AI가 실제로 추천하는지 여부에 대해 100개 기업을 득점했습니다. 대형 브랜드가 인디에게 지고 있습니다.

Classic SEO measures whether you rank on Google. But AI answers are quietly eating the ten blue links, and that raises a different question with different signals: when someone asks ChatGPT, Claude or Google AI for a recommendation, can the model actually understand, trust and cite your business?

I wanted real data instead of vibes, so I ran a controlled audit of 100 businesses — 25 local, 25 national, 25 international, 25 niche B2B/SaaS — scoring how well each site gives AI systems what they need to recommend it. (Audit date: 2026-05-31.)

Full disclosure up front: I build a tool that does this scoring (GetVisus). But this post is the data and the takeaways, not a pitch — you can act on every finding here without any tool.

The surprising part: brand size didn’t save anyone

• Bold Street Coffee, Mowgli, a local charity → 100/100
• Pret A Manger → 32. Greggs → 52. giffgaff → 53. Smarty → 29
• A well-known local venue → 14/100

Category averages:

Category Avg score International 91.5 B2B / niche SaaS 81.1 Local 79.8 National brands 75.5 (worst)

Big national brands often have beautiful, JS-heavy sites that are hard for an LLM to extract facts from — while a tiny coffee shop with a clear, text-first page wins outright.

The most common problems (out of 100)

Problem How many sites No FAQ / no answer-ready sections 79 Reviews/testimonials not visible as on-page text 70 Pricing not visible 49 Business not recommended for its own core prompt 29 About / who-you-are unclear 27 Weak or missing structured data 9

What actually moves the needle (all free to do)

1. Put a real FAQ / Q&A block on key pages, phrased the way customers actually ask. Answer-ready text is the single biggest gap.
2. Show reviews and testimonials as text on the page — not just a Trustpilot logo an LLM can’t read.
3. Make pricing, or at least a price range/scope, visible. “Contact us” tells an AI nothing.
4. Make the entity obvious: who you are, what you sell, who it’s for, where you operate — in plain text near the top.
5. Add basic schema (Organization, LocalBusiness, FAQ) — but only for facts that are visible on the page.

The pattern across every 100/100 site was identical: the business entity, the offer, the proof, and the “who it’s for” boundaries were all trivially easy to extract. The low scorers made the AI guess — and models don’t guess in your favour.

The honest caveat

This is a directional benchmark on a controlled sample, not a market-wide study, and not proof of exact revenue causation. The scoring reflects a model of what LLMs reward, not ground-truth from the model providers. But as AI answers replace the classic search results, “can an LLM cite you?” is becoming a real commercial question — and most sites, including big ones, are quietly failing it.

Full methodology and the 100-row dataset are in the benchmark report. Happy to answer methodology questions in the comments.