Gemini for Construction: What It's Actually Good For

I’ve already put Claude and ChatGPT through their paces for construction work, comparing them task by task. This is the third one: Google Gemini. I ran it against the same construction tasks I use every week, drawings, correspondence, reusable workflows, and recorded exactly what I found in the video below. This is for anyone deciding whether Gemini deserves a seat in their AI stack, or where it’s worth keeping around as a second tool.

Key takeaways

• Gemini’s clearest edge is file size: it took an 8 MB set of construction drawings straight into context in my test, while Claude has failed to upload drawing sets smaller than that.
• On a visual-reasoning benchmark, Gemini 3.1 Pro scored 71% against Claude Opus 4.6’s 55%, though Claude’s newer Opus 4.7 wasn’t on that chart.
• Deep research is Gemini’s other standout: it searches your own Drive, Gmail and connected files, useful for chasing down old correspondence on a live job.
• Gemini needs far more explicit prompting than Claude, and its Gems, its reusable workflow templates, only connect to data inside the Google ecosystem.
• None of this moves the trust boundary. Gemini still shouldn’t count or measure off a drawing. Extraction is the job, not judgment.

What is Gemini for construction?

Gemini is Google’s general-purpose AI model, the same category as Claude and ChatGPT, built into Google Workspace and accessed by typing “Gemini” into Google search. For construction use, its standout features are handling much larger file uploads than Claude or ChatGPT, deep research across your own Drive and Gmail, and Gems, Google’s version of a reusable AI workflow.

If you already run a Google Business Account, Gemini comes included. There are separate paid plans too, up to $400 a month, though I have no real idea what you’d use that tier for. In months of use I have not hit a rate cap on Gemini the way I regularly do on Claude or ChatGPT, and there’s a phone app for asking quick questions from the ute. The chat interface itself looks like every other AI tool at this point: a conversation window, previous chats, a spot for Gems, and a link into Notebook LM, Google’s separate research tool.

Where Gemini actually wins: drawings and file uploads

Gemini’s single clearest advantage for construction is file size: it accepts far larger uploads than Claude or ChatGPT and, on my test, scored 71% on a visual-reasoning benchmark against Claude Opus 4.6’s 55%. That combination makes it the more forgiving tool for a first pass over a full drawing set, though the gap on reasoning is closing fast.

In the video I uploaded an 8 MB set of construction drawings and asked Gemini to extract the key material requirements and prepare a procurement register. It took the file straight in and ran the analysis. That is not a given. In a separate test on Claude, an 11 MB drawing set failed to upload entirely, even though it sat under Claude’s own stated 30 MB per-file limit, and the error simply said the file format might not be supported. Gemini, in my experience, “almost always successfully” handles sets that size, and I’ve pushed it as far as 20 MB drawing bundles.

The benchmark backs up what the upload test shows. Pulling up a general visual-reasoning leaderboard, Gemini 3.1 Pro sat at 71%, with Claude Opus 4.6 at 55% and no listing yet for the newer Opus 4.7. As I said on camera, “Gemini is really a step above the different models at being able to process visual information.” That said, I’d already noticed the gap narrowing in the weeks before I recorded this, and Claude’s Opus 4.7 has since made a real step up on visual analysis. Model rankings move fast enough that the number matters less than the pattern: whichever model you’re on, check its current file-size limit and its current visual-reasoning score before you rely on it for a full drawing set.

None of this changes where the line sits. Reading and extracting what’s written on a drawing is a different job to measuring it, and that split holds regardless of which model does the reading, covered in more depth in AI for reading construction drawings.

Gemini’s deep research: searching your own files and inbox

Gemini’s deep research tool can search inside your own connected files, not just the public web, which is what makes it useful on a live job. Point it at Gmail or Drive and it digs through months of correspondence for a specific detail, the kind of task that normally means scrolling an inbox by hand.

The feature itself works the same way it does in ChatGPT or Claude: give it a research question, it goes and finds sources, comes back with a summary. The difference with Google’s version is that once you’re logged in, it automatically connects to your own Gmail, and I can search through it in real detail for a missing email or something I can’t otherwise find. On a project with a heavy volume of correspondence, variations, RFIs, subcontractor emails, that’s a genuinely useful application, not just a demo trick.

Gemini also has scheduled actions built into the same settings area. I set one up to email me a summary of unread messages every morning at 9am, which connects to Gmail and just runs on its own from then on. It’s a small thing, but it’s the kind of recurring document-control task that otherwise eats ten minutes a day across a job, every day, for the life of the project.

Gemini vs Claude: skills, memory and the ecosystem lock-in

Gemini’s Gems are Google’s version of a Claude project combined with a skill, a reusable workflow you build once and run again. In my test they worked, but felt less structured than Claude skills, and they can only pull data from inside the Google ecosystem: no arbitrary third-party connectors.

Dimension	Gemini	Claude (my test)
Large drawing/file uploads	Handled an 8 MB drawing set easily, up to roughly 20 MB in practice	An 11 MB drawing set failed to upload, despite sitting under Claude’s stated 30 MB limit
Visual reasoning benchmark	Gemini 3.1 Pro: 71%	Claude Opus 4.6: 55% (Opus 4.7 not yet on this chart)
Prompting style needed	Explicit, prescriptive instructions	Infers intent with less detail
Reusable workflows	Gems, Google Workspace data only	Skills, broader third-party connectors
Usage on the standard plan	Rarely hits a cap	Chews through allowance quickly

I built a procurement Gem that takes a set of tender documents and returns a proposed set of procurement packages with scopes of work, and it worked. But I’m not a massive fan of Gems generally, I find them a bit less intuitive and structured than Claude skills, including the ones our ContractorOS community has packaged for exactly this kind of procurement workflow.

The other real difference is how much you have to spell out. As I put it in the video: “I find when you’re doing stuff with Gemini, you have to be very, very explicit about the task you want it to do, which I haven’t been finding with Claude.” Claude seems to carry a deeper read on what you’re actually trying to do; Gemini needs the instruction written out. It’s the same trade-off that shows up when you compare Claude and ChatGPT for construction: the tool with the deeper workflow layer asks more of you upfront, the tool with the shallower one asks less, until it gets something wrong.

Setting up Gemini prompts that don’t hallucinate quantities

Gemini needs a more explicit prompt structure than Claude: background information, the transformation you want, the exact output format, plus a hard instruction not to count or measure. That last part is not optional. Get loose with a Gemini prompt on drawings and it will happily estimate a quantity that was never actually counted.

I think of every AI task as a data transformation: you’re moving information from one format into another, so the prompt needs to name three things.

1. The background. What project, what role, what you’re actually trying to achieve. “I’m working as a HVAC contractor to deliver this project, I need a complete bill of materials to order from a supplier.”
2. The transformation. What you want done to the input. “Extract all the relevant information from the drawings and put it into a table.”
3. The output format. A table, a Google Sheet, a register, whatever the next step needs.

I caught myself mid-prompt in the video getting this wrong. My first pass said “do not try to extract or count quantities,” which is a bad instruction, it blocks extraction entirely. I corrected it live to “do not try to count or measure, simply extract what is easily available.” That’s the actual line: pull what’s written down, don’t estimate what isn’t.

On top of that structure, I add three things to the end of almost every prompt, whichever model I’m using: ask clarifying questions before starting, so the model tells you what information it’s missing instead of guessing; be hypercritical and blunt rather than agreeable, useful when you want real feedback on something like a variation you’ve drafted; and break big tasks into small steps, because every extra step of reasoning is another chance for the model to drift off track.

Common mistakes when choosing Gemini for construction work

• Assuming visual-reasoning strength means measurement strength. A higher benchmark score on reading a drawing doesn’t extend to counting or scaling one. That line doesn’t move.
• Skipping the explicit prompt structure. Gemini doesn’t infer intent the way Claude does. Underspecify a drawing extraction task and you’ll get a vague or wrong table back.
• Building a shareable workflow as a Gem when your team isn’t fully inside Google Workspace. Gems can’t reach into arbitrary third-party tools the way a Claude skill can.
• Chasing the newest benchmark chart. Model rankings move fast, Opus 4.7 had already narrowed the visual gap by the time I finished testing.
• Running everything through one model out of habit. For a high-stakes drawing set, it’s cheap to cross-check the same extraction through Gemini and Claude and compare what comes back.

Gemini earns a real place in a construction AI stack, particularly if your files already live in Google Drive and Gmail. It is not a reason to leave Claude, and it’s not a shortcut around the trust boundary that sits under all of this, wherever this fits inside the wider picture of AI for construction: you understand and measure, AI indexes, extracts, populates and cross-checks. I walk through the full drawing upload test, the Gems demo and the exact prompts I used in the video above.