Discoverability

GEO for AI-Built Apps: How to Get Cited by ChatGPT, Claude & Perplexity

AI app builders solved how to ship in ten minutes. They did not solve how to be found. There's a fix, and it takes about an hour.

Bnome May 27, 2026 9 min read
A polaroid-flash diagram of a funnel extracting a single structured node from chaotic unstructured data

You spent the weekend wiring up a new idea. You prompted a builder, hit deploy, watched the URL go green.

Crickets.

This is the default experience for an app shipped in 2026. The build problem got cheap. The being-found problem got harder, because the rules quietly changed underneath everyone.

The numbers nobody put on a slide

Search is splintering. Google still dominates raw query volume, but its share of the wider "discovery" landscape — the universe of places people ask questions to get answers — dropped from 89.3% to 57.6% between December 2022 and December 2025. AI assistants now generate sessions roughly equivalent to 56% of global search engine volume. ChatGPT alone accounts for traffic comparable to about a fifth of global search.

The form of the answer is different too. Google returns ten blue links and you pick one. ChatGPT returns one paragraph and cites three sources. If you're not in those three, you didn't lose the click. You weren't in the conversation.

This is what GEO — Generative Engine Optimization — tries to solve. The term comes from a 2023 paper out of Princeton (published at KDD 2024), which measured the techniques that move citation likelihood in generative responses. They found visibility lifts up to 40%, and citation lifts of 115% for mid-ranked pages, just from structural and rhetorical changes to the source content. Same product, same authority — different chance of getting quoted.

Why AI-built apps lose this twice

Apps that ship in ten minutes also tend to skip ten things. Title tags. Meta descriptions. A sitemap. The first miss is the traditional SEO layer. Most builders have heard of that and feel vaguely guilty about it.

The second miss is GEO. The output of an AI builder is usually a wall of nested generic containers — div soup — that compiles into a beautiful interface for humans and a featureless blob for any model trying to read it. The crawler doesn't see the gradient borders. It sees a flat text stream, and if that stream has no structural anchors, it leaves with nothing to cite.

A live URL is a building with no address on it. Anyone inside can use it. Nobody outside can find it.

This isn't a complaint about AI builders specifically. Lovable, Bolt, v0, Replit Agent, Bnome — none of them treat discoverability as table stakes yet. The first one that does will quietly out-ship everyone else over the long run, because the apps coming out of it will compound traffic instead of vanishing into the void.

Five things that move the needle

1. Structured data with Schema.org JSON-LD

Schema markup is the cleanest GEO signal available. A Nature Communications study from February 2024 found that LLMs extract information significantly more accurately from structured inputs with defined fields than from unstructured HTML — the difference between handing the model a filled-in form versus a blank page.

The schemas that move citation odds for most apps: SoftwareApplication, FAQPage, Organization, and Article if you publish content. Drop a block like this into your <head>:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "SoftwareApplication",
  "name": "YourApp",
  "description": "What it does in one sentence.",
  "applicationCategory": "BusinessApplication",
  "offers": { "@type": "Offer", "price": "29", "priceCurrency": "USD" }
}
</script>

That's it. Five minutes. Permanent.

2. llms.txt — the AI-readable brochure

An emerging standard, proposed by Jeremy Howard of Answer.AI in September 2024. The idea is a Markdown file at yoursite.com/llms.txt that hands models a clean, pre-digested summary of your site — bypassing the HTML, the ads, the JavaScript. Think of it as a brochure written for machines.

The spec is short and the format trivial:

# YourApp

> One-sentence description of what it does.

## Key features
- Feature one
- Feature two

## Pricing
- Free
- Pro: $29/mo

Adoption is real and growing. Documentation tools like Mintlify generate it automatically, and Python libraries have been adding it through 2025. It is also the cheapest GEO win available — fifteen minutes, zero ongoing cost.

3. Semantic HTML over div soup

Crawlers extract weight from tag names. Text inside a nav is treated as routing. Text inside an article is processed as the core claim of the page. Text inside fifty layers of <div class="flex flex-col gap-4"> is processed as nothing in particular.

Most AI builders default to div soup because component frameworks encourage it. Go back to your prompt window and tell the agent to refactor the page using HTML5 semantic tags: main, article, section, nav, aside, h1 through h3. Thirty seconds of work. Massive difference in how a generative crawler sees you.

4. FAQ blocks in real Q&A format

This one is almost cheating. Models are trained primarily on instruction-response pairs, so a page with literal Q&A structure maps directly onto how they want to read.

Write the actual questions someone would ask a chat window, not the questions a marketing team would invent. Use real <h3> elements for each question and answer them directly in plain language. Then pair this with FAQPage JSON-LD, and you have doubled your surface area for citation.

5. Mentions on the corpora that LLMs actually trust

Here's where most people get the wrong idea. Generic backlinks barely matter for GEO. What matters is being mentioned on the sources LLMs disproportionately weight.

That list is shorter than you'd think. A June 2025 analysis of 150,000 LLM citations found Reddit was the source 40.1% of the time, with Wikipedia second at 26.3% and YouTube third at 23.5%. Inside Perplexity specifically, Reddit alone accounts for 46.7% of top citations. ChatGPT leans the other way and pulls 47.9% from Wikipedia.

Practical move: when you launch, write one genuinely useful comment in the most relevant subreddit. Not a promotional drop. An actual answer to an actual question, mentioning your tool once at the end. The thread will die in a week. The citation surface will keep working for months.

What "shipping" should actually mean now

The old finish line was a live URL. The new finish line is a live URL that other systems can read.

That means a domain, OG tags that look right when someone pastes the link into a chat, a favicon that renders in tabs, and the GEO surface above. Most of it is generated once and forgotten. It pays for itself the first time a model picks you out of the answer pool — which, given that this happens via real-time web retrieval inside tools like Perplexity, can happen the day after you publish.

Sixty minutes, in order

  1. Generate a SoftwareApplication JSON-LD block from your homepage copy. Paste into <head>. Five minutes.
  2. Write a 200-word /llms.txt with your one-line pitch, three features, and pricing. Ten minutes.
  3. Refactor your landing page markup to use semantic tags instead of nested generic containers. Twenty minutes.
  4. Find the three real questions your users ask. Write a FAQ section with those questions as <h3> headings, answer them directly. Pair with FAQPage JSON-LD. Fifteen minutes.
  5. Pick the most relevant subreddit. Find one current thread where someone has the problem you solve. Write a real answer. Mention the tool once. Ten minutes.

Sixty minutes for a layer that keeps working as long as the app exists.

The internet stopped being a list of links a while ago. The companies still optimizing only for blue results will keep getting them. Everyone else is being read by something that doesn't return lists at all — and the rules for that game are already on the table.

Build apps that ship — and get found

Bnome generates the app and the discoverability layer. Live URL, OG tags, structured data, all done.

Try Bnome