How ChatGPT Decides Which Brand to Recommend (and Why It Is Not Reading Your Website)

It is Friday. Somebody in Oakville is standing in the LCBO with their phone out and types "good Canadian gin for a negroni under 50 bucks." The answer comes back in four seconds. Three brands, a sentence on each, one sounds like it was written by a bartender who likes them.

None of the three is yours.

There is no link to click. Nobody visited anybody's website. Your analytics for that Friday will show a perfectly normal day. And a bottle you spent two years getting onto that shelf just got skipped by a person who was standing three feet from it.

That is the part of AI search nobody is measuring, and it is the part I want to talk about.

The path your dashboard cannot see

My business partner Hayden at Plate Lunch Collective just published a long piece on this. His argument is that AI search has six buyer paths and your website sees one. It is worth the 27 minutes if you run marketing for a brand. Path three is the one above: the AI recommends you (or your competitor) without citing a page, so there is no click, no referral, nothing to attribute. He calls it recommendation without citation.

I am going to take that one path and answer the question a beverage or CPG brand cares about. When the machine picks a brand and does not link to anyone, who was it listening to?

Because it was not you.

First, the stakes, so you know this is not a niche problem

Two in five Canadians now bring generative AI into their purchasing decisions, and among those users, 31 percent use it specifically for buying products or services, with 80 percent saying it helps them compare options. That is a Léger survey from June, not a vendor deck.

A separate Omnisend survey reported that more than half of Canadians have used AI for a shopping task, and 48 percent said ChatGPT gave them better recommendations than a search engine. South of the border it is further along: 77 percent of US consumers used AI to help with a purchase in the last six months.

Here is the number that should bother you. In G2's 2026 buyer study, 69 percent of buyers chose a different vendor than they originally planned because of what an AI chatbot told them, and one in three bought from a brand they had never heard of. That is software, not gin. But the mechanism is the same. The assistant is not confirming what people already wanted. It is changing the pick.

So: who is it listening to when it changes the pick?

The sources, ranked

Peec AI looked at 30 million sources that were directly cited across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews. The top of the list was Reddit, then YouTube, LinkedIn, Wikipedia and Forbes, with review platforms showing up heavily on recommendation questions. Brand websites are not in the top ten. They are barely in the conversation.

Muck Rack went further and classified 25 million links that ChatGPT, Claude and Gemini cited. 84 percent of AI citations were earned media, meaning content the brand did not write and did not pay for. Paid and advertorial content came in at 0.3 percent. Not 3 percent. Zero point three.

Then Ahrefs ran the study that matters most to anyone who makes video. They took 75,000 brands and tested which signals predict whether a brand gets mentioned in ChatGPT, AI Mode and AI Overviews. The strongest signal in the whole dataset was YouTube mentions, at a 0.737 correlation. Their definition of a YouTube mention: the brand's name appearing in a video title, transcript or description. Branded web mentions came next at 0.66 to 0.71. Backlinks, the thing SEO agencies have sold for twenty years, sat at 0.218. The number of pages on your site was 0.194, which is a polite way of saying the machine does not care how big your website is.

Read those three studies together and the picture is not subtle. The AI is deciding who to recommend by reading what other people say about you, and the single most predictive place they say it is on video, out loud, where the transcript gets indexed.

Why video? Because that is what the models were fed. The New York Times reported that GPT-4 was trained on more than a million hours of YouTube transcriptions, and Google's AI surfaces cite YouTube more than any other domain. To a language model, a creator saying your product name into a camera is just text. Very trusted text.

I have seen this before, through one way glass

Before I made videos for a living I spent 25 years in software, a lot of it running usability tests. You put a person in a room, give them a task, and watch through the glass while they try to use the thing you built.

The finding that never changed, across every product and every decade: nobody reads the manual. They ask the person next to them. If the person next to them does not know, they ask a forum, a friend, a video. The official documentation was the last place anyone looked, and we had spent months writing it.

AI search is the person next to them. Except this person has read every other person next to them, on every forum and every video, and formed an opinion before you walked in.

Your website is the manual.

The reframe: you are not being ranked, you are being referenced

Old search ranked pages. You could win by making your page the best page. AI recommendation does something closer to what a hiring manager does. Your website is the resume. It gets a glance. The decision is made on references, and the references are whatever the machine can find other people saying about you.

That changes what "content" means for a brand. A product page is not a reference. A creator holding your product, saying the name, saying what it tasted like and what they did with it, in a video with a title and a description that repeat the name, is a reference. A review thread is a reference. A bar manager on YouTube making a cocktail with your gin and mispronouncing it is, unfortunately, also a reference.

I wrote earlier this year about how to be the AI answer with video. This post is the why behind it. The machine's reference check runs on third party video and community discussion, and creator content is the one channel where a brand can put real references into the world on purpose, at scale, without faking it.

The scoreboard problem

None of this shows up in Google Analytics, and that is the trap.

Semrush studied nearly 4,000 brand appearances in AI answers and found that 25 percent were brand mentions with no citation at all, and 62 percent were citations with no brand mention. Getting linked and getting recommended are two different events, and the one that sells product is the one your analytics cannot see.

So measure against the goal, which is the same goal it has always been: selling more. Views are inventory, not results, and I have said that at length in what four million views a month is really buying you. For AI recommendation, the leading indicators are different. Are you being named when someone asks the category question? Is branded search volume moving? Are retailers telling you people are asking for it by name? Brand lift feeds the sale. Watch the lift, then watch the sale, and stop looking for a referral line that will never appear.

Test it like a product, not like a press release

Open ChatGPT, Gemini and Perplexity. Ask each one the ten questions your buyer would ask. "Best Ontario craft gin." "What hard seltzer is actually low sugar." "Good non alcoholic beer for a Super Bowl party." Write down who gets named. Do it again in 30 days.

One check is not a test. It is a single data point pretending to be a pattern, and in this channel the pattern moves fast. Semrush tracked 230,000 prompts over 13 weeks and watched ChatGPT's Reddit citations drop from close to 60 percent of responses to about 10 percent in a matter of weeks after one Google change. If the sources can swing that hard, your position in the answer can too.

Same rule I use for creative. Run it small, see what moves, adjust, run it again and try to beat your own winner. I laid that out in does UGC actually work. The testing loop for AI visibility is the same loop with a different scoreboard.

How to get on the list

This is the part where I earn the post. Everything below is built from the Ahrefs, Peec and Muck Rack findings above, applied to a beverage or CPG brand.

1. Get your name said out loud, on video, by people who are not you. The Ahrefs signal is brand name in title, transcript or description. That means the creator has to say the name clearly (not "this gin"), the title should carry it, and the description should carry it again. I brief this into every shoot now. It costs nothing and most creators skip it.

1. Many creators beat one big creator. A single video is a single reference. Ten creators across ten channels saying the same product name in ten different contexts is a reference file. The model is counting mentions and the breadth of where they appear, not one creator's follower count.

1. Answer the question people actually ask. The machine is matching your product to a query like "gin for a negroni" or "low sugar seltzer." A creator video that says the brand name and answers that specific question, in plain language, is a direct hit. A brand anthem video is not.

1. Show up where the machine already reads. YouTube first, because the transcript is indexed and the correlation is the strongest. Then the community and review surfaces in your category. For beverage that is the review apps, the cocktail forums, the retailer reviews. You do not fake these. You give people a reason to post, and you make sure the product facts they need are easy to find and consistent.

1. Keep the facts identical everywhere. ABV, sugar, size, price range, where to buy. If the machine finds three versions of your product spec across the web, you are harder to recommend with confidence. Boring work, high leverage.

1. Be honest about sponsorship, and understand the machine treats it differently than Meta does. Muck Rack's 0.3 percent number is about paid placements and advertorials. A sponsored creator video on YouTube still counts as a YouTube mention in the transcript, and you are still legally required to disclose it in Canada. Disclose it properly. If you sell alcohol or cannabis, read the rules first, because the same video that earns you an AI mention can get an account wiped if the disclosure and age gating are wrong. I mapped every province and platform for alcohol and for cannabis.

1. Stop measuring backlinks as if they are the job. A 0.218 correlation is not nothing, but it is a third of what a mention is worth. If your agency's monthly report leads with link counts, ask them for the mention count instead.

Red flags

• Anyone guaranteeing you will be "the AI answer." Nobody controls this. I wrote about the lies marketers tell about AEO, and the guarantee is lie number one.
• A content plan that is entirely on your own domain. That is the resume with no references.
• Creator videos where the product name is never spoken, or the title says "my Friday cocktail" and nothing else.
• A single creator, a single video, and a report that calls it a campaign.
• An AI visibility report from one day, on one engine, with no rerun.

What good looks like

The brands I see getting named are not the ones with the biggest sites. They are the ones that have been talked about, in a lot of places, by people with no reason to lie. Creator content done properly is that, manufactured honestly: real people, real use, the name said clearly, the question answered, the video titled so a machine can read it, and then done again next month with a new angle.

My standard has not changed since the usability lab. Brief with a reason. Say the name. Answer the real question. Test small, measure against the sale, run it again. And only promote products I would put in my own glass, because the day a creator recommends something they do not use is the day their references stop being worth anything, to a person or to a machine.

Questions to ask before you spend

• When someone asks an AI assistant the category question, do we get named? On which engines?
• Where does our brand name appear in video transcripts, titles and descriptions today, and who put it there?
• How many independent sources are talking about us, versus how many pages we have published about ourselves?
• Are our product facts consistent across every place the machine can find them?
• Are we rerunning the prompts monthly, or did we check once and call it a strategy?
• Is the creator content briefed to say the name and answer a real question, or is it a vibe with a logo?
• Is the disclosure right for the category, so the video that earns the mention does not cost the account?

Any creator or agency you hire should be able to answer those before you sign anything. I answer them up front, because that is the whole point.

The machine has already decided who it likes. The only question is whether it has heard of you.

Let's talk.