{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-08-21T09:56:11.152Z"},"content":[{"type":"documentation","id":"f8cc6740-cfb1-4a2c-b14f-e9bd8a35113e","slug":"unasked-research-questions-suppressed-demand","title":"The Questions Nobody Asks You: How a Slow Research Function Hides Its Own Demand (2026)","url":"https://www.koji.so/docs/unasked-research-questions-suppressed-demand","summary":"A research function can have a healthy queue, sane utilisation, rigorous methods and high satisfaction while informing a minority of decisions, because requests are filtered before they are made: nobody queues for an answer they believe will arrive too late. The answerable fraction is the probability that a decision's notice period exceeds research lead time. With notice periods averaging three weeks, a three-week lead time reaches about 37 percent of decisions and a two-day lead time about 88 percent. Returns to cutting lead time are multiplicative, not linear.","content":"**Bottom line up front:** A research team can have a healthy queue, a sane utilisation target, rigorous methods, true findings, and delighted stakeholders — and still be informing a small minority of the decisions it exists to inform. The reason is that requests are filtered before they are made. A decision that must be taken in two weeks is never brought to a team known to need six, so it never enters the backlog and never appears in any metric. The queue is not a sample of your organisation's questions; it is a sample of the questions somebody already believed you could answer in time. Measure the fraction of decisions whose notice period exceeds your lead time, and you will usually find it is under 20 percent.\n\n## Everything upstream can be fixed and the answer can still be no\n\nTwo operational failures explain most slow research functions. The first is inventory: [lead time equals work-in-progress divided by throughput](/docs/research-lead-time-littles-law), so a team with twelve open studies finishing two a week takes six weeks regardless of how good any study is. The second is capacity: [queue time scales with utilisation divided by one minus utilisation](/docs/research-team-utilization-queue-time), so a fully booked team is not slightly slower but several times slower.\n\nSuppose you fix both. WIP is limited, utilisation sits at 75 percent, lead time has come down to three weeks and holds. Stakeholder satisfaction is high. Every study that enters the queue leaves it on time, with a defensible finding.\n\nThere is still a question worth asking, and almost nobody asks it: **what fraction of the decisions this organisation made in the last quarter did research inform?**\n\nNot what fraction of *requests* were served — that number is near 100 percent, and it is the number every research function reports. The fraction of *decisions*. These are wildly different denominators, and the gap between them is where the entire value of a research function is decided.\n\n## The filter that runs before the request\n\nConsider a product manager with a decision to make and two weeks before it has to be made. They know, from experience, that research takes about three weeks. They do not file a request. They do not complain. They do not appear in any report of unmet demand. They ask two customers they already know, read some support tickets, and decide.\n\nNothing about that is irrational or disloyal. It is a correct local decision. And it is invisible.\n\nThis is the mechanism: **the queue only contains requests that survived a feasibility check the requester ran in their head.** People do not queue for things they believe they cannot get in time. The backlog is therefore not a measure of demand. It is a measure of demand *conditioned on* the requester's belief about your lead time.\n\nResearch teams are unusually well equipped to recognise this error and unusually prone to missing it, because it is a selection effect and they spend their careers catching selection effects in samples. Studying only customers who renewed tells you nothing about churn. Studying only the requests that were filed tells you nothing about the decisions that were made. It is [survivorship bias](/docs/survivorship-bias-customer-research) pointed at your own intake.\n\nWorse, the satisfaction score is computed on the same selected population. The stakeholders who are happy with research are, definitionally, the ones whose questions fit inside your lead time. The ones who routed around you are not unhappy — they are absent. A research function can approach maximum measured satisfaction while its actual coverage falls, and no instrument in the standard ResearchOps toolkit will register it.\n\n## The rerouting is visible in the industry data\n\nIf demand that cannot be served through the research team simply evaporated, this would be a smaller problem. It does not evaporate. It reappears elsewhere.\n\nMaze's *Future of User Research Report 2026*, based on responses from nearly 500 practitioners, reports that research is now being produced well outside dedicated research teams: **39 percent say product managers are conducting user research, 35 percent say market researchers are doing it, and 23 percent say marketers are involved.**\n\nThat is the suppressed demand, resurfacing under other job titles. Some of it is healthy democratisation. Some of it is a decision that needed evidence in two weeks finding the nearest available substitute. From inside the research team the two are indistinguishable, because neither one entered the queue.\n\nThe same report finds demand still climbing — the share of participants reporting increased demand for research went from nearly 55 percent last year to **66 percent** this year — while **13 percent of organisations have no resources at all** to support non-researchers running studies. Rising demand, a hard capacity ceiling, and no safe channel for the overflow is precisely the configuration that produces a large, invisible, unmanaged research function operating outside the research function.\n\n## The metric: your answerable fraction\n\nThe measurement problem is that suppressed demand leaves no record. But it can be modelled, and the model is simple enough to compute on the back of an envelope.\n\nEvery decision has a **notice period**: the time between the moment the question becomes askable and the moment the decision must be made. Your research function has a **lead time**. A decision is informable if and only if its notice period is at least your lead time.\n\n**Answerable fraction = P(notice period ≥ lead time)**\n\nNotice periods are strongly right-skewed. A few decisions are visible months ahead (annual pricing, roadmap planning); most surface with days or a couple of weeks of warning. Modelling notice periods as exponentially distributed with a mean of three weeks:\n\n| Your lead time | Answerable fraction |\n| --- | --- |\n| 6 weeks | 13.5% |\n| 4 weeks | 26.4% |\n| 3 weeks | 36.8% |\n| 2 weeks | 51.3% |\n| 1 week | 71.7% |\n| 2 days | 87.5% |\n\nTwo things fall out of this table, and both are more important than the specific percentages.\n\n**First, the well-run three-week team is serving about a third of its addressable decisions.** Not because it is failing — it is not — but because two-thirds of the decisions never had three weeks available. Its request-fulfilment rate is 100 percent and its coverage is 37 percent.\n\n**Second, the returns to cutting lead time are multiplicative, not linear.** Halving lead time from six weeks to three does not add a fixed increment of coverage; it multiplies coverage by 2.7. Cutting it from six weeks to one multiplies coverage by 5.3. Each week you remove is worth more than the week before it, because you are moving into the fat part of the distribution where most decisions live.\n\nThat convexity is the whole strategic case for research speed, and it is invisible to any team measuring itself on requests served.\n\nThe obvious objection is that the exponential assumption is doing the work. It is not. Repeating the calculation with a lognormal distribution of notice periods (median two weeks, sigma of 1) gives 13.6 percent at six weeks, 34.3 percent at three weeks, and 75.6 percent at one week — materially the same picture. The conclusion is driven by the right skew of notice periods, which is a robust feature of how organisational decisions actually surface, not by the particular distribution chosen.\n\n## Why nobody owns this number\n\nThe reason this failure survives in otherwise well-run functions is structural, and it is worth naming precisely.\n\n**The requester knows the notice period. The researcher knows the lead time. No artifact in the system contains both.**\n\nThe intake form does not fix this, and this is the subtle part. Intake forms do ask \"when do you need this by?\" — but the answer is given by someone who has already adjusted their expectations to what they believe is possible, and who has already declined to file the requests that failed their own feasibility check. The date on the form is conditioned on the very constraint you are trying to measure. The data is contaminated at the point of collection.\n\nSo the situation is: a healthy queue (leg one), a healthy capacity buffer (leg two), rigorous studies, and a number that determines the function's entire value which nobody computes, nobody is accountable for, and which the standard instrumentation actively obscures. No component of the system failed. The system does not have this component.\n\n## How to actually measure it\n\nYou cannot observe suppressed demand directly. You can observe it three indirect ways, and all three are cheap.\n\n**1. Sample decisions, not requests.** Take the last quarter and list the twenty most consequential decisions the organisation actually made — from roadmap changes and pricing moves to positioning and hiring. For each, record whether research informed it, and if not, why not. The \"why not\" answers sort into three piles: nobody thought of it, research was too slow, and research could not have helped. Only the middle pile is your problem, and it is usually the biggest.\n\n**2. Ask the question that surfaces the filter.** Add one item to your stakeholder survey: *\"In the last quarter, was there a decision you would have liked research on but did not request?\"* followed by *\"Why not?\"* This converts an absence into a datum. Expect the answer volume to be uncomfortable the first time.\n\n**3. Watch the tell.** Very high satisfaction combined with a long lead time is not a healthy signal — it is the signature of a function serving only the easy end of the distribution. Satisfaction measured on a selected sample tells you about the selection, not the service.\n\nThen report the ratio that matters: **decisions informed / decisions made**, alongside lead time. Requests served / requests received should be retired, or at minimum never reported alone. It is a measure of the filter, not of the function.\n\n## How Koji changes the shape of the problem\n\nEverything above says the same thing: coverage is a function of lead time, and the relationship is convex. That makes speed a strategic property rather than a convenience — and it is why the ceiling on human-moderated research is not a scheduling annoyance but a hard cap on how much of an organisation's decision-making research can reach at all.\n\nKoji attacks the term the model is most sensitive to:\n\n- **AI-moderated interviews run in parallel**, so fieldwork stops being a multi-week scheduling exercise. Twenty conversations can happen at once rather than across three weeks of calendar negotiation.\n- **Voice and text interviews** let participants respond asynchronously, removing the scheduling round-trip that puts a floor under every traditional qualitative timeline.\n- **Automatic thematic analysis** and **real-time reporting** mean findings exist as the interviews land, rather than a week later.\n- **Customisable AI consultants** encode your standards into the instrument, so a fast study is not a lower-quality study, and quality does not wait on a scarce senior reviewer.\n- **Structured questions** let one study do the work of two. Koji supports six types — `open_ended`, `scale`, `single_choice`, `multiple_choice`, `ranking`, and `yes_no` — so a single instrument returns a quantified distribution and the reasoning behind it. Two studies collapsing into one removes an entire lead time from the path to an answer.\n\nMoving a function from a three-week lead time to a two-day lead time is not a 90 percent efficiency gain. On the model above it takes answerable coverage from roughly 37 percent to roughly 88 percent — the function becomes able to serve most of the decisions it previously could not reach.\n\nThere is a second-order effect that matters more over time. Once the organisation learns that research is fast, the internal feasibility filter relaxes, and requests that were never filed start arriving. Expect the backlog to *grow* after a successful speed programme. That growth is not a problem to be managed down; it is the previously suppressed demand becoming visible for the first time, and it is the clearest available evidence that the programme worked. This is also the point at which [democratisation](/docs/research-democratization-playbook) stops being a slogan: self-serve templates and guardrails give the short-notice questions a safe channel instead of an improvised one.\n\n## Common mistakes\n\n**Reporting requests served as coverage.** It is the one metric guaranteed to look good regardless of how much of the organisation you are failing, because the denominator is filtered by the very constraint you should be measuring.\n\n**Reading high satisfaction as high impact.** Satisfaction is computed over served requests. A function serving 15 percent of decisions can score near the top on satisfaction and be strategically irrelevant.\n\n**Treating the growing backlog as a failure.** After lead time falls, request volume rises because the internal filter relaxes. Reading that as \"we are overwhelmed\" and re-imposing a slow, heavyweight intake process rebuilds the filter you just dismantled.\n\n**Trusting the date on the intake form.** It is given by someone who has already adjusted to your reputation, and it excludes everyone who did not file at all. Use it for scheduling, never for demand estimation.\n\n**Solving it with prioritisation.** Ranking the backlog better cannot recover a decision that never entered the backlog. Prioritisation operates on the surviving sample.\n\n**Assuming unserved demand disappears.** It becomes a product manager interviewing two customers on a Thursday. The research still happens; it just happens without method, sampling, or anyone checking whether the conclusion holds.\n\n## Frequently asked questions\n\n### What is suppressed research demand?\n\nSuppressed research demand is the set of decisions that would have benefited from research but were never requested, because the requester judged in advance that research would not arrive in time. It leaves no trace in the backlog, in intake data, or in satisfaction scores, which is why research functions systematically underestimate it.\n\n### How do I calculate my research team's answerable fraction?\n\nEstimate the distribution of notice periods for decisions in your organisation — the time between a question becoming askable and the decision being due — then compute the share of that distribution at or above your current lead time. With notice periods averaging three weeks and a three-week lead time, roughly 37 percent of decisions are informable. Cutting lead time to one week raises it to about 72 percent.\n\n### Why does cutting research lead time have multiplicative returns?\n\nBecause notice periods are right-skewed: most decisions surface with days or a couple of weeks of warning, and only a few are visible months ahead. Reducing lead time moves you into the dense part of that distribution, so each week removed captures more decisions than the week before. Going from six weeks to three multiplies coverage by about 2.7; going from six weeks to one multiplies it by about 5.3.\n\n### Is a growing research backlog a bad sign?\n\nNot after a speed improvement. When lead time falls, stakeholders stop filtering out requests they previously assumed were impossible, so volume rises. A backlog that grows after lead time drops is previously invisible demand becoming visible, and it is the best available evidence that the improvement was real.\n\n### Why is stakeholder satisfaction a misleading measure of research impact?\n\nBecause it is measured only among people whose requests you accepted, and those are the people whose questions happened to fit your lead time. Everyone who routed around research is absent from the sample. Satisfaction therefore measures the quality of your selection, not your coverage of the organisation's decisions.\n\n### What should I measure instead of requests served?\n\nMeasure decisions informed divided by decisions made, sampled from an independent list of the organisation's actual decisions rather than from your own intake records. Report it alongside lead time, since the two are causally linked, and add a survey question asking stakeholders about decisions they wanted research on but did not request.\n\n## Related Resources\n\n- [Research Lead Time: Why a Six-Day Study Takes Six Weeks](/docs/research-lead-time-littles-law) — the inventory arithmetic that sets lead time.\n- [Why a Fully Booked Research Team Is Slower](/docs/research-team-utilization-queue-time) — the capacity curve behind the same number.\n- [Expected Value of Information](/docs/value-of-information-research-decisions) — whether a study is worth running once you can run it in time.\n- [Evidence or Ammunition](/docs/research-as-ammunition-legitimation) — the other reason a rigorous study can be worth nothing.\n- [Structured Questions Guide](/docs/structured-questions-guide) — the six question types that collapse two studies into one.\n- [Research Democratization Playbook](/docs/research-democratization-playbook) — giving short-notice questions a safe channel.\n","category":"Research Operations","lastModified":"2026-08-21T03:26:55.189567+00:00","metaTitle":"Suppressed Research Demand: The Questions Nobody Asks You (2026)","metaDescription":"Requests served is a filtered metric. Compute your answerable fraction, why returns to speed are multiplicative, and how to measure demand you cannot see.","keywords":["research demand vs capacity","suppressed research demand","research coverage","research impact measurement","answerable fraction","research selection bias","research lead time"],"aiSummary":"A research function can have a healthy queue, sane utilisation, rigorous methods and high satisfaction while informing a minority of decisions, because requests are filtered before they are made: nobody queues for an answer they believe will arrive too late. The answerable fraction is the probability that a decision's notice period exceeds research lead time. With notice periods averaging three weeks, a three-week lead time reaches about 37 percent of decisions and a two-day lead time about 88 percent. Returns to cutting lead time are multiplicative, not linear.","aiPrerequisites":["Familiarity with research intake and how requests reach your team","Understanding that research lead time is a measurable quantity"],"aiLearningOutcomes":["Explain why a research backlog is a selected sample of organisational demand","Compute an answerable fraction from notice periods and lead time","Measure suppressed demand using decision sampling rather than intake data","Replace requests-served with a coverage metric"],"aiDifficulty":"advanced","aiEstimatedTime":"13 min"}],"pagination":{"total":1,"returned":1,"offset":0}}