AEO: how LLMs find and recommend your business

The Shift From Search Engines to Answer Engines

Something unusual started happening at BlackBox Vision early this year. Qualified leads began showing up through a channel we hadn't planned for: LLMs.

Twenty leads in three months. Every time we asked how they found us, the answer was the same — "ChatGPT told me about you," or Gemini, or Claude. Some of them even had our direct contact details handed to them by the model.

This didn't happen by accident. Last year, we noticed a paradigm shift forming. People were moving from searching on Google to asking AI models for recommendations. We decided to get ahead of it.

What is Answer Engine Optimization?

Many are now calling this AEO — Answer Engine Optimization. For us, it's really SEO with a technical and strategic twist. The core idea is simple: if LLMs are becoming the new front door to your business, your site needs to be readable and meaningful to them.

Traditional SEO optimizes for crawlers and ranking algorithms. AEO goes further — it optimizes for comprehension. You're not just trying to rank; you're trying to be understood, cited, and recommended.

The Technical Playbook

When we redesigned the BlackBox Vision site, aesthetics weren't the priority. Semantic clarity was. Here's what we actually implemented — no theory, just what's running in production today.

Strict HTML Semantics Across 80+ Pages

Every single page — 40 in English, 40 in Spanish — uses proper semantic elements: <header>, <nav>, <main>, <section>, <article>, <aside>, <footer>. No <div> soup. We also added 543+ ARIA attributes across the site: role="group" on interactive regions, aria-label on every button and navigation element, aria-expanded states on toggles, and aria-current="page" on active nav links.

Every page includes a <a href="#main" class="sr-only">Skip to main content</a> link. Over 554 images carry descriptive alt text. This isn't just accessibility compliance — it's machine-readable context that LLMs use to understand what's on the page.

Layered Structured Data (JSON-LD)

We didn't just sprinkle one schema type. We implemented multiple layered schemas across the site:

Organization schema on the homepage: includes founder information with LinkedIn URLs, an aggregate rating (5.0/5 from client reviews), individual review objects with author names, our physical office location with GPS coordinates, countries served, a full services catalog via OfferCatalog, contact points, and links to six social profiles via sameAs.
BreadcrumbList schema on every inner page: proper hierarchy from Home → Section → Page. On case study pages, this pairs with an Article schema that includes headline, description, author, publisher, and language declaration.
FAQPage schema on service and about pages: structured Q&A pairs that LLMs can extract and cite directly.

This layered approach means models don't just know we exist — they know what we do, who runs it, what clients say about us, and how to navigate our content.

Complete Meta Tags and Open Graph

Commercial pages carry the meta suite we want models and social crawlers to see: description, Open Graph fields (og:title, og:description, og:type, og:url, og:image with image dimensions and MIME type where available), Twitter Cards with summary_large_image, content language declaration, and page-specific preview images. We also run a build check so pages with an Open Graph image keep the matching twitter:image tag.

These aren't just for social cards. When a model encounters structured metadata, it uses it to build a summary of the page before parsing the body. Good metadata is like handing the model an executive summary.

Bilingual Architecture with hreflang

We serve 37 page pairs in English and Spanish using a subdirectory approach (/es/services/mvp-builders/). Every page pair is connected via proper rel="alternate" hreflang attributes in the HTML, and the sitemap includes full hreflang entries with x-default fallbacks.

This matters for AEO because models serving users in different languages need explicit signals about which version to reference. Without hreflang, a Spanish-speaking user might get recommended the English version — or worse, neither.

LLM-Specific Discovery Files

This is where we added a secondary discovery layer for non-Google AI assistants. Google has since clarified that its generative Search features do not require llms.txt, special AI markup, or AI-only text files. For Google, the source of truth is still the crawlable, indexable HTML page and the normal SEO signals around it. We still generate two files because they are useful context for other AI crawlers and assistants:

/llms.txt — A markdown-formatted overview of the company, services, and page index. Think of it as a structured README for AI crawlers.
/llms-full.txt — A comprehensive context file with our message hierarchy, service descriptions, page index, related surfaces, and founder profiles.

These files are generated automatically at build time. When a compatible assistant crawls our domain, it can find a clean, machine-readable summary of everything we do. The HTML pages, sitemap, metadata, and visible content remain the canonical source.

Clean URL Structure and Internal Linking

Our URL hierarchy is RESTful and descriptive: /services/mvp-builders/, /case-studies/reduc/, /about-us/. No query parameters, no cryptic slugs. Every service page links to related case studies. Every case study links back to the relevant service. Navigation uses descriptive anchor text — never "click here."

Combined with BreadcrumbList schema, this creates a navigation graph that models can traverse to understand relationships between our offerings.

Permissive Robots and Structured Sitemap

Our robots.txt is three lines: allow everything, reference the sitemap. No disallow directives. Our XML sitemap includes priority tiers (1.0 for homepage, 0.8 for main sections, 0.7 for sub-services, 0.6 for individual case studies) and weekly/monthly change frequencies. Every entry includes hreflang alternates.

We explicitly welcome AI crawlers. If you want to be recommended, you can't block the models from reading you.

Why This Matters Now

Google is still important. But the discovery funnel is branching. When a founder asks ChatGPT "Who can build my MVP?", you want to be in that answer. And unlike paid ads, you can't buy your way into an LLM recommendation. You earn it through clarity, structure, and substance.

The signals that make your site recommendable to AI are largely the same ones that make it excellent for humans — clear value proposition, useful non-commodity content, well-organized information, and technical execution that doesn't get in the way of comprehension.

What You Can Do Today

If you want your site to be easier for search engines and AI assistants to understand, here's a prioritized checklist based on what actually worked for us:

Create non-commodity content — Add firsthand experience, proof, specific tradeoffs, examples, and strong opinions that a generic model summary would not have.
Make pages crawlable, indexable, and snippet-eligible — Keep canonical URLs, sitemap entries, robots configuration, and metadata clean.
Implement layered structured data — Start with Organization schema (include founders, reviews, services, contact), then add BreadcrumbList on inner pages and FAQPage where natural and visible on the page.
Audit your HTML semantics — Replace unnecessary <div> wrappers with <header>, <nav>, <main>, <section>, <footer>. Add ARIA attributes and descriptive alt text where they help users and machines understand the page.
Set up bilingual hreflang — If you serve multiple languages, connect page pairs with rel="alternate" hreflang in both HTML and sitemap. Include x-default.
Clean your URL structure — RESTful, descriptive, hierarchical. /services/mvp-builders/ beats /s?id=42.
Interlink with intent — Every page should connect to related content via descriptive anchor text. Pair this with BreadcrumbList schema for a complete navigation graph.
Add llms.txt and llms-full.txt as optional support — Use them as structured summaries for compatible AI assistants, not as a replacement for strong pages or as a Google AI requirement.

The potential of getting ahead of how machines read your business is enormous. And the window to build this advantage is still wide open.