Most AEO pitches sound like this: "AI assistants are the new search engine. If you're not optimized for them, you're invisible."
That's true. But there's a harder argument that nobody's making yet, and it comes from a benchmark published last month.
Reflex compared two approaches to the same automated task. Path A: a vision agent that navigates your website by screenshot. Path B: an API agent that calls your endpoints directly.
550,000 tokens vs. 12,000 tokens. 53 steps vs. 8 calls. 1,000 seconds vs. 19.7 seconds.
The cost difference is 45x. And it's structural. Not a model problem, not a configuration problem. Vision agents must render every intermediate state to interpret it. Better models reduce error rates but can't reduce step counts. The gap is baked into the architecture.
What This Means for Your Business
AI agents are now software workers with token budgets. A token is a unit of cost. Businesses that make their data easy to extract pay 12,000 tokens to serve an agent. Businesses that don't pay 550,000, or don't get served at all.
This isn't about whether your business appears in a ChatGPT answer. It's about whether agents can afford to use you.
Here's how the decision plays out in practice: an AI agent searching for a dental clinic to recommend needs clinic details, opening hours, services offered, and a way to book. If your site has LocalBusiness schema and an API endpoint, the agent retrieves it in 8 calls. If your site has a JavaScript-rendered React page with no structured data, the agent either extracts it using 45x more compute, or moves to the next result that's cheaper to read.
At scale, frugal agents don't use expensive websites.
What "Structured Data" Means in 2026
When we say "optimize for AI agents," we mean four layers:
1. Schema markup. JSON-LD structured data that describes your business, services, FAQs, hours, and reviews in machine-readable format. Agents extract this without rendering anything.
2. llms.txt. A simple text file at your domain root that tells AI systems what you do, who you serve, and what content they can use. Like robots.txt — but written for language models.
3. Agent card. A .well-known/agent-card.json file that describes your business as an agent-callable service. Enables agent-to-agent discovery.
4. No AI crawler blocks. Agents respect robots.txt. Most businesses have no idea whether GPTBot, ClaudeBot, or PerplexityBot are blocked on their site. If they are, agents see nothing.
None of this is technically hard. All of it is currently neglected by most Norwegian SMBs.
The Real Cost of Doing Nothing
The 45x cost difference doesn't mean your business is taxed more. It means your business costs more to serve. So agents serve you less.
As agent usage grows and token costs stay real, this creates selection pressure. Structured sites get recommended more. Unstructured sites drift toward invisibility. Not because AI is biased against them. Because economics favor the efficient path.
You don't have to believe AEO is "the future of search" to care about this. You just have to believe that if AI agents become common tools for recommendation and discovery (which they already are for some audiences), the businesses they can efficiently read will have a structural advantage over those they can't.
What We Do
Synlig Digital runs a live 100-point AEO audit on your site. We check all five layers: schema markup, meta signals, content structure, technical foundation, and AI discoverability signals. You get a grade (A-F), a component breakdown, and ranked recommendations.
Most Norwegian SMBs score between 40-65. Most of the gap is fixable in a single engagement.
Try the free audit: https://aeo-mcp-server.amdal-dev.workers.dev/audit?url=yoursite.no
Or reach out: hei@synligdigital.no
Appendix: The Benchmark
The original research is from Reflex (April 2026): Computer Use is 45x More Expensive Than Structured APIs
The benchmark tested a browser-use vision agent vs. an auto-generated API agent on an identical admin panel task. Same model (Claude Sonnet), same application, same task.
Vision agent: 550,976 input tokens, 53 steps, ~1,000 seconds, failed first attempt (missed paginated content below the fold).
API agent: 12,151 tokens, 8 calls, 19.7 seconds, succeeded on first attempt.
The authors' framing: "An agent that must see to act will always pay for the seeing, regardless of how good the model gets."
This framing holds for external websites too. The structured API is what AEO builds. The vision cost is what businesses without AEO impose on every agent that tries to use them.
Top comments (0)