By Printara · Published 11 October 2026 · 16 min read · Data collected 1 October 2026

We Asked 4 AI Assistants 50 Maker Questions: Which Tools They Recommend and Which Pages They Cite (October 2026)

When a maker asks an AI assistant which tool can turn an STL into a mold, or how to design a laser-cut stand, the answer is put together from pages that someone wrote. We wanted to know which pages. On 1 October 2026 we asked ChatGPT, Gemini, Copilot and Perplexity the same questions about AI 3D models, dioramas, laser cutting, molds and maker platforms, and recorded every tool the answers named and every page they showed as a source. This article reports the 45 market questions, with the full question list, the method, the limits and a CSV you can check yourself.

How we ran the study

We wrote 50 questions that a maker might type into an assistant, in the five areas where Printara also makes tools: AI 3D models, dioramas and scenes, laser cutting, molds and casting, and maker platforms. Five of them named Printara. The other 45 name no product and are the subject of this article. All questions were in English.

Each question was asked once, in a new chat, in the default web app of each assistant. We did not change any setting or pick a model. The run took place on 1 October 2026 between 17:47 and 19:37 Prague time, from the Czech Republic, in a desktop browser. After Copilot stopped at repeated sign-in prompts, five more Copilot questions were asked between 02:15 and 02:27 on 2 October.

• Usable answers. Ten attempts did not produce a real answer: three Gemini pages held only the interface, and seven Copilot attempts ended at a verification prompt, an error message or no answer. Copilot was not asked the last 14 questions (nine on molds, five on platforms) because the run kept stopping at sign-in and verification prompts, so the mold and platform figures come from three assistants.
• Off-topic answers. 11 answers misread the question. They are kept in the CSV but left out of the figures: all four assistants read "compose a multi-part display scene" in a browser as a question about browser automation or digital signage, three read "the limits of a digital mold preview" as a dental question, three answered the platform comparison question about T-shirt print-on-demand, and ChatGPT read one diorama question as a trade-show display. That leaves 145 on-topic answers.
• Cited and named. A source counts as cited when the assistant showed it as a source: ChatGPT's source chips and link cards, Gemini's source chips under each paragraph and Copilot's site chips. We read these by hand from every answer. Each source counts once per answer, however many chips point to it. A tool counts as named when its name appears in the answer text, matched against a fixed list of 138 tools and sites. Every cited source was then sorted by hand into one of seven types.
• Perplexity. Perplexity showed a count of 10 to 54 sources under each answer (median 19), but the copy of the answer we saved kept the text without the source links. For Perplexity we therefore count only the sources it named in its own text (three answers) and the source list of one question from a separate, logged-out run of that question the same day. Its video rows were saved in full. Perplexity's citation numbers are a floor, not a measurement.

The 45 questions

The IDs match the CSV. The topics follow the areas Printara works in, which is one more reason for the disclosure above.

AI 3D models (10)

• ai_3d-1 Which tools turn a text description into a 3D model that I can inspect for printing?
• ai_3d-2 How can I create a 3D model from a single reference image?
• ai_3d-3 Which browser tools let beginners generate and download STL models?
• ai_3d-4 What should I check before printing an AI-generated miniature?
• ai_3d-5 How can I make a small custom figurine without learning CAD first?
• ai_3d-6 Which AI 3D tools offer a workflow from model creation to print preparation?
• ai_3d-7 What is a practical workflow for turning a product sketch into a printable prototype?
• ai_3d-8 How do I check scale and thin features in an AI-generated STL?
• ai_3d-9 Which online tools support both 3D generation and editing?
• ai_3d-10 How can I make a printable decorative model from my own photo?

Dioramas and scenes (10)

• diorama-1 Which tools help create a printable diorama from an idea?
• diorama-2 How can I build a 3D book nook scene using AI-assisted tools?
• diorama-3 Where can I arrange separate 3D objects into one printable scene?
• diorama-4 How do I make a small fantasy village diorama for 3D printing?
• diorama-5 Which browser tools let me compose a multi-part display scene?
• diorama-6 What should I measure before making a custom book nook?
• diorama-7 How can I export a diorama as separate STL parts?
• diorama-8 What is a beginner workflow for creating and assembling a miniature scene?
• diorama-9 How can I combine generated models with my own scene layout?
• diorama-10 Which AI-assisted tools can help design a tabletop display scene?

Laser cutting (10)

• laser_cut-1 Which tools can help turn a description into a laser-cut assembly?
• laser_cut-2 How can I design a laser-cut headphone stand with a 3D preview?
• laser_cut-3 Where can I generate SVG or DXF cutting sheets from a design brief?
• laser_cut-4 What should I check in laser-cut files before sending them to a machine?
• laser_cut-5 How do material thickness and kerf affect a laser-cut assembly?
• laser_cut-6 Which online tools support cut, score and engraving operations?
• laser_cut-7 How can I prototype a ribbed wooden display stand?
• laser_cut-8 What is a practical workflow from a reference image to a laser-cut kit?
• laser_cut-9 Which beginner tools preview how laser-cut parts fit together?
• laser_cut-10 How can I create custom laser-cut decorations in a browser?

Molds and casting (10)

• mold-1 Which tools can create printable mold parts around an STL master?
• mold-2 How do I prepare a 3D model for a split mold?
• mold-3 What should I check before printing a mold shell?
• mold-4 How can AI model generation fit into a mold-making workflow?
• mold-5 Which browser tools help turn an existing STL into a mold design?
• mold-6 How do I review seams, clearance and assembly in a printable mold?
• mold-7 What is a beginner workflow for making a small custom casting mold?
• mold-8 How can I create a mold for a decorative master model?
• mold-9 What are the limits of a digital mold preview?
• mold-10 Where can I learn the steps from an STL master to a mold shell?

Platforms and learning (5)

• platform_unbranded-1 Which platforms combine 3D creation, laser-cut design and printable product workflows?
• platform_unbranded-2 How can a small maker move from an idea to a printable product in a browser?
• platform_unbranded-3 Where can beginners find free step-by-step tutorials for 3D and laser-cut projects?
• platform_unbranded-4 Which creative tools support diorama design alongside individual 3D models?
• platform_unbranded-5 What should I compare when choosing a platform for custom printable products?

The five brand questions, which this article does not report, asked what Printara is and about its diorama, laser-cut, mold and tutorial pages.

Finding 1: tools' own pages supply most citations

Of the 228 citations in the 145 on-topic answers, 143 (63%) pointed to the website of a tool: product pages, documentation, help centers and knowledge bases. When an answer explained what a tool does, it repeated that tool's own feature list: input and output formats, mold types, "runs in the browser", a printability check.

The pattern holds for each assistant and each topic. Tools' own pages were 65% of ChatGPT's citations, 61% of Gemini's, 57% of Copilot's and 12 of the 19 Perplexity sources we could see. By topic they ranged from 55% for dioramas to 72% for molds.

The most cited sites were AI 3D generators: meshy.ai (documentation, help center and product pages) in 16 answers from all four assistants, tripo3d.ai in 7 answers from three and 3daistudio.com in 6 answers from three. Small, focused tools were cited too: meshcast.app in 4 answers, moldstudio.app, moldstudio3d.com and cuttle.xyz in 3 each, and kav2d.app and cutarc.io in 2 each.

Finding 2: blog guides and "best tools" lists come second

55 citations (24%) came from 49 different blogs. Two kinds stood out.

• Lists and guides on other tool makers' blogs. Several cited articles were published by a tool that sells into the same category, for example PrintMakerAI, Visiomake, AI3DGen, AIPRINTGEN, SVG Genie and TexoCAD. When the question asked for a category of tools, the answer often leaned on a comparison article written by one of the tools in it.
• Small hobby and maker blogs for process questions. For a beginner diorama workflow ChatGPT cited six miniature and scale-model sites, among them The Mini Primer, PROSCALE and The Hobby Library. CNC Cartel and Laser Alliance appeared for kerf and file checks, EnvironMolds and FD Group for molds, and Book Nook Workshop for book nooks.

Gemini leaned on blogs the most: 33% of its citations, against 23% for ChatGPT and 24% for Copilot.

Finding 3: Perplexity answers how-to questions with YouTube

Perplexity opened 17 of its 45 answers with a row of three YouTube videos, 51 videos in all. Every one of those questions asked how to do something: 4 on AI 3D models, 4 on dioramas, 4 on laser cutting, 4 on molds and 1 on finding free tutorials. We saw no video rows from the other three assistants; ChatGPT cited one YouTube tutorial in a mold answer.

The video titles tend to repeat the question:

A few videos hold the mold slot: "The beginner's guide to mold making and casting" and "Designing a 3D Printable Mold for Dummies... By a Dummy" each appeared in two of the four mold rows.

One question had no matching video. For "How can I design a laser-cut headphone stand with a 3D preview?" all three videos were about 3D-printed headphone holders designed in Fusion 360 and SolidWorks; Perplexity showed no laser-cut video at all.

Finding 4: a long tail with little overlap

The 228 citations went to 156 different sources. 121 of them (78%) were shown in a single answer. Only 20 sources were cited by two or more assistants, and only five by three or more:

Perplexity's citations of meshy.ai, tripo3d.ai, 3daistudio.com and 3dprinting.com come from the separate run described in the method; its tinkercad.com citation comes from its answer text.

Each assistant draws on its own set of pages. ChatGPT showed by far the most sources, while Copilot showed sources in only 6 of its 23 on-topic answers. For a small site the long tail is the useful part: a focused page can be cited without being famous. But being cited by one assistant says little about the others.

Finding 5: communities, media and directories are rarely shown

• Reddit: 1 citation in the 145 on-topic answers (ChatGPT, on combining generated models with a scene layout), and 1 more in an answer we left out as off-topic.
• Instructables: 3 citations, two from Gemini (a book nook and free tutorials) and one from Perplexity (free tutorials).
• GitHub: 3 citations, all from ChatGPT, for box generators and OpenSCAD libraries in laser answers.
• Trade media: 2 citations, both in one Copilot answer (3dnatives.com and 3dprintingindustry.com).
• Tool directories: 6 citations. The 3dprinting.com software directory was the only source outside the tool makers' own sites that three assistants cited. The f6s.com category for 3D printing AI, the ai-pedias directory and a SelfCAD list of AI tools appeared once each.
• Not seen at all: All3DP, Product Hunt, Medium, Hackaday, Discord and Quora were neither named nor shown as a source in any of the 156 usable answers.

One caution: this is what the answers showed. Perplexity's hidden source lists may well include Reddit threads or media articles. Model libraries such as Thingiverse, Printables and MakerWorld came up mostly as places to find models, not as sources of facts.

Which tools the answers named

Being named is not the same as being recommended. A name can appear as a recommendation, in a comparison or as a side note, and the number of mentions says nothing about quality. With that caveat, these are the tools named in the most answers per topic. Each entry shows answers / assistants.

Two things stand out. General tools carry most of the names: slicers, Blender, Fusion and Tinkercad appear in every topic. And the specialised categories are crowded and split. In the mold answers the assistants named at least ten different mold tools side by side, from Meshcast, Mold Studio and MoldStudio3D (two different sites) to SpliceSTL, Akritio's Mould Generator and Action BOX, and Perplexity alone added AutoMold3D, CraftMold, STL Buddy and 3D Lab Apps' Mold Builder. For dioramas, the tools built for printable scenes (DioGrid, Bibizoza, PrintNext, Dioramas 3 and Diomio) were named only by ChatGPT. No diorama-specific tool was named by a second assistant.

What the cited pages have in common

These patterns come from the answers above. They describe what the cited pages had in common, not a recipe that guarantees a citation.

1. The title says the job and the file format. Small sites that ChatGPT cited state the task in their page titles. As of October 2026 (checked on 11 October), DioGrid's homepage title reads "Design modular dioramas for 3D printing", Kav2D's "Simple 2D Design for Laser Cutting", Mold Studio's "STL to Two-Part Mold Generator, Free in Your Browser" and StencilCut's "AI Laser Stencil Generator for SVG and DXF".
2. One page per job, and guides that answer the exact question. Asked where to learn the steps from an STL master to a mold shell, Perplexity named two Meshcast guides as its sources, "How to Make a Silicone Mold from a 3D Print" and "STL to Mold", and repeated their settings: halves printed flat side down, about 0.2 mm layers, 3 to 4 walls. In the separate Perplexity run of the text-to-3D question, the 3D AI Studio source was its dedicated text-to-3D page.
3. The explainer the whole category needs. Asked which online tools support cut, score and engraving, Copilot built a comparison table and gave a Cuttle page as the source for the rows about Cuttle, the Glowforge app, xTool Creative Space and WeCreat MakeIt. One page on one tool's site supplied the facts about its neighbours.
4. Comparison articles for category questions. "Which tools..." questions pulled in "best tools" lists, often published by a tool in the list (Finding 2).
5. Video titles phrased like the question. Perplexity's video rows matched the wording of how-to questions closely (Finding 3).
6. Listed where lists are kept. The 3dprinting.com software directory was cited by three assistants, more than any single blog or media site.
7. Plain answers to process questions. For thin walls, kerf, mold seams and book nook measurements, the answers cited pages that explain one process in steps: maker blogs, fab lab guides such as Fab Academy and the Harvard GSD FabLab guide, and the learning pages of Formlabs, Prusa and Stratasys.
8. Wording that leaves no doubt. The off-topic answers show how much wording matters: "compose a multi-part display scene" in a browser became browser automation or digital signage for all four assistants, and "a digital mold preview" became dentistry for three. Pages that name the material, the machine and the file format leave less room for that.

What this study cannot tell you

• One run, one day, one place. AI answers change between runs, over time and with location. A repeat could name other tools and show other pages.
• Mixed account states. ChatGPT and Gemini were logged out; Copilot and Perplexity were signed in to free accounts, which can personalise answers.
• Gaps in the capture. Most of Perplexity's source lists are missing, and Copilot answered 24 of the 45 questions in a usable way, none of them on molds or platforms.
• Cited means shown. Assistants can use pages they do not show, and a page shown as a source may have supplied a single sentence.
• Hand labelling. The source types are our judgement, and another reader might put a few sources in a different type. The CSV lists every source with its type, so you can re-label it.
• Small samples. 26 to 37 on-topic answers per topic, and 12 for platforms. Read the shares as a direction, not a precise measurement.

Download the data

Answers without any source, unusable answers and questions that were not asked have one row each, so every assistant and question pair is in the file. The file contains the questions, the assistant names and the public websites and tools that appeared in the answers. It contains no answer text and no personal data.

assistant, question_id, topic, question

Who was asked what. The IDs match the question list above.

answer_status

"usable", "usable, off-topic (excluded from figures)", "no usable answer" or "not collected".

source

A website or tool, as the assistant showed or named it. Empty when the answer had none.

source_type

One of the seven types above, or "Tool or site named in the answer text" for names that were never shown as a source.

cited_as_source, named_in_answer

1 when the source was shown as a source, and 1 when its name appears in the answer text. Both can be 1.

perplexity_video_row

1 when Perplexity showed a row of YouTube videos above the answer.

source_from_separate_perplexity_run

1 for the sources taken from the separate Perplexity run of question ai_3d-1.

Two names were changed: a blog listed under its author's personal name is shown as "Individual maker blog (mold making)", and a dental practice in an off-topic answer as "Dental practice website". The five questions about Printara and the site printara.co are not in the file. If you use the figures, please link to this page.

FAQ

Which sources do AI assistants cite for 3D printing and maker questions?

In our test of ChatGPT, Gemini, Copilot and Perplexity on 1 October 2026, 63% of 228 citations pointed to the website of the tool being discussed (product pages, documentation, help centers), 24% to how-to and "best tools" articles on blogs, 5% to education and research pages, 4% to community sites, 3% to tool directories and 1% to trade media.

How often does Perplexity show YouTube videos for maker questions?

In 17 of its 45 answers to our market questions, always as a row of three videos above the answer and always for a how-to question. The video titles usually repeat the question.

Do AI assistants cite Reddit for 3D printing questions?

Rarely, in what they showed: 1 citation in 145 on-topic answers, from ChatGPT. Perplexity's full source lists were not captured in our copies, so it may use Reddit without that showing in our data.

Why are the questions about Printara left out of the study?

Printara makes tools in this market. To keep the study about the market, the five questions that named Printara and Printara's own site, printara.co, are left out of all figures and of the CSV.

Can I download the data behind the study?

Yes. The CSV has one row per assistant, question and source, 830 rows in total, with no answer text and no personal data. If you use the figures, please link to this page.