Drew Miller Logo

rew Miller

Notes for the Conveo deck

The reasoning behind the slides

The deck keeps each slide short. This page holds the evidence, the caveats and the sources, so you, or your AI agent, can check the full reasoning.

The move, and how it makes money

Turn the people who read Conveo’s research into the source of its next studies. Readers ask questions; the agent answers what past research can, groups the real gaps by the decision each one blocks, the researcher decides what to run, and the answer goes back to everyone who asked.

  • Questions are already usage. Talk to Your Data has a per-member allowance with billed overage, and Foundations prices questions at $2–5 per study a question touches after a free tier, or through annual question plans. Answered questions are value delivered; beyond an allowance they are also billed usage, and inside a plan they support renewal.
  • Gaps become studies. The researcher approves the ones worth running (Bet 1).
  • More people, more reasons to renew. Shared studies with buyers, answers that hold up in AI assistants, and a record of what research changed (Bets 2–4).

A research library can also replace new studies, so the loop is measured both ways. Being where a company’s research questions start could become hard to copy as decisions and context build up; an intake screen alone would not be.

Why now, and what has to be true

  • One research team got 55 requests with room for 15 (Julian Della Mattia, Finders to Builders ep. 29). That shows more demand than capacity, not hidden demand. Cheaper studies can clear part of the backlog; a ranked list says which part.
  • Shared reports began inviting reader questions on Sep 19, 2026, and readers sign in to ask.
  • “Speed and cost now is reaching the peak of its value proposition” (Nick Rich, Insighter’s Club, Jun 2026). Insight teams are often among the first cut when they are “seen as a service” (Transforming Insight ep. 73).
  1. Readers ask real questions the research can’t answer yet.
  2. The researcher finds some of them worth a study.
  3. Those studies are new or better work, not just a reshuffle.

Bet 1 (the lead): let every shared report suggest the next study

An agent answers reader questions from past research, collects the real gaps, and drafts a study for the researcher to approve.

Goal
More useful studies per account, from questions nobody had to spot. Questions answered from past research count as value too.
Trap to avoid
Treating every question as demand. Rank gaps by the decision each one blocks, and allow “use another method” or “not researchable as asked”.
First step
Ship it on one account’s shared reports, with the researcher approving every draft. Then four accounts that opt in.
It works if
Two accounts commit budget and a date to a study that began as their readers’ questions.
I’d drop it if
Readers rarely ask what the research can’t answer, or researchers don’t think the gaps are worth a study.
  • It builds on what Conveo already ships: follow-up questions on shared reports (Sep 19, 2026), where “picking one asks them to sign in with their email”, so every question has someone to answer later; the research library and Talk to Your Data; study setup; and Foundations.
  • Chat history already stores reader questions. Nothing public describes grouping the unanswered ones into a research agenda the researcher owns.
  • A restricted study is an access question, not a gap. Its content never reaches the answer; it is shown only as possibly relevant, with details the owner allows, and not at all if the owner keeps its existence private. The question stays open until access is granted, because permission may never come.
  • Before drafting, the agent pins down the decision, category, audience and comparison the askers meant, so a question about household budgets doesn’t turn into a snack study.
  • Ranking uses the decision each gap blocks, its date and owner, then how many people and teams asked. Three people from one meeting are grouped as one question, keeping each decision and person it serves.
  • A signed-in reader is known, but isn’t necessarily the decision or budget owner, so the researcher confirms who owns the decision before anything is drafted.
  • Starting assumptions to test, not facts: learning each team’s preferences needs a history of decisions (I’d start looking at about 30); pooled demand across brands is topic counts only, opt-in, with a minimum number of brands per topic (I’d start at five). Five brands alone is not a privacy guarantee, since a narrow topic can still reveal plans.
  • Conveo already suggests follow-up questions, so a click on a suggestion is weaker evidence than a question someone types, and both are weaker than one tied to a decision with an owner. They are counted separately.
  • Answered means the research covers the question, not that it is enough for the decision. The researcher can mark a covered question “not enough for this decision”, which keeps it open.
  • Both errors matter. A false “not covered” invents demand. A false “answered” can stop research a team needed.
  • The pilot measures researcher time, and whether approved studies are new spend or replace planned work. Study count, customer value and expansion revenue are measured separately, since a new study can use prepaid capacity. Swapping a weak planned study for a better one counts as a good outcome too.
  • I found no published figure for how often a reader’s question turns into a paid study, so that number has to come from Conveo’s own data.

Bet 2: shared studies with buyers

When several brands ask the same question, offer it as a Foundations study with one agreed scope.

Goal
Time to signed commitment and revenue per program. Fill rate alone would reward listing only easy wins.
Trap to avoid
Reading “proposed” as “stalled”, or exposing one brand’s plans. Pool only questions any competitor would ask the same way.
First step
Run it by hand with the Foundations team on one topic, with buyers agreeing on market, audience and timing.
I’d drop it if
Brands asking about the same topic won’t buy the same study.
  • Foundations already lists proposed programs that launch when enough founding subscribers commit, and founding buyers already get 50% off for life. The addition is choosing and selling topics from the questions brands actually ask.
  • Conveo’s own rule decides what belongs there: “if a competitor would phrase the question identically, it belongs in Foundations. If the answer is your advantage, it belongs in research that is yours alone.”
  • From outside I can’t see how long programs have been listed, how much sales effort they got, or whether buyers agree on scope, so the catalog count says little about why programs haven’t launched.
  • A shared topic isn’t a shared purchase: brands may need different markets, people, timing or depth. Count only signed commitments.

Bet 3: answers that hold up in AI assistants

When Claude or ChatGPT retells a Conveo answer, keep who was asked, how many, and what’s missing.

Goal
Conveo stays in the room where decisions happen, and questions asked there reach the same demand list.
Trap to avoid
Caveats written as labels the assistant drops. Put the caveat inside the sentence, and test named assistants rather than promise “any AI”.
First step
Add sources, counts and gaps to the answers Claude and ChatGPT get.
I’d drop it if
Nothing important gets lost. Then keep the tests and skip new fields.
  • Conveo’s MCP server (the connection that lets AI assistants query Conveo) already returns answers with sources and cited quotes, so citations are not new. The question is whether the answer a brand manager finally sees keeps scope, counts and gaps.
  • First check: whether the analysis tool already returns participant IDs and study scope. That decides whether this is a small schema change or a new retrieval step.
  • The answer must separate “not found in this search” from “not in the research”, since retrieval doesn’t read every transcript.
  • A missing group, like lapsed buyers, is disclosed. It becomes a study only if a decision needs it.
  • Re-test whenever the model, prompts, retrieval, permissions or data change. One 2025 study found state-of-the-art models reached incorrect conclusions on about 31% of hypotheses in its annotation tasks (Baumann et al.), a reason to treat any configuration change as a reason to re-check.
  • Other data vendors set the expectation: NielsenIQ customers asked to connect its data to Claude, ChatGPT, Gemini and Copilot, and YPulse’s CEO reported AI questions up 14x in 2026.

Bet 4: show what changed because of the research

Record the plan before the research, what changed, and what happened, so teams can defend their budget.

Goal
Insights teams can show what Conveo’s research changed, which makes the budget, and the renewal, easier to defend.
Trap to avoid
Turning it into a scorecard for Conveo. Research often earns its value by changing the question, not by predicting a launch.
First step
Pilot it as a service with a few customers, alongside Conveo’s Research Impact team.
I’d drop it if
We can’t capture the plan beforehand, even from existing briefs or a short conversation, or the outcomes never come back.
  • The plan is written before the research, because people rate decisions better after a good result.
  • It shows how often research changed the plan and how those plans did. It does not show that the research caused the result: budget, execution, timing and which projects get tracked all differ.
  • It credits unchanged decisions, launches avoided, uncertainty reduced and questions reframed, not just wins.
  • Survey and ad-test companies already publish validation of their own tests (Kantar 2022, Zappi 2024, System1 2025). I found no AI-interview company publishing a prospective record like this.

How I chose these

Set the goal line (Conveo’s job posts name net revenue retention and expansion ARR as the scoreboard), go wide, cut hard, try to kill the survivors with independent AI reviewers and fact checks, then build the smallest proof. 18 ideas went in; these were cut:

  • Already built (4): Run a study again; share findings across the company; a company-wide research library; alerts when a trend moves. Conveo’s changelog already has them.
  • Not where I’d start (3): A better AI interviewer; badges for researchers; out-volume a rival’s interview count. Interview quality is Conveo’s core system. I think the bigger opening is on the demand side.
  • Failed a fact check (3): Bridge studies for brands replacing trackers; pooled sales benchmarks; matching shoppers to purchases. Conveo adds to trackers; data licences and consent block the rest.
  • Merged into Bet 3 (4): Flag model changes; a catch rate for what AI missed; a blind-spot flag; answers inside AI assistants. Together they became one bet.
  • Needs Conveo’s data (2): Score accounts ready to expand; price credits on outputs, not minutes. Can’t judge from outside.

The ideas behind Bets 2 and 4 make 18. Bet 1 came later, from a conversation at AI Tinkerers.

Sources

Building something complex? Let's talk.

Looking for my next Product Engineer role at an AI-native company, on a small team with real ownership.

Get in Touch

Designed & Built by Drew Miller

© 2026. Version 3.3