When evaluating AI visibility work, a more decisive question than “Which keywords do you track?” is “On whose behalf, and with which questions, do you measure?” In Recro simulations, the same brand appeared in every answer to one decision-maker’s questions and in none of another’s. Measurement that doesn’t set up the persona, the question and the number of runs correctly shows a brand as better or worse than it is. The seven questions below separate strong work from weak.
AI visibility has quickly become a crowded market. Prompt-tracking tools, GEO agencies, “AI modules” added to SEO suites and consulting services make the same promise with different methods. For a marketing director or CMO, the hard part is finding a common yardstick to compare these offers.
This article builds that yardstick. For each question, we share a finding from Recro’s own simulations that shows why it matters. The findings are anonymized from 460 question-answer runs for 6 brands between April and September 2026.
1. On whose behalf is the measurement done?
In a simulation for a coffee supply brand, the brand was mentioned in 8 of 8 runs for the questions of a tourism and restaurant group’s F&B director persona. The same brand wasn’t mentioned at all in the questions of hotel chain, café chain and restaurant chain procurement directors. Had a single persona been chosen, the report would have said either “excellent” or “zero”.
Ask: Are personas derived from the brand’s real buyer structure, or is a generic “marketing manager” profile used? Are different roles in the buying group measured separately?
2. In whose language are the questions written?
In our simulations, brands appeared in 33% of answers when asked about by name and in 5.5% of decision questions that didn’t name them. Questions that include the brand’s name, or are written in the marketing team’s own category language, make results look better than they are.
Ask: Do the questions grow out of the buyer’s need and conditions? Are discovery and elimination questions without the brand’s name reported separately?
3. How many times, and over what period, is it measured?
In the coffee supply simulation, 236 runs produced 50 different “dominant competitor” names. In an enterprise software category, 112 runs produced 28. A single screenshot shows only one moment of that spread. SparkToro and Gumshoe.ai’s January 2026 research also found that the chance of two answers to the same question giving the same brand list is under 1 in 100.
Ask: How many times is each question run? Are results given as a rank, or as an inclusion rate and a time series?
4. Are the answer’s sources shown?
In a marine equipment simulation, one source that carried a competitor onto the shortlist was that competitor’s company profile on Kariyer.net, a job board. In other simulations, catalogs uploaded to document-sharing sites years earlier stood out as sources. An action plan made without seeing sources usually focuses only on the brand’s own website.
Ask: Does the report show the links that source the answer and which sources carry competitors?
5. Does it give the cause and the action alongside the number?
“Your visibility is 8%” is an observation, not a decision. The most common causes in our simulations were pricing and ROI information never being published (in all 73 reports), proof sitting only on the brand’s own site and absence from the category lists competitors appeared in. Each cause calls for a different action.
Ask: Are findings separated into gap types such as Communication, Content and Authority? Are actions prioritized and sector-specific?
6. Are AI’s claims verified?
When the model can’t find information, it fills the gap with generalizations. In our simulations, 27 answers followed “no public information available” directly with a sector generalization. In the European Broadcasting Union and BBC’s 2025 research, 45% of AI answers had at least one significant issue. An unverified AI output shouldn’t enter a report as a finding.
Ask: Are standout claims checked by a person against what exists online?
7. Are results retested?
Unless the same persona and question are measured again after an action is implemented, the work’s effect can’t be known. Model versions change too, so the measurement date and version should be recorded.
Ask: Is there a retest plan after actions? Are results reported with the model version?
Evaluation table
| Question | Strong answer | Weak answer sign |
|---|---|---|
| On whose behalf? | Role-based personas derived from the real buyer structure | One generic profile, or no persona |
| Which questions? | Decision questions in the buyer’s language, without the brand’s name | A keyword list or the brand’s own category language |
| How many times? | Repeated runs, rates and time series | A one-off screenshot, “you’re #3 in AI” |
| Which sources? | Links that source the answer and competitors’ sources | No source information |
| Cause and action? | A prioritized plan classified by gap type | A generic “produce more content” recommendation |
| Verification? | A reality check with human review | Model output goes into the report as is |
| Retest? | Post-action measurement with the same persona and question | No impact measurement |
This table works for evaluating an in-house measurement as well as an outside service. We explain the core logic of measurement in how to measure AI visibility, and how decision questions are mapped in what is a decision question.
Keywords or key people: which are you buying?
The core difference between offers is this. Some track keywords and count how often a brand appears in prompts related to them; that’s valuable for following broad category trends. The other approach sets up the key person with buying intent and explains why the brand is present or absent in that person’s decision questions; that puts investment where sales decisions are made.
The two can complement each other. But the approach that brings the board an answer to “Why aren’t our buyers choosing us?” is the second one.
Where does Recro stand in this evaluation?
The Recro Insight Model was built to answer each of the seven questions structurally. Measurement starts with the key person with buying intent.
One simulation · 4 key pillars · 90+ steps
What sets Recro apart is four pillars built from scratch for each brand. Together they form a single simulation of 90+ steps, with each pillar producing the input for the next.
- Persona builder: Decision-maker profiles that represent the brand’s real buyers, specific to its sector and sales structure.
- Question builder: The brand-specific decision questions these people ask AI along their buying journey.
- Report builder: A report that reads the answers against the brand’s goals and competitors, together with the sources behind them.
- Action recommendation builder: A prioritized list of actions that closes the sector-specific signal gaps.
Against the evaluation questions, each pillar maps as follows:
- The persona builder answers question one: measurement is done on behalf of decision-makers that represent the brand’s real buyers, specific to its sector and sales structure.
- The question builder answers question two: questions grow out of persona attributes and follow a natural buying-journey flow.
- The report builder answers questions three, four and six: daily outputs, the links that source the conversation, estimated visibility probability and a monthly reality check with human review.
- The action recommendation builder answers questions five and seven: a sector-specific signal list, prioritized actions and a retest after actions.
Executive summary
- In AI visibility work, the decisive question is “On whose behalf, with which questions?” more than “Which keywords?”
- In Recro simulations, the same brand showed 100% visibility for one persona and 0% for others; questions naming the brand inflated results up to six times.
- Strong work offers role-based personas, questions in the buyer’s language, repeated measurement, sources, causes and actions, human verification and retesting.
Frequently asked questions
Aren’t prompt-tracking tools enough?
They’re useful for following category-level trends. Explaining why a brand is missing from a specific buyer’s decision question takes brand-specific persona, question and source analysis.
How many personas are enough?
At least one for every critical role in the buying group. In our simulations, when the role changed within the same sector, both visibility and the dominant competitor changed.
When should we expect to see results?
Diagnosis comes with the first report. The effect of access and content actions is measured within weeks, authority actions within months.
Can we do this in-house?
Yes; the seven questions apply to an internal team too. The real challenge is building personas and questions independently of the brand’s own language and sustaining repeated measurement.
What drives the price?
The number of personas, the size of the question set, measurement frequency and language coverage. When comparing proposals, make sure these four variables are held equal.
Ask Recro these seven questions too. We’ll build a sample simulation with brand-specific personas and decision questions and share the result in a demo report.
Request a demo insight report →
Sources
- Recro simulation data · April–September 2026 · 6 brands, 460 question-answer runs, 73 reports (internal, anonymized; run on a single AI model)
- Search Engine Land · AI recommendation lists repeat less than 1% of the time (SparkToro, Gumshoe.ai)
- EBU · AI assistants misrepresent news content 45% of the time (EBU and BBC, 2025)
This article was prepared with AI assistance and published under the review of Mehmet Semih İpek.
