ChatGPT picks its brands before it searches. What new data means for CPG

Most AI visibility advice starts from one assumption: get your page cited and the AI will recommend you. Two pieces of research published in the last two weeks say that’s only half the story.

When ChatGPT searches the web to answer “what’s the best…”, it usually already has brands in mind before the first result comes back. The search mostly checks a shortlist the model built in training. If your brand isn’t on that shortlist, being mentioned on the pages it cites helps a lot less than you’d hope.

That’s a big deal for food, beverage, beauty and household brands, and the details are more hopeful for physical products than the headline suggests. Here’s what the data says and what we’d do with it.

What did Graphite find about ChatGPT and brand names?

Graphite, a growth agency that sells AEO services, published an analysis on October 16 comparing logged-out ChatGPT answers to comparison prompts with search and without it. Search Engine Journal covered it on October 26. Three numbers stand out:

• The no-search answer predicts the search answer. Across 3,115 brand and question pairs from 57 questions, brands ChatGPT rarely named without search were rarely named with it. Brands that appeared on at least 15% of the cited pages, but that ChatGPT named in under a quarter of its no-search answers, had a median visibility of just 11% when it searched.
• Brands go into the search query itself. Graphite counted about 957,000 background searches (fan-out queries) from 22,953 prompts that named no brand. 59.6% of those queries contained a brand anyway, and 22.7% used a site: operator.
• ChatGPT cites far more often than it did. On comparison prompts, the share of logged-out ChatGPT answers with citations rose from 56.6% in the week of August 3 to 88.5% in the week of September 28.

Graphite’s argument is that brands should work on what models learn in training, not only on citations. It is correlational data from a vendor with something to sell. Still, the direction is hard to ignore.

Did anyone reproduce it?

Yes, on a small scale. On October 27, SaaS Link Building ran 20 unbranded prompts through OpenAI’s gpt-6-luna model via an API, once with search on and once with it off. 77 of the 82 fan-out queries (93.9%) named a brand. For the top pick, search on and search off agreed in 15 of 20 prompts.

The part CPG teams should read twice is the split by product type. Sixteen prompts were about software, four about physical products: running shoes, robot vacuums, noise-cancelling headphones and meal kits.

For software, search worked like a fact check. The model looked up pricing pages for vendors it had already picked. For physical products, search changed the answer. Search swapped stale model names the model remembered for current ones (Brooks Ghost 17 became Ghost 18), and review sites carried real weight. Only 50% of the physical products in the searched answers also showed up with search off, against 90% for software.

Four prompts isn’t a market-wide rate, and the authors say so. But it fits what we see with consumer products: shelves change, flavors launch, packs get resized, and the model has to go looking.

How much does being a known brand matter?

A lot. The largest dataset on this comes from a June paper by Ranqo, another visibility vendor, covering 102,025 AI answers across 102 brands between March and May 2026. On each brand’s first tracking run, for prompts that didn’t name the brand, visibility fell into three clear tiers.

Olipop sits in that middle tier, which is roughly where a lot of growing CPG brands live. The same paper found that only 2.9% of citations pointed at the brand’s own site, while ranked “best of” listicles made up 21% of all citations, and YouTube was the top non-corporate source at 4.2%. The author co-founded Ranqo and the sample is 102 brands, so treat the exact numbers as a baseline, not a law.

Why can’t one test tell you where you stand?

Because the answers move. Archipel’s October AI Visibility Observatory looked at 406,626 answers from six AI engines to more than 1,700 buying questions between July 12 and October 11. The brand cited first changed from one week to the next 63.3% of the time. ChatGPT read 14.5 pages per answer and cited 3.8. Google AI Overviews appeared on 86.7% of tracked buying searches in October, up from 56.3% in July. Most of those questions were in French, about France, so the US picture may differ.

Put the three studies together and you get a fairly simple model of how a CPG product gets recommended:

• The model starts from brands it already knows for the category.
• It searches to confirm facts and, for physical products, to update models, flavors and prices.
• It leans on rankings, review sites and retailer pages far more than on your own site.
• The result reshuffles week to week, so one screenshot proves nothing.

What should a CPG brand do about it?

First, measure the shortlist, not just the citations. Run your category’s unbranded questions (“best high-protein snack bar for kids”, “gentlest laundry detergent for sensitive skin”) through the API with search turned off. Graphite suggests at least 50 responses per prompt. We’d repeat the test whenever a major model ships. If you never appear with search off, a better product page alone probably won’t get you in.

Second, earn your way onto the pages models learn from. That means category listicles, review sites, retailer pages and press, the places the Ranqo and Archipel data say engines lean on. Graphite estimates two to five months between a model’s knowledge cutoff and its release, so this is a slow lever. Start now, before the next model ships.

Third, win the search step where physical products have an edge. Because search changes answers more for physical goods, current facts matter: the new flavor name, the current pack size, the price, the claims in the first lines of the description. Make sure retailer listings and your own pages agree.

Fourth, check your robots.txt. OpenAI’s search crawler (OAI-SearchBot) and its training crawler (GPTBot) are controlled separately. Blocking the training crawler may keep you out of the shortlist you’re trying to get into.

If you want to see which of your SKUs ChatGPT, Gemini and Perplexity name today, run a free SKU audit or book a demo and we’ll walk through your category. More on how we score products is on the Searchpanel homepage.