpenAI and Anthropic each return real publisher URLs through a documented API field, and none of their public terms contains an equivalent clause forbidding index building. Google AI Overviews, the largest AI answer surface by reach, has no official API at all.

Does this mean published Gemini and Copilot citation tables are unlawful?

No, and we are not saying that. It means those tables were not built using the grounding APIs. They came from somewhere else: a rendered browser session, a scraped result page, a licensed panel, or a data partnership. Which one is not a footnote, because it determines whether the numbers describe the product a user sees or an API that approximates it.

---

Terms change. Google's general terms were scheduled to update on 30 July 2026. Read the primary sources yourself before you build anything. If either clause changes, tell us and we will log the change in corrections.