Older Consumers on Synthetic Panels: What Holds at 65+
By PersonaHive Team
TL;DR
Older consumers are the cell where a census-grounded synthetic panel covers the most people an online panel misses, and where the model answering for them has the weakest record. Online panels under-reach older adults: Pew Research Center found 14 percent of Americans 65 and older online almost constantly in 2025, against 63 percent of adults 18 to 29, and the UK's ONS counted 54 percent of people 75 and over as recent internet users in 2020. A census counts everyone, so the composition of a synthetic 65+ cell is right. The answers are another matter: Stanford and Columbia researchers found the 65+ group poorly reflected by every model they tested, and steering did not fix it. Validate older cells against probability or mixed-mode data, never against an opt-in online panel.
Can a synthetic panel represent older consumers aged 65 and over?
Partly. A census-grounded synthetic panel gets the composition of an older cell right, including the people who rarely or never go online, because a census counts them. It is weaker on what those people think, because the model answering for them represents the 65+ group poorly, and steering it by age does not close the gap.
Older consumers buy in almost every category you might test, from banking and travel to groceries and insurance. They are also the age group an online panel reaches least well.
That makes the 65+ cell a strange place for a synthetic panel. It is where the method covers the most people that an online panel misses. It is also where the published evidence on model answers is thinnest.
Both halves are true at once, so this post takes them in order: who an online panel loses, what census grounding gives back, what the model gets wrong, and which benchmark can tell you whether a synthetic older cell is any good.
Why do online panels under-reach older consumers?
Because an online panel can only recruit people who are online, and older adults are online less often and less heavily. In Pew Research Center's 2025 reference survey, 14 percent of Americans 65 and older said they were online almost constantly, against 63 percent of adults 18 to 29. Lower use leaves a smaller pool to recruit from.
Pew Research Center surveyed 5,022 US adults between February and June 2025. "A majority of adults ages 18 to 29 (63%) say they're online almost constantly. This share falls to 14% of those 65 and older." Smartphone ownership follows the same slope: 97 percent of adults under 50, 78 percent of those 65 and older.¹
The older the band, the steeper the drop. The UK's Office for National Statistics reported in April 2021 that 99 percent of adults aged 16 to 44 were recent internet users, "compared with 54% of adults aged 75 years and over."² ONS asked readers to treat the 75+ figure with caution, and discontinued the series in 2023, pointing readers to Ofcom's annual Technology Tracker for similar data.²
Weighting does not repair this. A panel can be raked so its 75+ count matches the census, but every person inside that cell was recruited online. You get the right number of the wrong people, a limit set out in what weighting a synthetic panel fixes.
Probability panels, recruited from a random sample of addresses rather than from volunteers, pay to close the gap. Pew's American Trends Panel has used address-based recruitment since 2018, inviting a random sample of US households, and "Panelists who do not have internet access at home are provided with a tablet and wireless internet connection."³ An opt-in access panel, by design, recruits only people who are already online.
| Indicator | Younger adults | Older adults | Source and date |
|---|---|---|---|
| Online almost constantly, US | 63% of 18 to 29 | 14% of 65 and older | Pew Research Center, fielded Feb to Jun 2025 |
| Own a smartphone, US | 97% of under 50 | 78% of 65 and older | Pew Research Center, fielded Feb to Jun 2025 |
| Recent internet users, UK | 99% of 16 to 44 | 54% of 75 and over | ONS, 2020 data, released April 2021 |
What does census grounding fix for the 65+ cell?
Composition. A national census counts people whether or not they use the internet, so a panel drawn from census totals includes the offline 80-year-old in the right proportion. That is the part an online panel cannot buy at any weight. It does not touch the answers.
A census-grounded panel starts from the population, not from a list of volunteers. PersonaHive builds its personas from aggregated public statistics, grounds them in each country's own national census data, and validates them against real surveys.
For the older cell, that matters more than anywhere else. The draw contains people in the oldest age bands and people who never go online, in the proportions the census reports. No recruitment channel filters them out first.
Two cautions apply. The census itself has a reference date that ages between rounds, covered in what goes stale on a census-grounded panel. And composition is only the first layer. Every persona's answer still comes from a language model.
How well do language models represent older adults' opinions?
Poorly, on the best published evidence. A Stanford and Columbia study compared nine language models with 60 US demographic groups across 1,498 Pew questions and named 65+ among the groups poorly represented by all models. Prompting a model to answer as a group helped modestly, and none of the representation problems were resolved by it.
In 2023, Shibani Santurkar and colleagues built OpinionQA from 15 waves of Pew's American Trends Panel, then measured how closely each model's answer distribution matched each group's.⁴ The groups that came out worst were consistent: "poorly represented by all models: e.g., 65+, Mormon and widowed."⁴
They then tried the obvious fix, telling the model which group to answer as. "Most models do tend to become better-aligned with a group when prompted to behave like it," they wrote. "However, these improvements are modest: none of the aforementioned representativeness problems are resolved by steering."⁴
A second effect stacks on top. In September 2026, Pew Research Center's Data Labs team replicated questions from three of its own 2026 survey waves with synthetic respondents and found that "AI respondents tended to eliminate a great deal of this diversity." Fully 47 percent of synthetic questions had at least one answer choice no synthetic respondent picked, which never happened in the human waves.⁵
One plausible reason the two findings meet in the older cell, which neither study set out to test: a model learns opinion from text, and text is written mostly by people who are online often. A census can count a person who never posts. A model has little of that person to read.
A census can count a person who never goes online. A model can only read the people who do.
Which benchmark can validate a synthetic older cell?
One that reached older adults offline. If your human benchmark is an opt-in online panel, its 75+ cell is the online minority, so a synthetic cell that matches it may be matching the wrong people. Use a probability-based panel that equips offline households, or a government survey that is nationally representative by age.
Validation is a comparison, and a comparison is only as good as its reference. For most age bands an online benchmark is a defensible reference. For the oldest bands it carries the exact coverage gap you are trying to test, so agreement proves less than it appears to.
This is a different problem from fraud or bots in a human panel, which grading your human benchmark covers. A perfectly clean online panel still under-covers people who are rarely online.
The practical rule is to ask how the benchmark reached its oldest respondents. A probability panel that sends offline panelists a tablet clears that bar.³ So does a mixed-mode survey, meaning one with at least one route, such as mail or phone, that does not need the internet. So does a national statistics or regulator survey designed to compare age groups.
PersonaHive's own validation report uses one such reference: the Consumer Financial Protection Bureau's 2024 National Age-Friendly Banking Survey, a "nationally representative survey of adults (ages 18 and older)" built to compare how different age groups experience banks.⁶ PersonaHive reports that its probabilistic method "reduced the overall spread difference by 68%" against a majority-consensus baseline, and describes the test as "one controlled benchmark: one survey, 12 questions, one country, and 200 personas."⁷ That is a starting point for older-cell validation, not a finished one. It is also the question to put to any vendor: name the reference, and say how it reached its oldest respondents.
What can you run on a synthetic panel for an older audience?
Jobs that rest on composition, and comparisons inside the older cell that you will check later. Sizing and quota design come straight from census totals. Ranking written claims among older personas works as a screen. Absolute levels, digital behavior and anything about the 75+ band need older people answering.
Sort the brief by which layer each output depends on. Composition comes from the census and is the strong layer. Opinion comes from the model and is the weak one for this age group.
Composition jobs run with confidence. How many 65 to 74 and 75+ people sit in each region, household type and income band of your market is a census question, and a census-grounded draw starts from exactly those totals. Use it to set quotas for the live study, then take the offline share from your statistics office's internet-use survey.
Ranking jobs run as screens. Putting ten written claims in front of an older cell and keeping the top three is a reasonable first cut, as long as the three then go to older people.
Level readings do not run. A synthetic share of 75-year-olds who would pay for a grocery delivery subscription is a number the evidence above does not support for this age band.
| Older-audience output | Which layer it rests on | On a synthetic panel |
|---|---|---|
| Size of 65 to 74 and 75+ cells by region and household | Census totals | Runs, and sets your live-study quotas |
| Ranking written claims within the 65+ cell | Model, relative | Runs as a screen, then confirm with older respondents |
| Share of 65+ who would buy, switch or pay | Model, absolute | No, the 65+ group is poorly represented |
| Anything about the 75+ band on its own | Model, absolute | No, the weakest cell with the thinnest benchmark |
| Digital habits of rarely-online consumers | Model, absolute | No, the model has little text from them |
| Validation of any of the above | Benchmark | Only against a probability or mixed-mode survey |
Frequently asked questions about older consumers on a synthetic panel
Five questions come up once the two layers are separated: whether a richer persona profile closes the gap, whether 65+ counts as a low-incidence audience, whether the internet-use figures travel across countries, how to report a synthetic 65+ result, and whether the same logic applies to younger audiences.
Does adding more detail to an older persona's profile fix the gap? The evidence says not by itself. Santurkar and colleagues found that prompting a model as a group improved alignment only modestly and resolved none of the representation problems.⁴
Is 65+ a low-incidence audience? No. Older adults are a large, well-counted population. The difficulty is reaching them online and modeling their answers, not finding them in the census.
Are the internet-use figures the same in every country? No. The Pew figures are US and the ONS figures are UK.¹ ² Check the statistics office for each market, and note whether its survey covers the oldest bands at all.
Can I report a synthetic 65+ result in a deck? Label it as simulated, state what it was calibrated and validated against, and keep it apart from rows that came from older people.
Does the same logic apply to younger audiences? The shape differs. Under-18s raise consent and census-counting questions instead, covered in child and teen audiences on a synthetic panel.
What should you do before your next study of older consumers?
Split the brief in two before you field anything. Put composition and claim screening on the synthetic panel, where census grounding carries them. Put every level reading for the 65+ and 75+ cells on a probability-based or mixed-mode sample of older people, and use that same sample as the benchmark for the synthetic cell.
Take the questionnaire you plan to field with older consumers and mark each output by layer: census, relative model reading or absolute model reading. Only the first two belong on a synthetic panel.
You can run the composition and screening pass on a free PersonaHive account, which comes with free credits and needs no credit card. Defining a 65+ audience in one of the nine national panels is where to start.
Then spend the live budget where it buys the most: on older people, reached by a method that does not require them to be online.
Sources
- What we know about internet use, smartphone ownership and digital divides in the U.S., Colleen McClain and William Bishop, Pew Research Center
- Internet users, UK: 2020, Office for National Statistics
- Methodology (American Trends Panel Wave 127), Olivia Sidoti and Emily A. Vogels, Pew Research Center
- Whose Opinions Do Language Models Reflect?, Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang and Tatsunori Hashimoto, arXiv
- How well synthetic polls capture the diversity of public opinion, Athena Chapekis, Arnold Lau, Samuel Bestvater, Sono Shah, Andrew Mercer and Aaron Smith, Pew Research Center
- National Age-Friendly Banking Survey Data, Consumer Financial Protection Bureau
- PersonaHive Validation Report: Beyond the Average Answer, PersonaHive