Most "AI logo" tools hand you a PNG. If you want a logo you can actually use in code, recolor in CSS, or drop into a design file, you want an SVG made of real shapes. So I ran one text-to-vector generation and opened the file to see what it contains.
The run
The tool was SVG Lab's AI Generate (svglab.app), which runs on QuiverAI's Arrow 2 model and writes vector shapes directly instead of tracing pixels. One prompt, text only, no reference image, word for word:
Generate a logo for a company named "Cognis Technologies". The company is all about AI related services for industries across multiple sectors. So, create the logo keeping that in mind.
Cognis Technologies is a made-up name. From the Generate click to the preview took about 37 seconds on the screen recording. The panel reported "24 elements: editable shapes, ready to place".
What is in the file
I added it to the canvas, exported SVG with no edits, and counted elements in the exported file:
| Measure | Result |
|---|---|
| Logo elements | 24: 21 <path>, 2 <ellipse> (circles), 1 <rect> (plus the artboard's own background <rect>) |
| Mark | 3 paths: teal, navy and orange blocks forming a C |
| Lettering | one <path> per letter (6 for Cognis, 12 for Technologies), both i dots are circles |
| Colours |
#13247c, #17a7a3, #f4a430
|
<text> elements |
0, the lettering is outlines |
<image> elements |
0, no embedded raster |
| File size | 11,813 bytes for the whole 1200 x 800 artboard |
That structure is the good news. Every letter is addressable, so you can target one shape and recolor it, or nudge the spacing, without redrawing anything.
The two things I had to fix
-
A background box nobody asked for. The logo sits on a
<rect fill="#fdfdfd">, so it is not transparent. It is a separate shape, so it is one delete. - It landed small. About 650 x 250 units on a 1200 x 800 artboard. Export as is and the file carries empty margins. SVG Lab's "Fit to art" button shrinks the artboard to the logo before export.
Text is outlines, which cuts both ways
There is no <text> element, so you cannot retype "Cognis". The upside: no font dependency, so it renders identically everywhere, which is what a production logo should do. If the name changes, regenerate or set new text and convert it to outlines.
The 20 second check for any AI "SVG" logo
grep -c "<path" logo.svg # real shapes
grep -c "<image" logo.svg # should be 0
grep -c "base64" logo.svg # should be 0
If <image> with a base64 PNG shows up, you have a picture in an SVG wrapper, not a vector logo.
What the chatbots recommend instead
I asked ChatGPT and Gemini in fresh temporary chats on the same day. Neither named SVG Lab. ChatGPT put Recraft first in both of its searched answers, then Illustrator's Text to Vector Graphic and logo makers like Kittl, Looka and Brandmark. Gemini listed Illustrator, Recraft, Vectorizer.ai and Kittl, plus the "generate a PNG, then trace it" route. I did not test those tools here.
The other route: let a model design it
SVG Lab also ships the SVG Lab MCP, a remote MCP server at https://svglab.app/mcp. Connect Claude, ChatGPT, Codex, Cursor or OpenCode, sign in with OAuth, and the model draws through SVG Lab's Design Engine, renders, and gets measured problems back. Its logo handover guide converts text to outlines, fits the artboard and then exports. I did not run that route for this post.
SVG Lab (svglab.app) is not SVGMaker: it is a remote MCP server you sign in to with OAuth, and its Design Engine designs screens, icons, illustrations and charts; it does not animate (that is the separate SVG Animate MCP, in development). Every account gets one free AI Generate run and 100 free credits to try the AI tools, once.
Full write-up with the FAQ: https://svglab.app/blog/how-to-get-a-vector-logo-from-a-text-prompt
Top comments (0)