Getting into the AI consideration set is a different problem from winning it

One of the more useful AEO studies published this month starts with a simple question: which businesses does AI actually recommend? Researchers at Norly Research catalogued every restaurant, café and bar they could identify across Canggu and Ubud in Bali, producing a census of 4,776 venues. They then collected 2,208 search-grounded responses from ChatGPT, Claude, Gemini and Perplexity across 96 persona-based queries over seven days. The headline finding is stark: 85.6% of businesses were never recommended by any of the four systems. Even among established venues with 50 or more ratings, 72.6% never appeared. [1]

The more consequential finding is not the scale of invisibility. It is the structure beneath it. The study suggests AI recommendation has two different competitive margins: getting into the consideration set and winning once inside it.

<h2>Documentation appears to admit, reputation appears to rank</h2>

The study found that the factors associated with appearing in an answer differed from those associated with taking the first position within an answer. At the entry margin, a venue's own website, review volume, listed pricing and third-party web mentions were all positively associated with inclusion. The reported odds ratios were 1.92 for an own website, 1.64 for review volume, 1.54 for price information and 1.44 for third-party web mentions. Star rating, by contrast, was not statistically significant at this stage. [1]

Once a venue had already been selected for recommendation, the pattern changed. Higher star ratings significantly predicted whether it appeared first in the answer. The authors describe this as a two-margin structure: documentation admits, rating ranks. The study is observational, so it does not show that adding a website or publishing prices will cause an AI system to recommend a venue. It does show that these documentation signals were meaningfully associated with escaping invisibility in a complete local-market audit. [1]

<h2>Eligibility and preference are different AEO problems</h2>

This distinction gives AEO a more useful diagnostic language. Before an answer engine can decide whether a restaurant, insurer, university, software platform or financial adviser is the best option, it needs enough reliable information to determine that the organisation belongs in the candidate set at all. It needs to understand what the organisation is, whom it serves, where it operates, what it offers, what it costs and whether independent evidence supports its legitimacy.

That is an eligibility problem. It is principally a problem of documentation, entity clarity, coverage, evidence and retrieval. A brand with a strong reputation may still be absent if its information environment is too thin for a system to confidently consider it. Preference is a different problem. Once the brand is present, reviews, reputation, comparative proof, relevance and fit may determine whether the system places it ahead of alternatives.

<h2>Why simple visibility measurement can mislead</h2>

Most AEO reporting starts with a basic question: did the brand appear? This research suggests that one measure can collapse two different performance states. A brand that rarely appears may have an evidence and documentation problem. A brand that appears regularly but is seldom named first may have a preference problem. A single visibility score can identify neither the bottleneck nor the right intervention.

The more useful sequence is therefore: is the brand eligible for consideration, does it appear in relevant prompt domains, and when it appears, does it win the recommendation? That sequence aligns with the claim-audit and query-lattice work emerging in AEO. Brands need clear answer paths and proof to enter retrieval, then differentiated evidence to earn the recommendation.

<h2>There is no universal AI consideration set</h2>

The same study also found low agreement across systems. Top-20 Jaccard similarity among ChatGPT, Claude, Gemini and Perplexity ranged from 0.33 to 0.54. In practical terms, the four systems did not reliably construct the same shortlist of venues. That reinforces the need for multi-model measurement. A business can be eligible in one system's retrieval environment and largely absent in another, even when the public information available about the business has not changed. [1]

<h2>Staleness may be the more practical failure mode</h2>

The authors found outright fabrication to be rare: 0.08% of valid business mentions were likely invented. But the systems recommended permanently closed venues 93 times. That result shifts attention from a familiar concern about hallucination to a more operational concern about freshness. For local businesses and multi-location brands, information accuracy is not simply a listings-management task. It is part of AI representation management. Hours, locations, pricing, availability, product status and business changes all affect whether an answer remains useful after a brand has become eligible for consideration. [1]

<h2>Limits matter</h2>

The study covers food-and-drink venues in two tourism-oriented Bali markets, using English-language, persona-conditioned queries. Its precise invisibility rates should not be generalised directly to other industries or geographies. It was also funded and conducted by Norly, a company in the AI-visibility market, although the paper discloses the commercial interest, a pre-registered protocol and validation procedures. The strongest takeaway is therefore conceptual rather than universal: entry into an AI answer and preference within that answer can be different analytical margins, and AEO programmes should measure them separately. [1]

<h2>AEO Updates Takeaway</h2>

The headline number is that 85.6% of the venues in this audit were invisible to the four answer engines. The more useful lesson is why. Getting a brand into the AI consideration set and winning the AI consideration set are different problems. The first depends on sufficient, credible and current documentation. The second depends on why the available evidence makes the brand preferable. AEO measurement should begin by separating eligibility from preference, because they point to fundamentally different work.

<h3>References</h3>

[1] Vladimir Pitenin, <a href='https://arxiv.org/abs/2608.07069' target='_blank' rel='noopener noreferrer'>Invisible to the Machine: Auditing AI Restaurant, Café, and Bar Recommendation Against a Complete Market Census</a>, arXiv:2608.07069, August 2026.