How PaperPMF translates language into purchase intent.

Research basis

The underlying paper was authored by researchers from PyMC Labs and Colgate-Palmolive. It evaluated semantic similarity rating across 57 personal-care product surveys containing 9,300 human responses and reported that SSR achieved 90% of human test-retest reliability.

That is a scoped result from an arXiv preprint, not a PaperPMF accuracy claim. PaperPMF uses generated personas, two separately generated reactions per persona, one generic purchase-likelihood anchor set, and the all-mpnet-base-v2 embedding model. Those differences mean the paper's reported result does not automatically prove that PaperPMF will produce the same result.

Read the underlying paper

Diagnostic sequence

1. Submitted product facts and an optional single image, or a compatible public HTTPS Shopify product page, become a factual snapshot. Shopify import is not universally compatible.
2. An LLM is prompted to rewrite the snapshot into one product concept intended to be neutral and to add no claims. This is a prompt objective, not independent fact-checking or a neutrality guarantee.
3. An LLM constructs a varied panel with factual demographic attributes conditioned by the free-text audience hypothesis. The personas are not recruited, population-weighted, source-backed, or guaranteed to satisfy every requested bound.
4. Each generated respondent produces two separately generated first-person reactions to the prepared concept and optional image.
5. Semantic embeddings compare each reaction with five purchase-likelihood anchors.
6. The resulting similarities are normalized into respondent and aggregate probability mass functions.

How PaperPMF is configured

Persona generation

openai/gpt-4o-mini

Concept and reactions

openai/gpt-4o

Responses per respondent

2

Maximum reaction concurrency

30

Semantic translation

PyMC Labs ResponseRater

Default embedding model

all-mpnet-base-v2

What the preview and full report include

Free preview

The preview shows the prepared product concept, the normalized primary image when one is usable, and one aggregate five-point probability mass function from 10 generated respondents. It does not expose the generated reactions, respondent profiles, or respondent-level PMFs.

Full evidence report

When the commercial purchase path is available, the full report runs a separately generated 100-respondent panel. It is not an extension of the preview's ten respondents. The report can include the aggregate PMF, mean intent, combined probability for “Likely” and “Very likely,” exact age, gender, location, occupation, and annual income filters, respondent and reaction-sample PMFs, two generated reactions per respondent, fixed rule-based full-panel themes, and method metadata.

Demographic filters recompute the displayed PMF over matching generated records. They do not create calibrated segments, and the stored full-panel themes do not recompute when a filter changes.

Report access

Reports use permanent, unguessable links rather than user accounts. Anyone possessing a report link can open it. “My reports” remembers visited links only in the current browser, and printable output uses the browser's print feature.

Inspect the recorded sample report to see these output fields. The sample is a recorded example created during product testing, not a customer case study, benchmark, or accuracy claim.

Method boundaries

• Synthetic reactions are model outputs, not statements from real customers.
• Results depend on the submitted stimulus, generated panel, model versions, embedding model, and purchase-likelihood anchors.
• More generated respondents add inspectable model output; they do not create the precision or representativeness of a human probability sample, repair prompt bias, or predict conversion or sales.

Independence

PaperPMF uses the open-source semantic-similarity-rating package. PaperPMF is not affiliated with PyMC Labs or Colgate-Palmolive.