Last updated: 19 August 2026
Response rates for telephone surveys collapsed once it became cost-free to ignore a call. Computer-assisted telephone interviewing, or CATI, was built for a period when a ringing phone was usually answered. Caller ID, call filtering and sheer nuisance volume made most people unwilling to pick up an unfamiliar number.
The scale of the fall is documented rather than anecdotal. Pew Research Center's own telephone response rates dropped to 6% by 2018, down from around 9% a few years earlier, and Pew is an organization with more methodological resource than almost any commercial buyer.
This matters beyond telephone research, because the same failure is now available to any method that assumes reaching people is the easy part.
How Far Did Response Rates Actually Fall?
To roughly 6% at a well-resourced public research organization by 2018. That is the cleanest public marker available, and commercial fieldwork with less callback budget generally sits below it rather than above.
Two features of that number matter more than its size. It is a floor observed by an institution with strong incentives and deep method expertise, which makes it a generous rather than pessimistic reading of the industry.
And it is the endpoint of a long decline rather than a sudden break. Response rates had been falling for years before, which is why the industry adapted incrementally and largely without alarm until the arithmetic stopped working.
What Caused the Decline?
Four things compounding, none of which the research industry controlled. The method did not get worse. The environment it depended on changed underneath it.
- Caller identification. Once an unknown number is visibly unknown, declining costs nothing.
- Mobile substitution. Landline frames aged out, and mobile numbers carry different reachability and different regulation.
- Spam and fraud volume. Legitimate research calls became indistinguishable from the thing everyone was already screening.
- Carrier-level filtering. Calls now get labeled or blocked before a human decides anything.
Each was survivable alone. Together they changed who was still willing to answer, which is the variable that actually determines whether a sample means anything.
Does a Low Response Rate Mean the Data Is Biased?
Not automatically, and this is the most misunderstood part. A 6% response rate is not evidence of bias on its own. It becomes bias when the people who answer differ systematically from the people who do not, on the variables the study is measuring.
The distinction is practical rather than academic. A survey about breakfast cereal preference among a broadly reachable population may survive a low response rate. A survey about time poverty almost certainly will not, because the people hardest to reach are disproportionately the people the study is about.
The AAPOR standards treat coverage and nonresponse as distinct components of total survey error for exactly this reason. Conflating them leads teams either to panic about a number that does not matter or to ignore one that does.
Weighting helps and does not rescue. Adjusting a sample to population margins corrects for who you know is missing. It cannot correct for a difference on a variable you did not measure and cannot see.
Mode substitution carries its own measurement question. A 2026 study in the Journal of Medical Internet Research examined mode effects between mobile web and telephone surveys on the same instrument, finding differences that were present rather than negligible. The same selection effect is examined in sample validity and who you miss.
What Did the Industry Do About It?
Three responses, in rough order of adoption: more calls per complete, then mode-switching, then rebuilding the frame. The last of those meant moving to methods that did not depend on someone answering an unknown number.
More dialing was the first and least effective. It preserved the method while multiplying its cost, since every additional attempt is a paid interviewer minute at a declining hit rate.
Mode-switching came next. Online panels absorbed a large share of quantitative work, which solved the cost problem and introduced a different coverage problem, since panel membership is its own filter on who participates.
Automation is the current response. One published deployment ran an LLM-based telephone survey system across a United States pilot of 75 participants and a large-scale Peru deployment of 2,739. Separately, Gallup has begun publishing research on AI phone interviewing as a methodological question rather than a product.
| Response | What it fixed | What it did not fix | Cost profile |
|---|---|---|---|
| More call attempts | Completed sample size | Who was willing to answer | Rises steeply with each attempt |
| Online panels | Cost per complete, speed | Coverage; panel membership is a filter | Low per complete, hidden frame cost |
| Mixed-mode designs | Coverage across segments | Comparability between modes | Higher design and analysis overhead |
| Automated calling | Cost per attempt, parallelism | Willingness to answer, which is unchanged | Low per attempt, young evidence base |
Read the third column as the honest one. Every row solved a real problem, and not one of them solved the problem that caused the collapse.
Who Answers the Phone Now?
A narrower and older group than the general population in most markets. That skew is what makes a low response rate dangerous rather than merely expensive, because the people missing are not missing at random.
Reach is not the same as connectivity. The ITU's Facts and Figures 2025 reports mobile broadband coverage as nearly universal while quality and affordability gaps persist, and counts 2.2 billion people still offline, most in low and middle income countries. DataReportal's Digital 2026 report counts more than 6 billion people online, which is a different statement from being reachable for a scheduled interview.
Why Does This Cut Both Ways?
Because the populations hardest to reach by phone in high-income markets are often the easiest to reach by phone everywhere else. In markets where browser-based research selects hard for connectivity, income and urbanization, a voice call remains the wider frame rather than the narrower one. What replaced the traditional phone room is covered in AI CATI software replacing phone interviews.
That is the case for treating mode as a sampling decision rather than a procurement one. Interviews run inside WhatsApp are asynchronous and need no app install, which reaches people a fixed live slot excludes, and a managed program can field several modes against one instrument rather than forcing a single channel to carry every segment. The platform landscape for that channel is set out in AI phone interview platforms.
Does Automated Calling Fix Any of This?
No, and claiming otherwise is the most common overstatement in the category. Automation changes the economics of dialing. It does not change whether a person picks up an unknown number.
What it genuinely changes is throughput and cost per attempt. A study is no longer capped by interviewer headcount, calls run in parallel, and the marginal cost of an additional attempt falls sharply. At a 6% response rate, that is a meaningful commercial difference.
What it does not change is the composition of the people who answer. If the willing 6% differ from the market on the variable being measured, an automated dialer reaches the same skewed 6% more cheaply.
Treat any vendor claim of improved response rates through automation as the thing to verify first, because it is the claim least supported by anything published.
Automation has a published record rather than only vendor claims. A paper in JAMIA Open described automated survey collection with LLM-based conversational agents, covering both the interview and the extraction of structured answers from the transcript. How those voice approaches differ is worked through in AI phone interviews against IVR and voice bots.
What Replaced Phone, and For Which Segment?
Nothing replaced it wholesale. Different methods absorbed different segments, which is why the honest description of the last decade is fragmentation rather than substitution.
Online panels took the high-connectivity middle of most markets. That worked commercially and introduced a coverage question of its own, since panel membership selects for people willing to join a panel, which is not a neutral filter.
Messaging-based research absorbed a segment panels struggle with: respondents who are reachable but not on a schedule. An interview running inside WhatsApp is asynchronous and needs no app install, which reaches shift workers, caregivers and anyone whose connectivity is intermittent rather than absent.
Voice retained the segments the others cannot reach at all. Feature-phone users, low-bandwidth households and older respondents in many markets remain reachable by call and by very little else, which is why AI phone interviewing is a reach instrument rather than a cost-saving one.
Does Any Single Vendor Cover the Whole Frame?
Few do, and the ones that claim to are usually describing panel partnerships rather than fieldwork through browser-based interviews, WhatsApp and phone calls. The practical test is whether a vendor can run the same instrument across all three without separate procurements.
Alchemic runs managed fieldwork across fourteen markets, including the USA and the UK alongside South and Southeast Asia, the Gulf and Africa, across 57+ languages including Spanish, Hindi, Tamil, Bangla, Arabic and Indonesian. Whether that breadth matters depends entirely on how much of your market sits outside the browser frame, and for a US-only study of high-connectivity consumers it very likely does not.
Where a study genuinely spans several reachability tiers, the coordination problem is larger than the tooling problem, which is the case for treating it as a managed research program rather than a stack of point solutions.
Which Studies Feel This Most?
Continuous ones. A one-off study with a coverage problem produces a wrong answer once, while a tracker with a coverage problem produces a trend line that reflects changing sample composition rather than changing opinion.
That is the quiet damage from the response-rate collapse. Trackers built on telephone frames in the 1990s kept reporting movement long after the movement was partly in who answered, and separating the two retrospectively is close to impossible.
Anyone running a brand tracking program today should hold mode constant across waves or document precisely when it changed, because a mode switch mid-program is indistinguishable from a market shift in the output. Explaining a movement once you see one is covered in adding qualitative follow-up to brand tracking.
The same reasoning holds for any AI-moderated design that runs in waves. What matters is less the absolute response level than whether it stays consistent over time, since comparison is what makes a tracker useful.
Where This Analysis Has Limits
- One organization's rate is not the whole industry. Pew's 6% is the best public marker available, not a universal figure, and rates vary by market, topic and incentive.
- Response rate is a weak proxy for quality. A higher rate on a badly designed instrument is not better data.
- Some populations still answer. Older and rural respondents in several markets remain reachable by phone at rates that would surprise anyone extrapolating from urban experience.
- Regulation shapes the numbers. Calling rules, preference registers and permitted windows differ by country and change what is comparable.
- The automation evidence is thin. Early deployments and institutional research are encouraging, not conclusive.
- Nonresponse is only half the problem. Coverage decides who could have answered at all, and it is the failure that never appears in a completion report.
The ESOMAR code and guidelines and the Insights Association both address these obligations independently of mode.
What Should You Do With a Phone Study Today?
Design it as one mode among several rather than as the study, and be explicit about which segment it is carrying. Phone is now a reach instrument for specific populations rather than a default for general ones.
Three practical decisions follow. Decide which segments phone is genuinely the best route to, rather than defaulting to it or abandoning it wholesale. Measure completion against quota by segment rather than in aggregate, because aggregate rates hide exactly the collapse that matters. And record the mode and its known coverage limits in the method note, so a later reader can judge the finding rather than inherit it.