Evidence quotes on clusters
When Usero classifies a feedback item, it also pulls out up to two exact sentences worth highlighting, checks each one is a true substring of what the customer wrote, and pins them to the cluster. You read the evidence in their words, then open the PR.
A cluster labeled "dark mode contrast issues, 6 reports" tells you a theme exists. It does not let you feel it. The sentence "Settings page text is invisible in dark mode" does, and someone wrote that sentence in your inbox already.
Tools built around customer evidence make a PM manufacture it by hand. In Productboard, the loop per snippet is: read the note, select the sentence, click a floating icon, search the feature tree, pick or create a feature, attach a customer record if the note lacks one, then rate importance on an unlabeled chip. We ran that loop in a trial workspace and a single two-snippet note took a dozen deliberate interactions. The output is worth having, a feature backed by verbatim customer sentences with names attached, but the price is a manual triage habit that dies the week the inbox gets busy. There is a second cost most teams never notice until it bites: when the highlight lives as an anchor inside editable rich text, editing the note can destroy the evidence. We watched an undo silently delete an insight link.
How it works
Disclosure: I build Usero, so weigh that. Usero already clusters feedback with AI, so the expensive question, which theme does this sentence belong to, is answered before a human arrives. Evidence quotes add the missing atom. The same AI pass that classifies each incoming feedback item extracts up to two verbatim quotes, the sentences a PM would have highlighted. The server keeps a quote only if it is an exact substring of the stored comment after whitespace normalization, because a paraphrase presented in quote marks is a fabricated customer quote. Survivors are stored as immutable rows referencing the source feedback, never as anchors in editable text, and the cluster detail page shows them in an "In their words" section with the person, the source channel, the date, and a click-through to the full report. From the same page you open a GitHub pull request for the cluster, with the customers' sentences sitting above the button.
No text selection, no feature-tree search, no importance chips. Extraction runs inside the classification pass that already touches every feedback item, so the evidence appears on the cluster the moment the item is filed.
Every extracted quote is validated server-side as an exact substring of the original comment. Anything the model reworded is dropped silently. If a quote renders on a cluster, the customer wrote those words.
Quotes are stored as their own rows referencing the source feedback, never as anchors inside editable rich text. No later edit, undo, or reformat can delete the evidence, which is a failure mode we observed first-hand in a competitor.
The quotes sit on the same cluster page as the Draft PR button. The theme, the verbatim sentences behind it, and the first pass at the fix live in one place instead of a triage tool and a repo that never meet.
The honest objection
Both are real limits today. The extraction cap is two quotes per item, chosen to keep the evidence section readable rather than exhaustive; a long email with five complaints will usually be split across clusters by classification anyway, but its quotes still cap at two. And quotes only exist for feedback classified after the feature shipped. There is no backfill pass for the old inbox yet, and no manual highlight-to-quote UI. If you want a PM hand-curating highlights across every historical ticket, Productboard does that job today and this deliberately does not.
FAQ
During classification, the AI pass reads the feedback item and extracts up to two verbatim sentences that best carry the request or the pain, the ones a PM would highlight. Some items yield zero quotes; boilerplate and pleasantries are skipped.
It can try, and the server throws the attempt away. Each candidate quote is checked against the stored comment and kept only if it is an exact substring after whitespace normalization. Rewordings, summaries, and stitched-together fragments all fail the check and are dropped.
An "In their words" section: each quote rendered verbatim with the person who wrote it, the source channel, the date, and a link to the full feedback item. Clusters with no extracted quotes omit the section.
No, and there is no UI for it. Clustering decides which theme a feedback item belongs to, and its quotes travel with it. The manual select-and-link loop other tools use is the step this feature exists to remove.
Backfill for feedback classified before the feature shipped, a manual highlight-to-quote flow, per-quote curation (pin, hide, reorder), and quote text inside the AI-generated PR description. The current slice is extraction at classification time, the substring guard, and the evidence section on cluster detail.
Free tier. No credit card. Two-minute install. The AI opens the PR, you merge it.
Keep reading