Home / Blog / Research

Research

My AI visibility score didn't predict which businesses AI recommends. Here's the data.

By Zeb Choudhry · Visus · 25 August 2026 · 6 min read

When we published our benchmark of 100 businesses, the headline was that big brands scored surprisingly low for AI readiness while local independents scored high. Fair enough. But two people pushed back with the same sharp point, and they were right, so this post is the follow-up they asked for.

The argument: an on-page readiness score (can an LLM extract your entity, offer, proof and buyer-fit?) is one measurement. Whether an answer engine actually recommends you in a live buyer prompt is a different measurement. Do not conflate them. We had the data to test it, sitting in a spreadsheet, and had not printed it.

So we ran the cross-tab. Here is what it shows.

The readiness score of the businesses AI refused to recommend

Of the 89 businesses that completed a full live audit, 60 were recommended for their own core buyer prompt and 29 were not. Here is how those 29 scored on our 0-100 readiness score:

They do not cluster at the bottom. The mean readiness score of the recommended group is 84.7; of the not-recommended group, 77.6 - a gap of only 7 points. Nearly half of the businesses AI would not recommend scored 90 or higher for readiness.

High readiness, still not recommended

The most interesting subset. These sites are highly extractable - an AI can read exactly who they are and what they offer - and were still not recommended for their own core prompt:

• The Brain Charity - 100
• OVO Energy - 100
• Motorway - 100
• Userflow, Userpilot, Recurly - 99
• Hope Street Hotel - 94
• Marks & Spencer - 92
• Starling Bank, Salesloft, Delifonseca Dockside - 90

Two patterns run through this group. First, pricing, plan or scope was not extractable - the page is readable, but it does not state the commercial terms an AI needs to make a confident recommendation. Second, there was no concise buyer-intent Q&A - the site does not directly answer the question a buyer would actually ask. In other words: readable, but it does not answer the purchase question.

Low readiness, recommended anyway

The mirror image is just as telling. Seven businesses scored under 60 for readiness and were recommended:

• Pret A Manger - 32
• Greggs - 52
• Liverpool Philharmonic - 54
• Everyman Liverpool - 56
• Moneybox - 57

Almost all household names. That points to something on-page work cannot touch: the model already knows these brands from its training data, so a thin live page matters far less. Familiarity carried the recommendation. (This is the retrieval-versus-training-data split we wrote about here - Pret does not need a great page to be recommended, because ChatGPT already knows what Pret is.)

What this actually means

On our own sample, the readiness score is necessary-ish but nowhere near sufficient. Being extractable is a floor, not a guarantee. Recommendation needs additional layers on top:

• Query fit - does the page answer the specific buyer prompt, not just describe the business?
• Commercial clarity - visible pricing, plan, scope, eligibility.
• Third-party corroboration - reviews, listings, and mentions the model can cross-check.
• Brand familiarity - whether the model already knows you (training data), which the biggest names get for free and everyone else has to earn.

So we were wrong to imply readiness drives recommendation in a straight line. The honest model is two separate measurements: a site-readiness score, and a live prompt-level outcome tracked on its own - mentioned, cited, shortlisted, recommended, reason given, source used. A 100/100 readiness site is not the same thing as "AI will recommend this business," and the data proves it on a hundred real sites.

That is a better way to think about AI visibility, and it is the direction we are taking our own scoring: readiness and live recommendation, measured and reported separately, never collapsed into one number.