How does an AI decide which business to recommend?

By Manish, SEO & AI Visibility Lead at DigiVino · Updated August 20, 2026 · 9 min read

When someone asks an AI for the best wine club that ships to Illinois, it is not judging wine. It is assembling an answer out of the most quotable facts it can find — and that is a competition you can actually enter.

The examples below are wineries and wine clubs, because a club makes the mechanics unusually visible. The mechanics themselves are identical for craft distilleries, specialty producers, boutique lodges, nonprofits, B2B tech consultancies and niche law firms.

What does an AI actually weigh when it names a business?

Five things, in rough order of leverage: a clear statement of what you are near the top of the page, specific evidence, structure it can parse, freshness, and corroboration from somewhere that is not your own website.

A clear definition, early. The first 150 words or so of a page carry outsized weight when a model summarizes it. "Founded in 1974 on a ridge above the valley" is lovely and tells a machine nothing. "Harper Ridge is a family-owned winery in Napa Valley with a four-tier wine club shipping to 38 states" is a sentence it can lift whole.

Specific evidence. Numbers, dates, named sources, prices, quantities. Anything a machine can repeat without having to interpret it.

Structure it can parse. Question-shaped headings, plain paragraphs that answer them, and schema markup underneath telling the machine what each fact is. Models retrieve in chunks; a page organized into self-contained answers gives them clean chunks to take.

Freshness. Dated pages beat undated ones, and a citation's hold fades surprisingly fast without updates. A club page that still shows 2024 pricing signals abandonment whether or not the price changed.

Corroboration. An engine is more willing to name you when your own site and other people's pages agree. Reviews, an accurate Google Business Profile, directory listings, coverage, honest answers in forums. See E-E-A-T for how this gets formalised, and which reviews AI actually trusts.

Why does specificity win?

Because an AI is not choosing the best business — it is choosing the safest sentence. It has to produce a claim it can stand behind, in front of a user who might click through and check. "We're the best winery in the valley" is unquotable; there is nothing to stand on. "Our club is $210 a year for twelve bottles and ships to 38 states" is a fact, and a fact is exactly what the machine is shopping for.

That reframing is worth sitting with, because it explains a lot of otherwise baffling results. The winery that gets named is frequently not the best winery. It is the winery that made the most checkable statements. Persuasion is what you write for people; verifiability is what you write for machines. You need both, on the same page, and they turn out to be surprisingly compatible — buyers like checkable facts too.

Do all the AI engines pick the same businesses?

No, and assuming otherwise is the most common and most expensive mistake owners make here. Every engine has its own taste in sources. Conductor tracked citations across seven engines — ChatGPT, ChatGPT Search, Perplexity, Google AI Overviews, AI Mode, Gemini and Claude — from September 2025 through March 2026, and found only around 11% of cited domains overlap between ChatGPT and Perplexity. BrightEdge's benchmarking found the engines disagreed about which brands to recommend 62% of the time.

The practical consequence: checking one engine tells you about one engine. Plenty of owners have typed their name into ChatGPT, seen themselves appear, and concluded they were fine — while being entirely absent from the tool their actual customers use. Whatever you check, check in at least four places, and check the question a buyer would ask rather than your own name.

Do citations come from my own site, or from somewhere else?

Both, unevenly, and the split surprises people. Muck Rack's analysis of more than a million AI citations found 82% came from earned media — articles, coverage, directories, third-party pages — and 94% from non-paid sources of some kind.

The useful way to read that is not "my website doesn't matter." It is that the two do different jobs. Your own site is what makes you describable. It is where the price, the tiers, the shipping states and the terms live, and nothing else on the internet is going to supply those. Other people's pages are what make you credible. They are the corroboration that turns a claim into something the machine will repeat.

A winery with immaculate schema and no mentions anywhere gets described accurately and recommended rarely. A winery with glowing press and an unreadable club page gets recommended vaguely, without the details that convert. You want the pair.

How does an AI decide which wine club to recommend?

By filtering on facts first and quality last — which is the opposite of how most club pages are written. Listen to how the question is actually asked. Almost nobody types "best winery." They type something closer to: "a wine club under $250 a year that ships to Illinois, no long commitment, mostly reds."

That is four filters and a preference. Price. Shipping state. Commitment length. Varietal mix. And every one of those filters is applied to a stated fact. A machine cannot filter on something it cannot see, so a club whose price sits inside a JPEG is not judged and rejected — it is never considered. By the time quality enters the conversation, the shortlist has already been drawn.

This is the part worth carrying away: your wine is the last filter, not the first. Everything upstream of it is logistics and terms, written in plain sentences. Which means the club page has a job it was probably never designed to do — not to seduce, but to declare.

What that looks like in practice is unglamorous and quick. One plain, complete sentence per fact, high on the page:

• "The Cellar Club is $210 per shipment, four shipments a year, three bottles each."
• "Members save 20% on all wine purchases and 10% in the tasting room."
• "There is no minimum commitment; you can cancel any time from your account or by email."
• "We ship to 38 states. [List them.]"
• "Members can collect shipments at the winery instead, and bring two guests for a complimentary tasting."

Complete sentences matter more than they look. A three-column comparison table is excellent for humans and often loses its column headers when a machine extracts it — leaving "$210" floating with nothing attached. Keep the table if it helps your visitors; add the sentence underneath it for everyone else. And note that none of the above is a marketing decision. It is a transcription problem: facts you already say out loud, written down where something can read them.

The pattern generalises cleanly. A gym membership, a CSA share, a coffee subscription, an IT retainer — in each case the terms are the product, and in each case the terms are usually the least readable thing on the site.

What can I do myself, and what needs a developer?

The left column is real work with real returns and no technical risk. Start there.

The right column is where the leverage concentrates, and also where the failure modes are quiet. Schema that disagrees with the page can cost you the result it was meant to earn. An entity that reads as three separate businesses splits the credibility you spent years building. A bot rule can turn away a crawler your robots.txt just welcomed, and nothing tells you. These are not hard problems for someone who does them daily; they are unforgiving problems to learn on a live site during club renewal season.

How do I know whether it is working?

1. Ask buyer-shaped questions, not brand-shaped ones. "Wine clubs that ship to Illinois under $250" — not your name.
2. Ask in four engines, the same day. ChatGPT, Gemini, Perplexity, and Google's AI results. They disagree; that is the point.
3. Look for the name, not the link. Being named in the answer is the win. A link is a bonus.
4. Log it monthly. One page, four engines, three questions. Trends beat snapshots, and freshness fades — a result you earned in spring can quietly go.
5. Note what it gets wrong. Every wrong detail points at a fact your site never stated clearly.

If you would rather see the whole picture at once — every major engine, what each currently believes about you, and where your competitors are being named instead — that is what a free AI Visibility Check produces.

Sources: Conductor, seven-engine citation analysis, September 2025 – March 2026 (source overlap); BrightEdge benchmarking (brand-recommendation disagreement); Muck Rack, analysis of over one million AI citations (earned vs. owned split). Current as of August 2026.

← Back to Yield & Press