From AI visibility evidence to decisions
PollyX shows how AI presents your business, identifies evidence-backed improvements, and tracks what changes after you act.
A check measures the answers collected for a selected set of questions. To decide whether to enter a market, set a price or launch a product, add evidence from actual customers and relevant market data. The Research workspace keeps those strands together with their sources and limitations.
Start with a decision and customer questions
Write down the decision, the audience, the market and the evidence you still need. Save questions from search, sales, support and interviews with a dated reference. Keep untested ideas labelled as hypotheses. Separate questions that name your business from questions that ask about the category: they answer different questions about visibility.
The workspace shows which stages have no saved questions. Its question counts describe your library, not the frequency of demand in the market. Select 3 to 12 saved questions to prepare a new check and review them before running.
Measure variation and preserve comparability
New checks record their question set, market, language and collection settings. Alongside the existing category-question repeats, up to two other questions receive one extra attempt from up to three selected assistants (normally ChatGPT, Gemini and Perplexity). The repeatability section reports how many observations were collected, whether prominence changed and its observed range. These extra observations do not alter the headline score.
Two or three observations are a small variability check, not a population estimate. Missing answers stay missing. Recorded settings cannot reveal every change a provider makes. Controlled comparisons require matching recorded settings and questions; older checks without that record remain exploratory.
Check the reading against the retained evidence
Human review opens a real answer and its recorded brand prominence, tone and citation references. Review brand identification, tone, local relevance or whether a citation supports the claim. Record the reasoning, reviewer and dated source. If the evidence does not settle the question, select uncertain.
Error share is the number marked incorrect divided by determinate judgments for that criterion. Repeated reviews of the same answer and criterion count once, using the latest judgment. These are owner-entered judgments on selected answers; they are not an independently certified accuracy rate or a random quality audit. Independent review requires an actual reviewer and an agreed sample, with disagreements and uncertainty retained.
Record actions before assessing their results
Choose a baseline, record an intervention date and hypothesis, and separate affected questions from questions you intend to leave unchanged. A later check compares only question/assistant pairs answered in both checks. The difference in changes subtracts the comparison group's movement from the intervention group's movement.
This is an observational comparison. Seasonality, other campaigns, news and provider changes can affect both groups. It does not establish that your action caused the change. Business outcomes require a stated source, definition and equal, non-overlapping periods. Revenue needs a currency; visits and leads are counts. Existing Analytics records can be linked for a whole-property comparison of identifiable AI referrals and key events. Only their references are saved: disconnecting Google removes its figures. They are not copied into research notes, case-study exports or AI prompts.
Add actual human and market research
Import aggregate results from real surveys, interviews, concept tests, pricing studies or licensed datasets. Record the provider, fieldwork dates, recruitment, population, sample selection, weighting and limitations. Every answer count has a denominator. Percentages describe that sample and question; interview counts do not estimate market prevalence.
PollyX does not recruit respondents through this workspace, supply a representative panel, verify a research provider or manufacture missing responses. A provider or your team must carry out that work. Do not substitute generated AI personas for people. A broad market decision can also require market sizing, competition, pricing, distribution and financial evidence beyond AI visibility.
Publish only what the evidence supports
Case-study drafts bring together the intervention, available comparisons, supplied business outcomes and attached human research. They retain limitations, consent references and review declarations. Exports remain private drafts; naming a reviewer does not certify independence and selecting permission obtained does not publish the draft.
We do not claim a demonstrated customer uplift or independent validation without the corresponding real evidence. Software checks can verify calculations and access controls; they cannot establish that a customer outcome occurred or that a research sample represents a market.