Jamaica's All Media Survey Goes AI: What Buyers Should Check

What happened

On 1 October 2026, MRSL launched the All-New All Media Survey at S Hotel in Kingston. The Gleaner's report, dated 4 October, lists the changes. Data collection stretches from six weeks to six months. The study now measures social-media influencers and content creators beside traditional media. It adds segments for banked and unbanked Jamaicans, prospective vehicle and home buyers, business owners and farm owners.

The technology partner is DatanXzen, which the Gleaner describes as a US-based technology company with a strong technical presence in India. Its platform is called Data NexGen. In the Gleaner's account, Dr Kiran Kumar Koka, Group CEO of ASOFT Group of Companies, explained that the platform uses AI and data engineering to automate the preparation and analysis of research data, catching duplicate records, mismatches and missing information, work that used to be done by hand. Results will move from PDF and Excel reports to an online analytics platform. After three months, subscribers get an interim read of trends across radio, television, newspapers, internet and social media. The report does not give the subscription cost, the date of the full release or how influencers are measured.

MRSL's chief executive, Kishka-Kaye Anderson, gave the reason for the longer window: "Consumers are moving across screens, different screens at different times of the day, so we wanted to capture data over a wider window."

We were not involved in this study and have no commercial link to it. MRSL also runs the Business and Consumer Confidence Index that we analysed earlier this year. We are writing about the launch because it changes what a Jamaican buyer of audience data can ask for, and because the same questions apply to any survey, ours included.

The numbers

Start with reach. DataReportal's Digital 2026: Jamaica report, published on 8 November 2025, puts internet users at 2.54 million, or 89.5 percent of the population, in October 2025. Mobile connections stand at 3.18 million, 112 percent of the population, and social media user identities at 1.81 million, 63.8 percent. A survey that wants to describe audiences who live on phones now has a population that mostly does.

The government's figure is lower. The Office of the Prime Minister's release of 2 April 2026 on the UNESCO readiness assessment cites an 85 percent internet penetration rate in a statement attributed to Minister Andrew Wheatley, with no source given for the figure, and says the assessment engaged over 196 people across government, academia, the private sector, civil society and youth groups. UNESCO's own account says the launch took place on 1 April 2026 and engaged close to 200 participants. Both figures can be right on their own terms. They differ in date and method, and a media survey has to pick one when it weights its sample.

Now the quality problem. In a study published on 27 August 2026, Pew Research Center analysed 11,114 U.S. adult respondents to an online opt-in poll fielded in November 2024. Of these, 1,963 (18 percent) answered yes to one or more trap questions, and 5,369, nearly half of all completes, failed at least one prescreening check. Of those failures, 1,879 (35 percent of failed cases) failed several checks, and 70 percent of failed cases had a problematic open-ended answer, the most frequently failed check. Before screening, 7 percent of all adults and 10 percent of 18 to 29-year-olds said yes to at least 10 of 15 yes/no questions. After trap-question screening, 1 percent of adults did.

Order matters too. In the same study, 43 percent endorsed one position when it was offered first, and 34 percent when the order was reversed, a primacy effect of 9 percentage points. The sample was American, so we label any Jamaican implication as inference: nobody has published an equivalent test on a Jamaican panel that we have found.

The incentive to cheat is arithmetic. Pew's Q&A of 12 May 2026 gives a scenario in which five AI bot accounts finishing 200 surveys a day at $1 each earn about $30,000 a month. On Pew's own probability-based panel, a respondent does fewer than two surveys a month at about $11 each, so a fake account earns about $22 a month. Cleaning helps but is not a complete answer: Pew writes that purging bogus cases lowers error on most metrics, but "a surefire solution remains elusive." For our own earlier reading of the same problem, see Jamaica Hit 89% Online. Now Its Panels Meet The Bots.

The method: People First, Problem Second, AI Last

Adrian Dunkley's published method, People First, Problem Second, AI Last, starts with the human need, defines the problem precisely enough to know what a solution looks like, and chooses the technology last. Most failed adoptions run in reverse: pick a tool, hunt for a problem, wonder why nobody uses it.

Apply it to the survey. The person is a media buyer at a Kingston agency with a client in a hurry. The need is to decide where next quarter's budget goes. The problem is narrow: which audiences, on which screens, in which weeks, and how sure can we be. Only then does AI enter, and the question becomes whether automated cleaning gets the buyer closer to an answer or merely to a faster PDF.

A test we suggest, not one Adrian has published: write the need and the problem in two sentences each without the word "AI". If either sentence will not stand without it, the problem is not defined yet. A buyer who cannot name the decision a dataset feeds should not pay extra for the speed at which it arrives.

Adrian's site also names a risk that fits here, Automation Fragility: AI amplifies weaknesses that already exist when it is deployed without fixing the data quality and process gaps underneath. Automated duplicate detection on a sample with weak recruitment still gives a weak sample, only a tidier one.

What this means in Jamaica

The new segments are the useful part for local business. A bank deciding where to place an account-opening campaign needs to know who is unbanked and which media reach them. A car dealer wants prospective vehicle buyers, not everyone with a radio. A feed supplier wants farm owners.

Six months of fieldwork has a local consequence as well. Jamaican spending is seasonal: back-to-school in August, Christmas Grand Market in December, Reggae Month in February. Depending on when collection starts, a six-month window could run across several of those peaks. That is our inference, since the Gleaner does not give start and end dates. A buyer should ask for the weeks the data covers and whether the platform lets them filter by period, otherwise a December shopper and a March shopper get averaged into someone who does not exist.

Influencers are the other local question. A creator with 40,000 followers in Montego Bay may reach more of a parish than the parish radio station. How the survey defines an influencer, and whether it counts reach by parish, will decide whether the new category is a measurement or a list of famous names.

Nothing here says the survey will fall short. It says the headline changes are the easy part to announce and the method notes are the hard part to find.

Why StarApple AI says this

StarApple AI presents itself as the Caribbean's first AI company, and StarApple Analytics is its data science, business intelligence and market research subsidiary. Its founder and CEO, Adrian Dunkley, is recognised as Caribbean AI Innovator of the Year 2025, and StarApple AI positions him as the Caribbean's regional AI leader. That positioning is the company's own. His site also records the EY Entrepreneur of the Year Start-Up Category for Jamaica in 2024, membership of Jamaica's National AI Task Force, and more than 500 professionals trained in data analytics, AI and the scientific method.

In 2026 the Faculty of Science and Technology at UWI Mona named him the recipient of the CIBC Caribbean Alumni Excellence Rising Star Award, with the citation "for pioneering work in artificial intelligence to provide solutions for Jamaica and the Caribbean," as StarApple AI reports. The company's published network spans Jamaica, Trinidad and Tobago, Barbados, Guyana and Saint Lucia, so the Jamaican cases we study can be compared with neighbours.

What that means for a buyer is practical. We run fraud detection, attention screening and, where the stakes call for it, verified fieldwork on our studies, as set out in our earlier article on bots and panels. The Omnibus starts from J$50,000 with results three weeks after you submit your questions. Details are on the services page.

What to do this week

1. Ask your media or research supplier for the rule list the platform uses to remove duplicate and mismatched records, and for the count of records removed, by parish. A supplier who cannot supply both is selling a result without its working.
2. Write the decision your data feeds, in two sentences with no mention of AI, and send it to the supplier before you buy anything extra.
3. Ask when the three-month interim read arrives, what it covers, what the subscription costs, the date of the full release and how influencers are defined. The Gleaner's report gives none of these.
4. Ask which weeks a six-month dataset covers, and whether you can filter by month. Do this before you compare any seasonal result with last year's.
5. Ask how the survey recruits respondents and how it screens for bots. Pew found trap questions largely removed yes-saying (1 percent of adults afterwards), yet Pew still calls a surefire fix elusive, so ask what else the supplier checks, such as open-ended answers.
6. Ask what the platform does with records it flags: removed, or reviewed by a person first. Pew found that matching an opt-in sample to voter files slightly increased error by removing mostly good respondents.
7. Test it: take five questions you asked of your last audience report, time how long each took to answer from the PDF, and set a target of under 10 minutes in total for the same five on any online platform you trial. Record both times by 19 October 2026 and hold the supplier to the second number.

Where this can go wrong

Automated cleaning can delete real people. A rule that treats two interviews from one address or one phone number as duplicates may remove households where several adults share a device, which in lower-income and rural areas can be the norm (this is our inference, not a finding in the Gleaner report or Pew). The groups most likely to vanish are the ones the new banked and unbanked and farm-owner segments are meant to count, and advertisers who buy on those numbers bear the cost in wasted budget. A longer window adds a second risk, because seasonal swings can look like trends. The guardrail is documentation: published removal rules, counts of removed records by parish and segment, and a validation sample checked by a person. A platform that replaces PDFs also makes it easier to run a filter that no one has examined, so ask who signs off each release before the data goes live. Pew's voter-file result shows this is not hypothetical: matching to voter files slightly increased error by removing mostly good respondents. Dr Koka's own framing is a fair test for the supplier: "It is not about replacing human intelligence. It is to give added advantage to human intelligence." A buyer can ask where the human sits in the pipeline.

Questions readers ask

What did the Gleaner leave unanswered about the All Media Survey?

The Gleaner's report of 4 October 2026 does not give the subscription cost, the date of the full results release, or how influencers are measured and ranked. It does say subscribers receive an interim read on radio, television, newspapers, internet and social media after three months, so a buyer can ask for that date in writing.

Can cleaning survey data remove real people?

Yes. Pew Research Center found that matching an opt-in sample to voter files slightly increased error by removing mostly good respondents, such as those who simply declined to give a name or address. Pew also says purging bogus cases lowers error on most metrics, but a surefire solution remains elusive.

How many bad cases failed more than one check in Pew's study?

Of 5,369 cases that failed at least one prescreening check, 1,879 (35 percent of failed cases) failed several. A problematic open-ended answer was the most frequently failed check, present in 70 percent of failed cases, so reading free-text answers catches more than any single trap question.

Did trap questions fix the yes-saying problem in Pew's data?

Largely. After screening with trap questions, 1 percent of adults, and no more than 2 percent of any demographic subgroup, answered yes to 10 or more of the 15 yes/no questions, against 7 percent of all adults before screening. Pew still describes a full fix as elusive.

Why do two sources give different internet penetration figures for Jamaica?

DataReportal's Digital 2026: Jamaica report puts penetration at 89.5 percent in October 2025. The 85 percent figure comes from a statement attributed in the Office of the Prime Minister's April 2026 release to Minister Andrew Wheatley, and the release cites no source for it. A buyer should ask which figure a survey's weighting uses.

The loose end

We have not seen MRSL's methodology document, only press coverage of the launch, so we cannot say how the survey recruits respondents, what its AI removes or how it defines an influencer. Everything above about risk is a list of questions, and the answers may be good. The weakest part of our argument is the Pew evidence: it measures American online polling, and no Jamaican equivalent has been published.