Mars is measuring how its brands show up in AI answers. Here's how to build a baseline that doesn't flatter you

Mars just said out loud what most CPG brands are quietly doing. In a Digiday interview published November 11, Neil Reynolds, Mars's global chief customer officer, said the company's first step in agentic commerce is to "get some real data around what's the start point for how our categories and how our brands show up."

That's the right first step. It's also the step most teams get wrong, because an AI visibility baseline can be built to flatter you without anyone meaning to. Two studies from the past few weeks show exactly where the flattery creeps in. Here's how to build a starting number you can actually manage against.

What did Mars say about AI visibility?#

Reynolds described two parallel road maps. One is about stores becoming places for discovery. The other is about agents: "how do we make sure our brands are discoverable by the LLMs and how do we make sure that once those brands are discovered, that they convert through to purchase?"

Mars is getting its starting data "through partnerships with emerging tech companies," and it wants to know two things: how it shows up, and "where's that data being sourced from."

The backdrop is Mars's future of shopping report from October 6, built with Kantar and Synthesis on more than 180,000 data signals. It predicts that "legacy browsing and impulse shopping behaviors give way to predictive fulfillment that happens in the background." For a snack company, that threatens the checkout-lane candy bar. If an agent restocks the pantry, a brand that isn't on its list gets no second chance at the register.

Note what Mars did not say. It hasn't shared a visibility number, a lift or a tool name. This is a company setting a baseline, not reporting results. That's exactly why the method matters.

Which prompts should a CPG baseline track?#

This is where most baselines go soft. The usual shortcut is to have a model write test questions from your own product pages. Those questions carry your claims, your flavors and your pack formats, so of course the answers mention you.

Proofsource measured the gap in a study published November 10. It took 105 brands that tracked both real searches (keywords and People Also Ask questions, word for word) and questions written by a model from the brand's website. Across 9,329 answers from ChatGPT, Claude, Perplexity and Google AI Overviews between October 16 and November 9, written questions named the brand in 7.7% of answers. Real searches named it in 5.2%.

ChatGPT showed the widest gap, 7.0% against 12.7%. Of the 61 brands named at least once, 46 did better on their written questions. The reason is length. The median real search had 5 words and the median written question had 15, and Proofsource says "the extra words are where the brand comes in."

Neither set is exactly what shoppers type into a chatbot. Semrush's U.S. clickstream study found the average ChatGPT prompt that triggered web search grew from 4.7 words to 8.7 words between early 2025 and early 2026. So the honest move is to track both kinds and report them as two numbers. The real-search number is your floor with shoppers who haven't found you yet.

Does the type of shopper question change the baseline?#

A lot. Yarnit's fall study found brands named in 37% of answers to direct recommendation questions, 17% of "what should I look for" questions and 8% of concern questions. We covered that split in our post on how ChatGPT ranks merchants, and PickedByRobots paired it with Euromonitor's data.

If your baseline is 50 "best protein bar" prompts, you're measuring the stage where brands do best. The criteria questions ("how much protein should a snack bar have") are where the shopper learns what good looks like, and the brand is usually missing.

Need matters inside a single category too. In Euromonitor's scalp-care test, Head & Shoulders appeared in 66% of answers to general dandruff prompts and 24% of answers about psoriasis and eczema, where Neutrogena reached 69%.

Should you measure each AI assistant separately?#

Yes, and each country too. Euromonitor found Nizoral in 58% of Gemini answers and 24% of ChatGPT answers. Gemini drew 58% of its citations from retailer pages like Target, Boots and Superdrug, while ChatGPT drew 32% from editorial and beauty media and 17% from clinical sources. Philip Kingsley showed up in 48% of UK answers and 2% of U.S. ones. CeraVe ran the other way, 33% in the U.S. and 1% in the UK.

That answers Reynolds's second question, "where's that data being sourced from," for one category: mostly not from the brand. Brand-owned sites were 3% or less of citations on both engines. Your retailer listings and your earned media are your most-quoted copy.

Filter the findings below by what each one measures.

How should a CPG brand build its baseline?#

Here's the version we'd run for a brand with a few dozen SKUs.

1. Pull real questions first. Search terms from your retailer and Google reports, People Also Ask, retailer Q&A and customer service logs.
2. Cover three stages. For each product line, write recommendation, criteria and concern questions. Don't let the "best X" bucket be more than half the set.
3. Add needs, not features. "Snack for a long drive with kids" is a need. "Peanut-free 100-calorie pack" is your spec sheet. Track both, labeled.
4. Run every engine and country separately. At least ChatGPT, Gemini, Google AI Mode and Alexa for Shopping, in the U.S. Report a number per engine.
5. Log the source, not just the mention. Which retailer, publisher or Reddit thread was cited, and whether the product details were right.
6. Freeze the set. Changing the prompt mix moves the number even when no answer changed. Add new prompts as a separate cohort.

Why log sources so carefully? A UIUC and IISc preprint summarized November 11 found that rewriting one sentence of an ordinary article to name a brand put that brand into 62.7% of GPT-5-mini answers to questions that never named it, in the strongest condition of a test retrieval pipeline. That's a lab setup, not live ChatGPT. Still, if your number jumps, you want to know which page moved it.

What does this mean for CPG brands right now?#

Mars is right that conversion is a separate problem from discovery. But neither gets fixed without a starting number that holds up. Then fix the pages the engines actually read, starting with your retailer listings.

That's the work Searchpanel does for every SKU. Our free SKU audit shows whether ChatGPT, Gemini, Perplexity and Google recommend each product and which sources they cite. The Searchpanel homepage explains the score, our glossary defines unbranded prompts and AI share of voice, and if you want a baseline set up across a portfolio, book a demo.