AI Grounding: Every AI Engine Grounds in Live Search, Except Gemini

Do all AI engines ground their answers in live search?

In our measurement across roughly 2,500 prompts per engine, five of six AI engines grounded their answers in live search 91–98% of the time. Gemini grounded just 41.1%.

AI grounding is what those numbers describe: the engine ran a live search, read real web results, and wrote its answer from them instead of from memory. Ask ChatGPT, Perplexity, or Google’s AI Mode a question today and the answer almost always comes back with citations, the little numbered links to pages the model read while answering.

It has quietly become the default. But not for everyone, and the engine that breaks the pattern is the one most people would not guess.

What AI grounding actually means

AI grounding is the step where a model connects its answer to live web content retrieved at query time. Instead of generating a response purely from the statistical patterns it learned in training, a grounded engine issues a search, pulls back real pages, and composes its answer from them, and attaches citations so you can check the sources. Google calls its version Grounding with Google Search and describes it this way: it “connects the Gemini model to real-time web content” and shows “the sources for the model’s claims.”

Grounding is how engines perform real-time extraction on AI-generated search citations: they search, read, and cite live pages at query time.

It helps to place AI grounding next to the three things people confuse it with:

• Fine-tuning bakes knowledge into the model’s weights before it ships. It’s expensive, slow to update, and frozen the moment training ends. That makes it useless for anything that changed yesterday.
• Long context lets you paste a lot of text into a single prompt. Useful, but you have to supply the text, and it doesn’t reach anything the model wasn’t handed.
• RAG (retrieval-augmented generation) retrieves passages from a corpus you control (your docs, your knowledge base) and feeds them to the model at answer time.

Grounding is retrieval where the corpus is the live web. Mechanically it’s RAG whose retriever is a search engine instead of a private vector store: the freshest, broadest source available, with citations built in. That’s why grounding specifically means live search data. Freshness, breadth, and verifiability are the whole point. When an AI answer cites a news article from this morning, that’s grounding; fine-tuning could never have known.

The natural question is how consistently AI grounding actually happens across the engines people use. So we measured it.

We measured grounding across six AI engines

cloro monitors how AI engines answer real prompts. It’s what the AI visibility product is built on. That gave us a direct way to quantify grounding: for each answer an engine returned, does it carry at least one live source, and how many?

We pulled a slice of cloro’s monitoring corpus (roughly 2,500 prompts per engine over a 7-day window) and scored every answer. “Grounded” means the response carried ≥1 live source (a cited web page). The prompts are real customer AI-visibility queries: conversational, commercial, and multilingual. They are the kind of questions people actually put to these tools rather than a synthetic benchmark set.

Here’s what five of the six look like.

Measured 2026-07-05. See the August re-measurement below before acting on these numbers.

Engine	Grounded (≥1 live source)	Avg sources per answer
ChatGPT	98.4%	14.1
Google AI Mode	98.2%	9.9
Copilot	97.5%	4.6
Perplexity	94.9%	9.5
Google AI Overview	91.2%	9.1
Gemini	41.1%	3.2

Five engines cluster tightly between 91% and 98%. For ChatGPT, AI Mode, Copilot, and Perplexity, AI grounding is how they answer almost every time rather than a feature that occasionally fires. This lines up with our earlier finding that ChatGPT uses web search in 80%+ of prompts under clean testing conditions; measured on this corpus, with citation parsing across the full answer, the grounded share is higher still. Google’s AI Overview sits slightly lower at 91.2% (some queries render an Overview that leans more on Google’s own knowledge layer than on cited pages), but it’s still grounded nine times in ten.

Then there’s Gemini.

Why Gemini grounds least of the six

Gemini grounded 41.1% of its answers. On the other roughly 59%, it answered from training memory, with no live search and no citations. That’s not a rounding difference from the pack; it’s a different behavior.

The reason is architectural. For the Gemini API, grounding is an opt-in tool a developer turns on. Google’s documentation describes enabling the google_search tool, after which “the model handles the entire workflow of searching, processing, and citing information automatically.” When that tool isn’t invoked, Gemini does what a language model does by default: it answers from its parameters. The consumer surface reflects the same tendency: Gemini reaches for live search selectively rather than reflexively. (If you want to see the grounded-vs-ungrounded split up close, we pulled it apart in scraping Gemini, and /gemini/ shows how cloro parses its citations.)

This matters more than a trivia stat. It means that when Gemini answers, there’s a better-than-even chance it’s telling you what it remembered, not what the web currently says. For a factual or fast-moving query, that’s the difference between a citation you can verify and an assertion you can’t. And for anyone doing AI-visibility work, it means live-source strategies, the ones that get you cited, reach Gemini least often of any major engine. In a later window where Gemini carried a live source on 55.8% of answers, it still cited a LinkedIn URL once in 56,423 answers, so a rising grounding rate does not mean an engine will cite any given domain.

Source depth is a visibility lever too

AI grounding isn’t binary. An engine that cites one source is grounded; so is one that cites twenty. Depth tells you how much of the live web an answer is actually built on.

ChatGPT is the heaviest reader in the set, pulling 14 sources per answer on average. It fans a query out across many pages before writing. AI Mode, Perplexity, and AI Overview land in a middle band of 9 to 10 sources. Copilot grounds nearly every time but reads shallower, around 4.6. And Gemini, on the minority of answers where it does ground, cites just 3.2: the least likely to ground and the thinnest when it does.

Depth is its own lever for visibility: a 14-source answer has far more citation slots than a 3-source one. We took the citation anatomy apart engine by engine in how each AI engine cites sources. It’s worth reading alongside this if you care about which pages get pulled rather than how many.

Re-measured, August 2026: these rates move a lot

We re-ran the panel of six engines on 2026-08-19 over a 30-day window at 3,000 prompts per engine. Almost every figure had moved, and one had moved enough to change the story.

Engine	Grounded, Jul 05	Grounded, Aug 19	Sources, Jul 05	Sources, Aug 19
Perplexity	94.9%	99.2%	9.5	13.0
Google AI Overview	91.2%	94.3%	9.1	8.2
Copilot	97.5%	91.8%	4.6	4.5
Google AI Mode	98.2%	91.0%	9.9	13.0
ChatGPT	98.4%	78.9%	14.1	12.7
Gemini	41.1%	70.7%	3.2	10.1

Gemini nearly doubled its grounding rate and tripled its source count. ChatGPT, which led the July panel, is now second from the bottom. The tight 91-98% cluster that made the original finding so clean has spread out.

The 41.1% was a trough, not a property

Pulling Gemini month by month explains it, and the lesson generalises:

Month	Grounded	Avg sources
2026-01	57.4%	4.2
2026-02	61.2%	4.2
2026-03	60.6%	4.5
2026-04	69.6%	4.9
2026-05	27.3%	1.9
2026-06	37.2%	2.7
2026-07	55.0%	6.0
2026-08	60.8%	6.6

Gemini has ranged between 27% and 70% in eight months. The July measurement caught it climbing out of a May-June trough, so 41.1% was an accurate reading of a moment and a poor description of the engine. We published it as the latter, and this section is the correction.

What actually survived

Two claims held across both measurements and the monthly series:

• Gemini grounds least of the six. True in July at 41.1% and true in August at 70.7%, when the next-lowest was ChatGPT at 78.9%. The gap narrowed; the ordering did not change.
• Copilot runs the thinnest citation set. 4.6 sources in July, 4.5 in August, against 8.2 to 13.0 for everything else. It is the most stable number on the page.

What did not survive is any precise grounding percentage as a durable fact about an engine. If you are building on these numbers, treat the ordering as the finding, re-measure before you quote a rate, and date whatever you publish. We now re-measure every figure we cite on a schedule, for exactly this reason.

Why AI grounding is the whole game for AI visibility

Across the six engines, most AI answers are now grounded in live web content. Five of the six engines people use every day run a search before they answer, and they cite between 3 and 14 pages when they do.

AI visibility means being one of the sources an engine cites. That’s precisely why it is a real, winnable channel rather than a black box:

• If an engine grounds, its citations are contestable. ChatGPT reading 14 live pages per answer means 14 slots your content could occupy. Grounding is what turns “getting mentioned by AI” from luck into an addressable surface.
• The engine’s grounding behavior is your strategy input. For ChatGPT, AI Mode, Copilot, and Perplexity, live-source optimization pays off on nearly every query. For Gemini, being cited helps only on the ~41% of answers it grounds. Reaching it more often means winning where it does search, and accepting that memory covers the rest.
• You can’t optimize what you don’t measure. Which engines ground on your queries, and whose pages they cite, is an empirical question with a different answer for every brand and topic.

That last point is the job cloro does. It runs your prompts across these engines and reports, per query, which grounded and who they cited. That’s the same measurement behind this study, pointed at your footprint. If you’re doing brand or AI search tracking, knowing which engines answer from the live web (and can therefore cite you) versus which answer from memory is the difference between a plan and a guess.

How to ground your own LLM or agent

If you’re building on models rather than monitoring them, grounding is a pattern you implement in four steps:

1. Take the user’s question and use it as the search query, rewriting it first if the question is conversational.
2. Call a search API for that query and take the top results.
3. Format those results as context (title, URL, snippet) and pass them to the model.
4. Have the model answer from that context, citing the results it used.

In pipeline form:

user question
   → call a search API for the query
   → format the top results as context (title, url, snippet)
   → pass that context to the model
   → model answers, citing those results

Either inject the results into the prompt as context, or expose search as a tool the model calls itself (which is exactly what Gemini’s google_search tool and ChatGPT’s web search do under the hood). The quality of the grounding is only as good as the search layer feeding it, so the retriever choice matters.

A Google SERP API like cloro returns the live results page as structured JSON (organic links plus the AI Overview, People Also Ask, and shopping units), so you can ground a model on what Google actually surfaces, features included. If you’d rather the search happen inside the model’s own tool loop, Copilot and Bing-based grounding and Anthropic’s Claude both wire a web index in directly; we measured how close one of those indexes sits to Google in Brave Search API vs Google. Pick the retriever whose output your product needs, then feed it in. That’s grounding.

Which API do you use to observe grounding in practice?

Watching grounding from outside means scraping the consumer answer, not calling the developer API. cloro returns the grounded answer with sources (ChatGPT 5 credits, Gemini and AI Mode 4, Perplexity 3), putting a 1,000-prompt daily ChatGPT panel at about $60 a month at Hobby rates. Tavily and Exa sit on the other side of the pipeline: they supply retrieval to your model rather than observing a consumer engine’s.

Methodology and honest caveats

A few things to hold this data honestly:

• It’s cloro’s own corpus, presented with methodology. These are first-party measurements from cloro’s AI-visibility monitoring on real customer prompts rather than a public benchmark. Numbers are within-engine and directional: they describe how each engine behaved on this prompt mix rather than serving as a universal census of all AI answers.
• The prompt mix has a shape. The corpus skews conversational and commercial, and spans multiple languages. A different query set (pure factual lookups, or code questions) would move the absolute grounding rates. The ordering (five engines high, Gemini low) is the robust finding.
• AI Overview is measured where it rendered. The 91.2% figure is computed only on captures where an AI Overview actually appeared, with sources read from the Overview’s own block (result.aioverview), not the surrounding organic SERP. Queries that render no Overview aren’t counted for that row.
• “Grounded” is a floor, not a grade. It means ≥1 live source was present. It doesn’t judge whether the source was good, whether the model used it faithfully, or whether the citation was relevant. It says only that the engine reached for the live web at all.

Within those bounds, the headline is stable and striking: grounding is now the norm across the AI engines people use. Gemini is the one that most often answers from memory instead. If you’re deciding where to invest in being seen by AI, that map of who searches and who remembers is where to start.

Want to know which engines ground on your queries, and whose pages they cite? See how cloro’s web search and AI visibility monitoring works.