Ask as your shopper, not as yourself: a persona-based AI brand check for challenger CPG brands

The most common AI visibility check we see at brand teams goes like this. Someone opens ChatGPT, types "best protein bar" or "best natural deodorant," reads the answer and screenshots it for the Monday meeting. If the brand is in it, good week. If not, bad week.

That check has two problems. It is one run, and assistants vary from run to run. And it asks the question the way nobody actually shops. Real people tell assistants who they are: "I'm training for a half marathon and I hate chalky bars," "my kid has eczema," "I'm trying to spend less on groceries." Two audits now show that this context changes which brands get recommended, and that it changes them most for exactly the brands that are not already the default answer.

This playbook turns that research into a test a small team can run in an afternoon.

What the research found

The newest evidence is from a paper posted on November 5, 2026 by researchers at Politecnico di Milano, The Popular Get More Popular. They asked 80 commercial questions in Italian across four sectors to GPT, Gemini and Perplexity, ten times each, both plain and preceded by one of three buyer personas. A persona was three sentences: life stage, experience with the category and attitude to price. The baseline persona wants to spend as little as possible; a price variant changes only the price sentence (for example, happy to pay more for quality or looking for the best value); a profile variant changes the life stage or experience instead.

They measured brand popularity bias as a slope: how many more answers out of 100 mention a brand for every tenfold increase in its Google search volume. A higher slope means the answers lean harder toward famous brands.

Show the data

Prompt	Popularity slope
No persona	0.98
Profile variant persona	0.89
Price variant persona	0.85
Baseline persona (spend as little as possible)	0.8

Source: Monaco, Lenti and Pierri, arXiv 2611.07614, Section 4.1 and Appendix C.3. All engines and sectors pooled. The baseline and price-variant drops were significant; the profile-variant drop was not.

Two of the three personas significantly reduced the tilt toward the most popular brands. The authors also looked at why. With the baseline or profile persona, at least one of the engine's web searches contained a price word in 86% to 96% of calls in fashion and beauty, sport and travel, against 25% to 46% without a persona. The persona changed what the engine searched for, and so what it found. Two answers to the same question with different personas also shared fewer sources than two repeats with the same persona.

An earlier audit, by Will Jack and colleagues, posted in May 2026, looked at the same effect from the brand's side. Across 2,000 runs with 10 personas, 8 prompts and three model setups, adding a persona lowered the overlap of recommendation sets in every setup measured. The effect was sharply split by brand size, according to the paper's abstract: category leaders kept about 80% of their recommendations across personas, while mid-market brands swapped up to 75% of the recommendation set as the persona changed.

Put those together and the logic for a challenger brand is simple. A bare prompt shows you the leader's home turf, where you will mostly lose. A persona prompt shows you the parts of the market where the answer is still up for grabs.

Step 1: Write four personas from real customers

Do not invent personas from a marketing deck. Pull them from what you know about the people who buy you and the people you want. Each persona is two or three plain sentences in the first person, mirroring the structure the Milan team used: life stage, experience with the category, attitude to price.

Persona	Text placed before the question
A: Budget	I'm a college student who lifts four days a week. I've tried a few bars. I want the cheapest one that still tastes fine.
B: Quality	I'm a college student who lifts four days a week. I've tried a few bars. I'll pay more for clean ingredients.
C: New parent	I'm a new parent and barely have time to eat. I've never bought protein bars. I want something I can keep in the car.
D: Dietary need	I'm vegan and avoid soy. I run most mornings. Price matters less than what's in it.

Keep A and B identical except for the price sentence, as the researchers did, so you can see what price attitude alone does. Make C and D the shoppers your brand is actually built for.

Step 2: Write five questions that name no brand

Use the questions shoppers ask, not the ones your SEO team targets. Mix one broad question ("What's a good protein bar?"), two use questions ("What should I eat before an early run?"), one comparison ("Are protein bars or shakes better after lifting?") and one constraint question ("Which protein bars have no added sugar?"). Never include your brand name or a competitor's. Tip 12 covers the basics of checking whether ChatGPT recommends you.

Step 3: Run each combination three times

Four personas, plus a bare version, times five questions is 25 prompts. Run each three times on each assistant you care about, in a fresh chat with no memory or history, logged out where you can. That is 75 answers per assistant. It sounds like a lot; in practice it is an afternoon for one person per assistant.

Why three runs? Because single answers are unreliable. In Visa's shopping benchmark, which we covered in today's data post, all three runs of the same scenario passed only 34% of the time. Visibility checks have the same problem, as we argued in our piece on personal AI answers.

Step 4: Record the same five things every time

For every answer, write down:

1. Which brands were named, in order.
2. Whether your brand was named, and in what position.
3. What the answer said about your product, and whether it was right.
4. Which sources were cited, if any.
5. Any requirement the assistant asked about before answering.

Keep the sheet's columns fixed from month to month. The value is in the comparison over time, not the single snapshot.

Step 5: Read the results by persona, not in total

This is where most teams go wrong. If you average across all prompts, you get one number and you lose the finding. Jack and colleagues say directly that a measurement that aggregates across personas "systematically obscures" the variation.

Instead, make a small grid: personas down the side, assistants across the top, your mention rate in each cell. Look for three patterns.

Where you win. Cells where you are named in most runs. These are the shoppers the assistant already associates with you. Protect them: keep the facts that earn those mentions accurate and prominent.

Where you are close. Cells where you appear sometimes, or where a direct competitor of your size appears and you do not. These are the best opportunities. Read the answers and the cited sources to see what the assistant is matching on.

Where you are absent but should not be. Cells for a persona your product is built for, where you never appear. That is a content gap, not a ranking problem. Usually the facts that would match that persona are missing from your pages or from the third-party pages the assistant reads. Tip 20 covers content for specific needs.

Step 6: Fix the facts, then retest

For each gap, write down the specific fact that would let an assistant match your product to that persona: soy-free, 20 g of protein, no refrigeration needed, under $2 a bar at a named retailer. Check whether that fact is on your product page, in your feeds and on your main retailer listings, stated plainly. If an assistant says something wrong about you, tip 25 covers how to trace and correct it.

Do not rewrite copy for tone. The same Milan paper found that fact-preserving rewrites for structure, fluency or authority moved a page by about 0.02 positions, which is nothing; we wrote about that in today's think piece. Add facts. Then rerun the same grid a month later.

What a good first month looks like

You will not move the bare prompt much in a month, and you should not judge yourself on it. A good first month is a grid that shows two or three persona cells where you went from absent to sometimes present, with a clear fact behind each change.

If you are doing this across many products and assistants, it gets heavy fast. Some teams move it into a tool such as SearchPanel, from XENA Intelligence, which records the actual answers per product; others keep a shared sheet. The method matters more than the tooling.

The caveats

The Milan audit used Italian prompts in four sectors and API access, not consumer apps. The Jack audit used software prompts such as "best CRM software," not grocery, and only a few model setups. Neither tested packaged goods directly. Persona effects in the consumer apps may be larger, because those apps also carry memory and personal context, or smaller, because they wrap the model in shopping data. We would treat the direction as solid and the sizes as unknown for your category.

What both papers agree on is enough to change how you test. The same question produces different brands depending on who the assistant thinks is asking. If you only ever ask as nobody in particular, you are measuring the one version of the market you are least likely to win.