Readers of shared reports ask follow-up questions. The ones past research can’t answer stay open, and nothing public describes grouping them.
An agent answers what it can, ranks the rest by the decision each blocks, and drafts a study for the researcher to decide on.
Answered questions are value delivered, and billable usage beyond the allowance. Open ones can point to the next study.
Shared reports already suggest follow-up questions, and picking one asks the reader to sign in. This starts after they ask.
Seven follow-up questions on one shared report.
41 interviews with current US buyers · shared by Maya
Readers never see each other’s questions. A restricted study’s title and owner show only if its owner allows it. Otherwise the reader learns only that a study exists and can request access, unless the owner keeps even that private; then the question simply stays open.
Open questions only, ranked by the decision each one blocks.
Nothing checked yet.
First by the decision a question blocks and when it’s needed, then by how many people and teams asked. Partly answered questions are logged but not pitched.
Pitches expire instead of piling up: once the “needed by” date passes, or a month goes by with no new asks. The count stays, so the topic comes back if people keep asking.
The team’s researcher. Her reason is what the agent learns from.
Nothing to review yet.
Every decision and reason is saved, and “no budget” counts as timing, not taste. Predicting what Maya will run comes later, once there’s enough history. About 30 decisions is a starting assumption to test, not a known threshold.
Talk to Your Data has a per-member allowance with billed overage. Foundations charges $2–5 per study a question touches after a free tier, or sells annual question plans.
When readers at several brands ask about the same topic, the agent can propose a shared study for Conveo’s marketplace. Accounts opt in.
Only a topic and a count leave a company, never question text or who asked. Conveo’s Foundations rule: “if a competitor would phrase the question identically, it belongs in Foundations”.
Starting assumption to test: at least five brands before a topic shows. Five brands isn’t a privacy guarantee: a narrow topic can still reveal one company’s plans, so narrow topics are reviewed before they show.
Both mistakes are costly. A false “not covered” invents demand. A false “answered” can stop research that’s needed. Each gets its own tests. A first test is below.
Each test must fail first, for the reason predicted, before the fix that makes it pass counts. Maya’s decline reasons become new test cases.
Answers show their source, sample size and date, and never stretch US evidence to a UK question. Drafted studies are checked for leading questions, and a researcher approves the guide and screener.
We tested both costly mistakes on made-up questions, with the right answers fixed before the AI saw them. Every run is reported, including the bad parts:
Weak spots: runs 2 and 3 both called questions answered by one person “partly” (Q02, Q03, Q05). The leak check looks for two or more words that appear only in restricted interviews, so it misses a paraphrase, a single telling word (it missed Q29) or a number. Small, self-written and with easy traps: a first check, not proof.
Details: the made-up library, the questions and answers, every AI answer.
“We got 55 requests. We only have capacity to tackle 15.”Julian Della Mattia, research consultant, on logging unmet demand to argue for budget · Finders to Builders, 2025
“Every study plan should be a download of what’s been known so far.”Jake Burghardt, author of Stop Wasting Research · NN/G UX Podcast, 2026
“No question is ever a finished question when you work on a client side.”Matt Roberts, VP Business Intelligence, Formula E · Insighter’s Club, 2025
Planning opens with the decisions waiting on research.
Every customer question gets an answer, now or from the next study.
It sees the next study before anyone books it.
Drew Miller · drewx.design/conveo · Example data throughout. Report features follow Conveo’s public changelog. Full notes