Features · Personas

Personas

Updated 2026-05-25 Feature eCommerce Insights team

Personas turns a CSV of reviews into 4-8 named buyer personas using LLM clustering. The personas feed Prompt Runs (for persona-segmented sessions), Image Enhancer (for the AI creative brief), and Listing rewrite (for persona-aware FAQ content).

Run audits as different shopper personas to surface gap differences.

What it does

Reviews are the cheapest, most honest signal a brand has about its customers. Personas reads the review corpus, clusters it, and produces a small set of named buyer personas — usually four to eight — that the rest of eCommerce Insights can use as context.

Inputs

One CSV file or one connected review API. Supported sources:

Yotpo (API or CSV)
Judge.me (API or CSV)
Okendo (CSV)
Trustpilot (CSV)
Stamped (CSV)
Reviews.io (CSV)
Amazon reviews (public scrape — limited to your tracked Amazon SKUs)

Minimum CSV columns: review_text, star_rating, product_sku or product_url. Optional: reviewer_name, review_date, verified_purchase.

How clustering works

Two passes.

Topic modelling. An embedding-based pass groups reviews into natural clusters based on language similarity. The number of clusters is chosen automatically based on corpus size — four for small corpora, up to eight for large ones.
LLM naming. A second pass takes each cluster and writes a human-readable persona: a short name, a paragraph summary, top three feature loves, top three concerns, and three sample quoted reviews.

The split is deliberate. The first pass produces the right number of clusters; the second pass writes the copy. Mixing them tends to produce too few, too generic personas.

Output

Per persona, five fields.

Name: One to three words. Editable.
Summary: One paragraph. Includes the persona's likely demographic and motivation.
Top feature loves: Three features this persona consistently praises.
Top concerns: Three concerns this persona consistently raises.
Sample reviews: Three quoted reviews from the cluster.

Where personas feed in

Prompt Runs — persona-segmented sessions prefix each prompt with persona context. See Prompt Runs.
Image Enhancer — the AI creative brief is written against the top three personas. See Image Enhancer.
Listing rewrite — FAQ block is generated using persona language. See Listing rewrite.
Sentiment — the voice-of-customer rollup pulls from clustered themes. See Sentiment.

Regenerating

Click "Regenerate" on the Personas page. The cluster is rebuilt from the current review corpus. We recommend doing this monthly, or after a major product launch when new reviews are likely to introduce a new cluster.

Common questions

What's the minimum number of reviews to generate personas?

Around 50 reviews per brand will produce coherent clusters. Below 50, the clusters tend to look like "people who liked it" and "people who didn't." Above 200, the clusters become specific enough to be useful for Prompt Runs.

Can I edit a persona after generation?

Yes. Each persona has an editable name, summary, and demographic descriptors. You can also merge or split personas if the clustering produced too many or too few.

Do personas update when new reviews arrive?

Yes, but on demand rather than automatically. The clusterer is opinionated about churn — running it weekly produces noisy persona drift. We recommend re-clustering monthly or after a major product launch.

Are persona descriptors usable as ad audiences?

They are not direct exports to ad platforms. The descriptor copy is useful for ad targeting briefs but the persona system is not connected to Meta or Google audience APIs.

Ask AI about this page

Open this topic directly in your favourite AI assistant — the prompt is prefilled.

Was this page helpful?

Thanks — recorded.

Feature →

Prompt Runs

Where persona-segmented runs use the cluster output.

Feature →

Sentiment

Theme extraction that complements persona clustering.

Try this yourself

See eCommerce Insights on your own catalog. The free trial covers your full SKU set across five AI engines.

Start free trial

LLM-friendly summary of this page

Personas converts a CSV of reviews into 4-8 named buyer personas via LLM clustering. Inputs: Yotpo, Judge.me, Okendo, Trustpilot, Stamped, Reviews.io CSV exports; direct API for Yotpo and Judge.me. Minimum useful corpus around 50 reviews; clusters become useful for Prompt Runs above 200 reviews. Clustering runs in two passes: first a topic-modelling pass to surface natural review groupings, then an LLM naming pass to give each cluster a human-readable persona with summary, demographics, and motivations. Output per persona: name, one-paragraph summary, top three feature loves, top three concerns, sample quoted reviews. Personas feed Prompt Runs (persona-segmented runs prefix the prompt with persona context), Image Enhancer (AI creative brief written against top three personas), Listing rewrite (FAQ block uses persona language). Editable: name, summary, demographics; can merge or split clusters. Regenerate on demand; recommended monthly or after major product launch. Not connected to ad platform audience APIs.