Gemini Enterprise and enterprise search

The returns team at Birchline answered the same questions all day. Can a customer return a tent that was used once? What if the receipt is from a different store? Does the 60-day rule apply to online orders shipped to a store? The answers were spread across a policy library on the intranet, a folder of memos in Drive, the ticketing system's knowledge base and the head of returns' memory. New staff took weeks to learn where to look. "Gemini in Gmail won't fix this," Owen said at the third session. "The answers aren't in anyone's inbox."

"Right," Farah said. "This is a different problem. People need to ask a question once and get an answer drawn from all the places the company keeps its knowledge, with links so they can check it. That's what Gemini Enterprise is for." This lesson explains Gemini Enterprise, enterprise search and the agents that come with it, and how it differs from both Gemini in Workspace and a custom build.

What Gemini Enterprise is

Gemini Enterprise is Google Cloud's workplace product that gives employees a single place to chat with Gemini, search across company data and use agents. It connects to many sources, including Google Workspace, Microsoft 365, common business systems such as customer relationship management and ticketing tools, intranets and databases, through connectors. It then answers questions using that data and shows where each answer came from. The exam guide lists Gemini Enterprise as a prebuilt offering and highlights three capabilities: multimodal search, custom agent capabilities and the Gemini Notebook API, which brings notebook-style research over chosen sources into the product.

Farah placed it on the stack: Gemini Enterprise is an application, but one that is designed to be assembled. An administrator connects data sources, sets who can use which agents, and can let staff build simple agents of their own without code. It needs more setup than Gemini in Gmail and much less than a custom build.

How enterprise search works

Ordinary web search finds public pages. Enterprise search finds company information that is spread across many private systems, and it must respect who is allowed to see what. Gemini Enterprise does this in four broad steps.

• Connect. Connectors link each source system. Some connectors copy content into an index that Google manages; others query the source live when a question is asked. The choice affects freshness, speed and where data is held.
• Index and understand. Content is processed so that it can be searched by meaning, not only by exact words. A question about returning a used tent can match a policy that talks about opened camping equipment.
• Check permissions. Results are filtered by the person's access in the original system, so a store associate does not see a document that only head office can open.
• Answer with sources. Gemini writes an answer from the most relevant passages and lists the documents it used, so the person can open them and check.

That last step is grounding, which GCP-110 introduced: anchoring a model's answer in specific, trusted sources rather than in whatever the model learned during training. Grounding reduces made-up answers and makes answers checkable. Lesson 8 explores grounding in more depth, including grounding with Google Search.

Multimodal search

Company knowledge is not all text. Birchline's product team keeps photos of every item, the warehouse has scanned delivery forms, and training material includes videos. Multimodal search means searching across these kinds of content together and asking questions that combine them, such as "show me the setup video and the instruction sheet for the two-person dome tent". For the exam, the key idea is that Gemini models understand images, audio and video as well as text, so enterprise search can draw on them too.

Agents in Gemini Enterprise

Gemini Enterprise is also where employees find and use agents. Some are made by Google, such as agents for deep research across many sources or for analyzing data. Some come from partners. Some are built by the company itself, either by developers on Agent Platform or by business users with a no-code designer inside Gemini Enterprise. Administrators decide which agents are available to whom.

Farah's rule was to start with the simplest kind that works. A returns helper that only reads policies and drafts replies can be built by the returns lead in an afternoon. An agent that changes data in another system, such as creating stock transfers, needs developers, testing and security review, because a mistake has real consequences.

Gemini Enterprise compared with Gemini in Workspace

Owen found the overlap confusing, and the exam tests it. Both use Gemini and both respect access. The difference is scope and purpose.

Search for customers is a different product family

Jules asked whether the same search could power the search box on the online shop. Farah explained the distinction the exam guide draws. Gemini Enterprise is for a company's own employees. Search for external users, such as customers browsing a website, is built with Agent Search on Gemini Enterprise Agent Platform, which lets developers create search and recommendation experiences for websites and apps using company data such as a product catalog. Lesson 8 covers it with other grounding tools. "Same idea, different audience," Farah said. "Employees get Gemini Enterprise. Customers get something we build with Agent Search."

Setting it up well

A good enterprise search project is mostly about the content, not the model. Farah gave Owen her checklist, built from a project at her previous employer.

• Choose a first use case with clear value. The returns questions were ideal: frequent, costly when wrong, and answerable from documents that already existed.
• Fix the sources first. Out-of-date or contradictory documents produce out-of-date or contradictory answers. The returns team retired 23 old memos before connecting the library.
• Check permissions in the source systems. Search inherits whatever access the source allows, including mistakes.
• Decide how fresh data must be. Live queries are always current but slower; indexed content is faster but updated on a schedule.
• Measure answer quality. Keep a list of real questions with known correct answers and test regularly.
• Tell people how to check answers. Sources are shown for a reason; staff should open them for anything unusual.

# A quick quality check: compare search answers with known correct answers.
test_questions = [
    ("Return window for online orders", "60 days", "60 days"),
    ("Used tent, unopened box missing", "Store credit only", "Store credit only"),
    ("Receipt from another store", "Accepted at any store", "Accepted at any store"),
    ("Bike helmet after a crash", "Not returnable; crash replacement offer", "Not returnable"),
    ("Gift card refund", "Not refundable for cash", "Refund to original payment"),
]
correct = 0
for question, expected, answer in test_questions:
    ok = expected.lower().startswith(answer.lower()) or answer.lower().startswith(expected.lower())
    correct += ok
    print(f"{'OK ' if ok else 'BAD'} {question}")
print(f"{correct} of {len(test_questions)} answers acceptable")

OK  Return window for online orders
OK  Used tent, unopened box missing
OK  Receipt from another store
OK  Bike helmet after a crash
BAD Gift card refund
4 of 5 answers acceptable

The illustrative test above shows the habit, not a real result: a short list of questions with agreed answers, rerun after every change to the sources. The wrong gift card answer would send the team back to the source documents, where they might find two policies that disagree.

Costs and licenses

Gemini Enterprise is sold per user, in editions with different features and limits. Dana asked how to decide who needed it. Farah suggested starting with the teams that spend the most time looking things up: returns, customer service, store support and the buying office, about 140 people, rather than all 1,100 staff. Staff who mainly need help writing in Gmail and Docs would be served by Gemini in Workspace. Editions, prices and included features change, so the course does not list them; a leader should compare the current editions against the actual use case.

Who owns enterprise search

Enterprise search touches many systems, so ownership is easy to leave unclear. Farah asked Birchline to name three owners before the pilot. A business owner, here the head of returns, decides which questions the search must answer well, approves the test set and signs off on quality. Content owners, one for each source system, keep their documents accurate and current and retire outdated ones; the policy library has an editor, and the ticketing knowledge base has a team lead. A technical owner on the IT team manages connectors, licenses and settings, watches usage reports and handles problems. Teresa's security team reviews access settings before each new source is connected.

Owners matter most when something goes wrong. When the search gives a wrong answer, the technical owner checks whether the right document was retrieved; if it was, and the document itself is wrong, the content owner fixes it; if the wrong document was retrieved, the technical owner adjusts the source settings or reports the case. Every wrong answer becomes a new test question, so the same mistake is caught next time.

Farah added a warning about expectations. Search across company data feels like magic in a demonstration, because demonstrations use clean, well-chosen documents. Real collections contain duplicates, drafts, outdated versions and documents that disagree. The first weeks of any enterprise search project are usually spent finding and fixing these problems, and that work has value of its own: the returns team discovered that two policies gave different time limits for opened camping stoves, a conflict that had been confusing staff for a year.

The returns pilot is scheduled for February 2027, after the old memos are cleaned out. The head of returns had already written down forty questions that new staff ask most often, and those would become the test set.