Key takeaways
- AI engines build answers from the question, the inferred intent, and whatever context the user gives. Change the person asking and the brands, order, and descriptions can change too.
- Being mentioned is not the same as being chosen. A brand can appear often and still be described in a way that loses a specific buyer.
- Test one intent through several personas, then compare visibility, position, and sentiment per persona to find where you are invisible or mis-described.
- Fix gaps with content and sources matched to each prompt type. Promptwatch Data shows commercial, informational, and branded prompts pull very different content.
- Several tools now offer persona-aware tracking. Pick based on whether you only need measurement or also need help fixing what you find.
Why one brand perception no longer exists
In classic search you had a ranking. Roughly, you were position four or you weren't. Buyers might see slightly different results by location, but the page was broadly shared.
AI answers don't work like that. The model takes the query, infers who is asking and what they need, and writes a synthesized answer. If the prompt says "I'm an email marketing manager at a mid-sized charity, I need a CRM that handles GDPR and costs under $20K a year," the model has a profession, an industry, a compliance need, and a budget ceiling to work with. Semrush's enterprise team makes this point when explaining why personas went from nice-to-have to necessary in AI search.
The result is that your brand doesn't have one perception in AI. It has several, and you probably have only looked at one of them: the generic prompt your team typed into ChatGPT during a meeting.
A recent Search Engine Land post framed it bluntly: a brand can show up in AI answers constantly and still lose the buyer. Their study of about 2,000 scored brand mentions found that mentions alone told you little about whether the buyer was persuaded.

I think this is the most underrated shift in AI visibility work. Most dashboards still report a single score. That score hides the buyer-level differences you actually need to act on.
Mention, position, and description are three different problems
Before getting into personas, separate what you are measuring.
- Inclusion: does your brand appear in the answer at all?
- Prominence: where does it land among the other brands?
- Framing: what does the model say you are good for, and what does it say you lack?
Peec AI's launch post on brand perception puts the third point well. Your brand might appear in 60% of answers across tracked prompts, but that number doesn't tell you what the AI gave the buyer as a reason to choose you. The model weighs sources, compares options, and tells the buyer what each brand is known for. That conversation is mostly invisible to you.
Personas make the framing problem sharper. A finance buyer might hear that you are "strong on security and audit trails." A startup founder might hear that you are "expensive and built for enterprise." Both can be true of the same product, and both can come from the same set of sources, just weighted differently by the question.
What actually changes between personas
A few things shift when the asker changes:
- Terminology. A compliance lead asks about "data residency." A marketer asks about "where my data is stored." Different words pull different pages.
- Constraints. Budget, team size, region, and tooling stack get baked into the prompt and filter which brands survive.
- Evaluation criteria. A developer wants API docs and SDKs. An executive wants case studies and pricing clarity.
- Conversation history. Chatoptic points out that answers also depend on signals accumulated in a conversation, such as earlier questions or implied constraints.
One honest caveat: not every prompt carries persona information. Many real prompts are short and generic. Persona testing is most useful for commercial and comparison queries where buyers tend to add context, and less useful for quick factual lookups.
How to build personas for AI testing
You probably already have buyer personas. Most are static PDFs that sit in a shared drive and say things like "Marketing Mary, 38, likes podcasts." Those are not useful for prompt testing. You need personas defined by how they ask questions.
Start from real buyer data
Pull from sources that already contain buyer language:
- Sales call transcripts and discovery notes
- Support tickets and onboarding questions
- Search Console queries and People Also Ask
- Reddit and community threads where buyers describe their situation
- Win/loss interviews
Define each persona with testable attributes
For each persona, write down the role, industry, company size, main constraint, and the three to five criteria they use to judge options. Then write the way they would phrase a prompt. Something like:
Persona: IT security lead, 500-person fintech
Constraint: SOC 2, EU data residency
Prompt: "We're a mid-sized fintech in Germany. What are the best [category] tools that meet SOC 2 and keep data in the EU?"
Keep the set small at first
Four to six personas is plenty. Teams that generate hundreds of prompts often end up with what Semrush calls volume without variety: lots of prompts that all reflect the same perspective. Fewer personas with several intent variations each will teach you more.
Run the test: one intent, many personas
The method is simple. Pick one intent, such as "find the best [category] tool." Rephrase it per persona. Run each version across the AI engines your buyers use, and repeat over time because answers shift.
Record, per persona:
- Whether you were mentioned
- Your position among the brands listed
- Which competitors appeared with you
- The claims the model made about you
- Which sources were cited
Chatoptic describes aggregating this per market, grouping personas by geography and language so you can see visibility in a specific region. That is a useful idea even if you don't use their product. A persona in Munich and one in Chicago should not be averaged together.

Spin Sucks makes a related point from the PR side: AI platforms themselves have become an audience. If ChatGPT doesn't know you, the argument goes, your customers won't either. I'd add that the model's picture of you is built from receipts such as citations, consistency across sources, and fresh signals, not from clever campaigns.
Reading the results
Look for patterns across personas, not individual results. A few common ones:
| Pattern | What it usually means | First fix |
|---|---|---|
| Visible for one persona, absent for another | Your content speaks one audience's language | Publish pages in the missing persona's terms |
| Visible but described as expensive or complex | Pricing and positioning pages are unclear, or third parties describe you that way | Clarify pricing pages, fix review-site profiles |
| Mentioned late in the list | Competitors have stronger citations | Close source gaps, get cited where AI looks |
| Wrong claims about features | Outdated or conflicting pages | Update the pages stating it, correct third-party listings |
| Strong with developers, weak with executives | Docs are good, business-case content is thin | Case studies, ROI and comparison pages |
Why the source mix differs by prompt type
Persona-specific answers draw on different sources, and the type of prompt matters a lot. One caveat first: Promptwatch Data doesn't split results by persona. It splits by prompt intent and prompt type, which is a related but different cut. It's still useful because buyers in different stages and roles tend to ask different kinds of questions.
In Promptwatch's content-type report (Sep 28 to Oct 4, 2026), product pages made up 22.4% of ChatGPT citations overall. But the mix moved by intent:
- Commercial prompts: product pages 26.0%, landing pages 20.6%, listicles 15.0%, comparisons 7.9%.
- Informational prompts: how-tos 10.7%, docs 8.5%, and articles 7.7% all rose above their overall share.
- Branded prompts: product pages 23.3% and landing pages 22.8%, with reviews at 5.6% (more than 3x their overall share) and social posts at 3.0%.
So the buyer who asks "best CRM for small businesses" sees a different source mix than the one who asks "is HubSpot good for startups." If your persona is at the evaluation stage, comparison pages and listicles matter. If they are checking you out by name, your reviews and FAQ carry weight.
There is a second finding worth knowing. In Promptwatch's unique-domains report for the same week, 47.6% of ChatGPT responses with citations cited 10 or more domains. But for brand-specific prompts, the most common count was just 5 domains (14.9% of responses), and 40.7% cited 5 or fewer. The report's takeaway: because brand-specific answers draw on few sources, the review sites and forums on that short list shape how ChatGPT describes you.
That is where perception gets decided. When a buyer asks about you by name, a handful of pages write your description. If one of them is an outdated review, that is the sentence the buyer reads.
Fixing the gaps
Match content to prompt intent
Plan content per intent rather than per keyword. Commercial prompts reward product pages and comparisons. Informational prompts reward guides and docs. Branded prompts reward your own site, FAQ, and reviews.
Write for each persona's criteria
If your security-lead persona never sees you, check whether you have a page that answers their exact questions in their words: certifications, data residency, audit logs. Then make sure it is crawlable. AI crawler access is a quiet cause of missing mentions.
Correct what third parties say
Find the review sites, directories, and forum threads that appear in brand-specific answers. Update your profiles there and answer the complaints that keep getting repeated. This is slow, unglamorous work, and it often moves framing more than a new blog post does.
Check claims against your own pages
When AI states something wrong, trace which page teaches it. Sometimes it is your own old pricing page.
Re-test on a schedule
AI answers change constantly. A persona test run once is a snapshot. Run it weekly or monthly and watch for drift, especially after launches.
Tools for persona-based perception tracking
Persona support varies a lot. Some tools only let you tag prompts. Others generate persona-specific prompts and compare results on a radar or heatmap. I haven't compared pricing across all of these, so check current plans on each vendor's site.
Promptwatch supports custom personas and persona stacking, so you can prompt like your actual buyers. Persona insights show visibility, position, and sentiment per persona against the general audience on a radar chart, and flag the persona with the biggest gap. Brand Narrative pulls out the individual claims AI makes about you and files them under themes like Pricing, Features, and Support, and Fact Check traces inaccurate claims to the pages that teach them (Business plan and above). It then goes further than tracking, with a Content Agent and Action Board for fixing the gaps. You can try it with the 7-day free trial or start with the free AI visibility report.

Gumshoe AI is built around persona-based tracking specifically.

Peec AI added a brand perception feature focused on how AI describes each brand, not just whether it appears.
Semrush's enterprise AIO product has persona-based prompt generation for teams already in that ecosystem.
Waikay centers on AI brand perception and visibility.
Profound is an enterprise option with deep analytics.
| Tool | Persona handling | Perception or framing analysis | Helps fix gaps |
|---|---|---|---|
| Promptwatch | Custom personas, stacking, radar comparison | Brand Narrative, sentiment, Fact Check | Content Agent, Action Board, CMS publishing |
| Gumshoe AI | Persona-based tracking is the core idea | Visibility by persona | Limited, check current features |
| Peec AI | Prompt tracking with suggestions | Brand perception feature | Mostly monitoring |
| Semrush AI Visibility Toolkit | Persona-based prompt generation (enterprise) | Within Semrush suite | Via Semrush content tools |
| Waikay | Check current features | Brand perception focus | Mostly monitoring |
For a wider view of the category, the GEO software directory at bestgeosoftware.com lists more options.
A practical 30-day plan
- Week 1: Define four to six personas from sales and support data. Write three prompt variations each for one or two core intents.
- Week 2: Run the tests across ChatGPT, Gemini, Perplexity, and Google AI Overviews. Log mention, position, competitors, claims, and sources.
- Week 3: Identify the biggest gap per persona. Trace the sources behind inaccurate or weak framing.
- Week 4: Publish or update pages for the top two gaps, correct third-party profiles, and set a recurring re-test.
Resist the urge to chase every gap. Pick the persona with the most revenue behind it.
What to avoid
- Treating one aggregate score as the answer. It averages away exactly the differences you need.
- Inventing personas in a workshop and never validating them against real buyer language.
- Reading single-prompt results as truth. The Promptwatch unique-domains report recommends measuring across many prompts, since one prompt tells you little.
- Ignoring sentiment and framing because inclusion numbers look healthy.
The bottom line
Your brand is described differently depending on who is asking, and most teams only see the version a generic prompt produces. Test one intent through several real personas, compare how you are included, ranked, and described, then fix the content and third-party sources behind the weakest persona first.


