Original insights by Allison Huang
30-second rundown
Key learnings:
- AI visibility changes with the customer: different personas shared only about one quarter of recommended brands and one fifth of cited domains.
- Income changed the recommendations and evidence: answer engines translated income into signals such as budget, premium, or luxury, then searched different parts of the market.
- One average score can hide the buyers you are losing: test important questions across your actual customer personas and track the brands, sources, and hidden assumptions.
This week: Test three buying questions using your most valuable customer persona and one contrasting persona across ChatGPT, Claude, and Gemini. Record the recommended brands, cited sources, and qualifiers such as budget, luxury, young, or professional. Flag any audience consistently sent toward competitors.
Coming 3 months: Use an AI search monitoring tool to track these questions across your most important personas. Repeat each test enough times to separate persistent differences from normal variation, then connect gaps to missing facts on your pages or weak evidence elsewhere online.
Profound compared 71,147 responses from ChatGPT, Claude, and Gemini and found that one buying question produced different brands, citations, and searches when the prompt changed the customer’s income, age, gender, or occupation.
Change the persona and the answer engine may redraw the market before the customer sees a single recommendation.
Your average AI visibility can look healthy while the buyers you want most are being driven straight to a competitor.
The Machine inspects the customer, guesses what they can afford and probably want, then quietly chooses which highway they are allowed to travel.
1. One Prompt Became Several Markets
Profound kept the buying question the same and changed only the customer’s gender, age, income, or occupation.
The Machine changed the brands, sources, and searches like a crooked concierge switching the guest list after checking everyone’s shoes.
Repeated answers for the same persona shared about two in five recommended brands. When the persona changed, the overlap fell to roughly one in four.
Citation overlap also fell from roughly one third for repeated runs of the same persona to around one fifth between different personas.
The customer description changed both the recommendation and the evidence supporting it.
Income produced the loudest split. Clothing recommendations stretched from resale platforms and single-digit T-shirts to luxury brands charging more than $1,000. Higher-income answers also mentioned more brands and sometimes relied more heavily on brand-owned webpages.
A brand can appear visible in a general test but vanish when the prompt resembles its actual customer.
The average score smiles politely while the sale leaves through the back door.
2. Hidden Searches Decide Who Belongs
ChatGPT and Claude revealed part of the translation racket through their recorded searches.
A hidden search query is a search the answer engine creates from the user’s prompt before writing its answer. The user asks one question, but the Machine may secretly run several searches containing assumptions they never supplied.
Income became affordable, budget, premium, or luxury. Age and occupation triggered additional guesses about style, needs, and suitability.
The Machine was rewriting its queries based on the customer.
Some shortcuts helped narrow the results. Others carried bias in the glove compartment.
Clothing searches for people in their fifties were far more likely to assume the shopper was female. A relevant retailer serving older men could disappear before its products were compared.
The visibility loss started inside the hidden query of the LLM.
If the Machine translates the customer into the wrong category, the right business can be locked outside the casino before the cards are dealt.
Clear, crawlable facts reduce the room for invention. State price ranges, use cases, accessibility features, service areas, eligibility requirements, and limits on suitability.
A hotel should not merely call itself “affordable luxury.” It should show typical rates, included services, room types, suitable occasions, and who each offer was designed to serve.
Describe the offer precisely enough for the retrieval system to establish who it fits.
3. Replace the Average With a Persona Map
An aggregate AI visibility score can show whether a brand appears frequently overall. It cannot show whether the brand appears for the people most likely to buy.
Replace the average with a persona map: a comparison of how the same buying questions are answered for each important customer group.
Test the same question across the answer engines your customers use. Repeat it enough times to separate persistent differences from the Machine’s ordinary dice game.
Track:
- Recommended brands
- Cited webpages and domains
- Price tiers
- Hidden qualifiers
- Personas repeatedly receiving the wrong recommendation
If your brand disappears because the answer engine cannot confirm a price, use case, feature, or requirement, strengthen the relevant page.
If competitors are supported by reviews, comparison articles, directories, or trade publications while your evidence comes only from your website, build stronger third-party coverage.
Missing facts call for better owned content. Missing trust calls for outside evidence.
One visibility score can tell you the highway is busy. It cannot tell you whether your best customer was placed in the wrong car and delivered to a competitor with the Machine grinning behind the wheel.
AI visibility is a distribution of answers shaped by the person the system thinks it is helping.
Measure the personas you serve. Publish the facts proving who your offer fits. Watch the hidden searches for signs that the Machine has invented the wrong customer.
Because the answer engine does not recommend brands in a vacuum.
It recommends them to someone.
