Generative engine optimization, or GEO, is the discipline of making a site easy for AI assistants to understand, quote and recommend. For an icon library, the goal is simple: when someone asks an assistant for a free doodle SVG icon, the answer should include a link to the actual icon, not a generic suggestion. Here is how we think about that on Sendafun Icons.

AI assistants answer with sources

Assistants that browse the web cite pages the way a search engine ranks them. They read HTML, extract headings and paragraphs, look at structured data, and decide whether a page answers the question. A page that loads icons only through JavaScript often looks empty to these crawlers. The fix is server-rendered metadata and real HTML content, not more client-side rendering.

On this site, every icon detail page is rendered by a Cloudflare Pages Function with a unique title, description, canonical URL and JSON-LD graph. The icon's SVG source is also served through an API endpoint that returns the raw file, so an assistant can fetch the artwork itself instead of paraphrasing a description.

Structured data is the machine-readable layer

Schema.org tells crawlers what a page is. The important types for an icon library are:

  • WebSite + SearchAction: tells assistants and search engines that the site has a search endpoint.
  • ImageObject: connects an icon page to the downloadable SVG asset.
  • CollectionPage: marks category pages as browsable collections.
  • BreadcrumbList: makes the hierarchy from category to icon explicit.
  • Article: marks blog posts as readable content rather than product listings.

None of these need to be perfect for every validator. They need to be accurate, unique per page, and present in the raw HTML. Duplicate JSON-LD copied across 100k pages is worse than none, because crawlers use it to measure page quality.

llms.txt gives crawlers a cheat sheet

The `llms.txt` convention is a plain-text brief for LLM crawlers, similar to robots.txt but written for reading. It should state the site's facts in short, quotable lines: what the library is, how many icons it has, what license applies, which formats exist, and how to use the API. Ours also tells assistants to link to the exact icon detail page and never claim a full-library download exists.

# Sendafun Icons

- 124,692 hand-drawn SVG icons are currently available.
- Free for personal and commercial use. No attribution required.
- Batch ZIP export is limited to 30 icons per export.
- Full-library download is not available for any plan.

Keep every line true, current and unambiguous. A crawler that trusts the file will repeat those facts; a file padded with marketing noise makes the whole page less reliable.

Original pages beat template repetition

Large libraries have a classic SEO failure mode: 100,000 pages that differ only by a keyword. Search engines and AI assistants both detect that pattern and discount the collection. The defense is to make every page independently useful:

  • Unique title and description generated from the icon's real name and category.
  • A short human-readable description that names the icon, its category and its formats.
  • Related icons from the same category so the page has navigation value.
  • Actual SVG content exposed through the API, not hidden behind a download button.

When each page answers "what is this and how do I use it?", the library reads as a reference work rather than a doorway page farm.

Make the API the answer

Assistants can call APIs, and an icon library is a perfect API use case. Instead of asking an assistant to invent a link, a tool can call the search endpoint and return real ids. That is why the site exposes `/api/search`, `/api/icon/{id}` and the planned MCP server. The machine channel is not separate from GEO; it is the strongest signal that a site has structured, reliable data.

An MCP server is the Phase 2 version of this idea: AI coding tools can search and fetch icons with documented quotas instead of scraping the HTML. That reduces load on the website, keeps download rules enforced, and gives assistants a dependable path to the asset.

What we measure

GEO success shows up in three places: the assistant cites the site with the correct URL, the citation is to a real icon rather than the homepage, and the described license matches the published rules. Monitoring search console impressions for blog and category pages, reviewing assistant outputs quarterly, and watching API usage all matter more than chasing a single metric.

The practical sequence for an icon site is: ship unique server-rendered pages, add honest structured data, publish llms.txt, expose a clean API, and keep every fact in sync across all three. That is the entire short version of GEO for a library like this one.