otb research
Capabilities statement · 2026

Research that teams trust enough to act on.

Oliver Barker · Lead User Experience Researcher · London

AI product research · Trust and safety · Integrity · Mixed methods · Research strategy

OTB Research is the independent practice of Oliver Barker, a lead user experience researcher with 13+ years across agency and in-house. At Meta I led research for misinformation, user reporting, fraud and scams: trust and safety and integrity work on AI products used by billions, operating under regulatory scrutiny and at global scale. I design and run mixed-methods programmes that turn ambiguous, high-stakes questions into decisions product, policy and commercial leaders can stand behind.

Start a conversation LinkedIn See the work Capabilities PDF
Where I fit

I get hired when the question is ambiguous, the stakeholders disagree, and the deadline is real.

Research delivered at
ReapitMetaDICEShpockRazor ResearchFuller Research Group
13+
years in research, agency and in-house
5
years at Meta, at global scope
5
continents where I have conducted international research
1,000+
in-depth interviews and focus groups moderated
01 · What I can help with

Six things teams bring me in to do.

AI products

AI product research and evaluation

How users understand AI features, whether they trust them, and when they decide to rely on them. Concept testing, conversational tone and guardrails, evaluation rubrics pairing blind human review with LLM-as-judge scoring, and measurement once it is in production and in front of users.

Discovery

Foundational research and strategy

Research that starts in genuine ambiguity and ends as roadmap direction. Understanding needs and behaviours, sizing the opportunity, aligning stakeholders who disagree, and turning the findings into a north star and a prioritised set of bets.

Evaluation

Product, journey and onboarding evaluation

Usability and concept testing, end-to-end journey audits, onboarding and verification flows, and funnel diagnostics that locate where users drop off and explain why. Recommendations are tied to the metric they should move.

Measurement

Survey programmes and tracking

Survey design at scale, in-product intercepts at the moment of behaviour, and longitudinal CSAT, sentiment and trust tracking built into product dashboards with data science, so the metrics can act as guardrails for ship or no-ship decisions.

Insight

Market, customer and partner insight

How customers decide, not only how they use the product. Category, brand, shopper and merchandising research, plus B2B audiences: advertisers, trade sellers, estate agents, venues, promoters and artists, each measured against the commercial goal that matters for them.

ResearchOps

Research operations and team capability

Recruitment strategy, screeners and panel management, insight repositories, and coaching teams to run good research themselves. Sample quality sets the ceiling on what a study can claim, so I treat recruitment as part of the method, not admin.

02 · Selected work

Seven programmes, filtered by what they evidence.

Every programme named here is one Meta has announced publicly, and each links to its public source. Internal business performance, commercial figures and unreleased work are deliberately left out. What is here is the reasoning: how the problem was framed, what the research found, and what changed as a result.

Seven programmes, most recent first. Use the arrows to move through them.

2026MetaLead researcherConversational AI0 to 1

Scam Protection Center

A conversational AI product for reporting scams and getting support, taken from foundational discovery to global launch in around six months.

The finding

What users wanted from reporting a scam was not reimbursement. It was acknowledgement that what happened to them was real, and confidence it would not happen again.

meta.com/en-gb/safety/scam-protection-center
The Scam Protection Centre on meta.com, showing reporting alongside recovery and prevention.
The shipped destination. Reporting sits alongside recovery and prevention rather than behind them, which is what the foundational work argued for.
Context

Meta needed a dedicated anti-scam destination spanning Facebook, Instagram and Meta.com, a category that did not yet exist for the company. The bet was one place for prevention, recovery and reporting, powered in part by conversational AI. The timeline was aggressive and the ambiguity was high.

Approach

I sequenced the research to de-risk the biggest unknowns first, starting with whether a single destination answered a real need at all, and ending with a measurement design that could tell us whether it had worked.

01Foundational discoveryIs a single destination a real need, or a company convenience?
02Name and brandComprehension, trust, category fit.
03Prototype, then in-productUsability of the AI interaction model, in prototype and then live.
04Post-launch trackingHoldout design, so it was measured against a real baseline.
Discovery to global launch, roughly six months
Impact

The research shaped the product's framing, its information hierarchy, its name, and the AI's conversational tone and guardrails. The destination is now live publicly, organised around Meta's remove, reduce, inform and improve approach.

Scam Protection Centremeta.com ↗
2024 / 25MetaLead researcherGuardrailsMulti-market

Conversational AI in reporting

How the language of an AI assistant changes what people believe happened to them.

The finding

High-empathy responses were read as the system confirming the user's suspicion. Empathy was not neutral. It worked as a truth signal.

Meta AI support assistant  /  Report a scam
The assistant acknowledging the report without confirming a verdict.
01Acknowledge, then act
The assistant explaining how the review process works.
02Explain the process
The assistant declining to imply a refund when asked directly.
03Set the expectation
The shipped assistant. It acknowledges what happened without confirming a verdict, explains how the review works, and is explicit that reporting will not get the money back. Note the third screen: asked directly about a refund, it declines to imply one, which is the same tension the Scam Protection Center research surfaced.
Context

A conversational AI layer was being built into reporting, helping people describe what happened and reach the right resolution. Its tone was a live design question with legal and regulatory exposure. Meta has since launched AI-powered support for reporting scams, impersonation and problematic content.

Approach

Moderated concept testing, iterative in-flow usability and cross-market comparison, focused on how people interpreted the assistant's language rather than on whether they could finish the task.

Before: be warm

That sounds really stressful. I am so sorry this happened to you.

Read as: the system agrees I was scammed. A verdict the assistant had no basis to give, before anything had been reviewed.

After: acknowledge, do not judge

Thanks, I have got what I need. Here is what happens next and when you will hear back.

Read as: my report is being handled. Support without a premature conclusion.

Illustrative of the tone principle the research produced, written for this portfolio rather than taken from the product.

Impact

Moved the design principle from "be warm" to "acknowledge without implying a verdict", and made tone a first-class guardrail question rather than a copy decision taken late.

Support and safety with AIabout.fb.com ↗
2024MetaLead researcherFoundationalRanking inputs

Feedback signals, a foundational reframe

Hide, Snooze, Block and Not Interested, read as intent rather than defence.

The finding

The signals were not reactive at all. People were making high-conviction statements about what they wanted to see less of, and ranking was under-using that information.

Content preferences  /  Facebook and Instagram
Facebook content preferences, under What you see and Who you see.
Facebook
Instagram content preferences, with Interested and Not interested under Your algorithm.
Instagram
Where these controls live today. Both apps present them as preferences that shape what you are shown next: Instagram groups Interested and Not interested directly under Your algorithm, Facebook under What you see and Who you see. The framing is forward-looking rather than defensive.
Context

These controls had been treated as reactive tools people use to defend themselves from unwanted content. The open question was whether they could also serve as inputs to ranking.

Approach

Deliberately foundational rather than tactical. Interviews and behavioural probes designed to interrogate the intent behind each signal, not the usability of the surfaces that expose them.

Read as defence
HideSnoozeBlockNot interested
Damage control. Measured as complaints.
Read as intent
HideSnoozeBlockNot interested
High-conviction preference. An input to what you are shown next.

Same four controls, a different question asked of them. The reframe is the deliverable.

Impact

Changed what these controls were understood to be for, and with that the brief for designing and measuring them.

User controltransparency.meta.com ↗
2023 / 24MetaLead researcherEU DSA Article 16Qual and quant

Reporting experience overhaul

A redesign under regulatory deadline, validated across seven markets.

The finding

Reporting was not a form problem, it was a comprehension problem. People failed because they did not understand what the categories meant.

Reporting an ad on Facebook  /  five steps, scroll to see them all
Entry point to the reporting flow.
01Entry point
Choosing a report category.
02Choose a category
Narrowing the category down.
03Narrow it down
The chosen categories read back in plain language before submitting.
04Confirm what you said
What happens after the report is submitted.
05See what happens next
The shipped flow. Steps four and five are where the comprehension work shows: before submitting, the categories chosen are read back in plain language, and afterwards the review is tracked through to a decision rather than ending at a thank-you.
Context

The existing reporting flow was under regulatory pressure and internal quality scrutiny, with several candidate redesigns on the table and a hard external deadline.

Approach

Directional evaluative testing to identify the strongest starting point, then a seven-market moderated validation study, escalated to unblock it inside a compressed timeline, followed by post-launch behavioural analysis with data science.

EUEUEULATAMLATAMAPACNorth America

Seven markets of moderated fieldwork inside a compressed regulatory timeline. Market-specific comprehension issues drove the fast-follow pivots.

Impact

Fast-follow pivots improved how accurately people categorised the harm they were reporting. That accuracy paid off twice: more harmful content was correctly identified and removed from the platform, and the review process itself became substantially cheaper to run, because fewer reports arrived mislabelled. The study also became the team's template for running international validation on regulatory timelines.

2022 / 24MetaLead researcherExternal oversightSensitive content

Content warning screens

Research on a live product control, presented directly to Meta's Oversight Board.

The finding

Warning screens work as decision aids rather than obstacles, provided they explain why they are there. People do not resent friction they understand.

transparency.meta.com/enforcement/taking-action
A content warning screen over sensitive content.
The control
The explanation layer behind the warning screen.
The reason behind it
Meta's public documentation of the control and the layer behind it. The screen on the left is the friction; the panel on the right is the explanation that determines whether people accept it.
Context

The interstitials that sit over sensitive, graphic or disputed content were under external scrutiny, with their value as a control being examined rather than assumed.

Approach

Multi-phase work combining moderated qualitative research using trauma-aware moderation, in-product behavioural analysis, and design-iteration testing across sensitive content categories.

Impact

I presented the research directly to Meta's Oversight Board. Formal engagement with an external oversight body at this level is rare for a researcher, and it required the evidence to hold up to scrutiny from people with no stake in the product.

Sensitive and misleading contenttransparency.meta.com ↗
2022 / 23MetaLead researcherB2BEnforcement and appeals

Advertiser enforcement and trust

Why legitimate businesses were refusing to verify, and what changed when they understood why.

The finding

Advertisers were not refusing to verify. They were refusing to verify without understanding why. Upstream context mattered more than downstream optimisation.

Context

Legitimate businesses were being caught in enforcement actions with little context on what they had done or how to recover, while verification friction compounded the problem.

Approach

I reframed the question from "optimise the flow" to "build trust before the flow", combining depth interviews with advertisers who had been penalised, funnel analysis, and pre and post experimentation.

Impact

Moved the work upstream, from optimising a verification flow to building the context businesses needed before entering it. The enforcement experience went from research submission to shipped in under six weeks.

2019 / 21DICESole researcherRetentionTwo-sided marketplace

Going out is not a solo decision

Why people with an event saved never booked it, and the social product that answered it.

The finding

The blocker was not the event and it was not the price. It was having nobody to go with. Going out is a social decision, and the product was asking people to make it alone.

DICE for iOS  /  the social layer, still live today
Finding friends already on DICE.
01Find friends
Seeing which events friends are interested in.
02See what they are into
Inviting friends to an event.
03Invite them
The shipped social layer, following the research: find who is already there, see what they are going to, then make the plan. Names are redacted; these are real contacts.
Context

DICE was growing quickly, but a meaningful share of people browsed and saved events and then never booked. The working assumptions were price, discovery and checkout friction, and each had a team behind it.

Approach

I aimed the research at the gap between intent and purchase rather than at browsing, combining depth interviews with lapsed and repeat bookers with behavioural analysis of the funnel.

The blockerInterested in the event, no one to go with, so the ticket never gets bought.
The bridgeFind friends already on DICE and see which events they are into.
The planInvite them, make it a group, transfer the ticket.

The research turned a retention problem into a product feature: make the decision social before asking for the purchase.

Impact

The finding led directly to a social layer in the app, and drove a double percentage point increase in user retention. On the supply side of the marketplace, research into the tools promoters, venues and artists use delivered a multi-million year-on-year increase in partner revenue. The feature set is still live today, covering contact sync, friends, groups and ticket transfer.

Find friends on DICEdicefm.zendesk.com ↗
03 · Methodologies

Choosing the method to fit the question.

Most stalled product decisions are not short of data: they are short of evidence the whole team trusts. My job is to produce that evidence, choosing the method to fit the question rather than the other way round.

Qualitative

Why users behave the way they do, in their own words and settings.

  • In-depth interviews, remote and in person
  • Focus groups and group moderation
  • Contextual research: shop-alongs, in-store intercepts and in-situ observation
  • Guerrilla and rapid-turnaround studies
  • Moderated usability and concept testing
  • Online communities
  • Multi-market programmes, including emerging markets
  • Research with sensitive and hard-to-reach audiences: users reporting harm or fraud, parents and teenagers, low digital confidence

Quantitative

How many, how much, and whether a change actually moved the number.

  • Survey design, scripting and analysis at scale
  • In-product intercept surveys triggered at the point of behaviour, such as exit from a flow
  • Funnel and drop-off analysis
  • Longitudinal tracking: CSAT, sentiment and trust
  • Test versus holdout comparisons and experiment readouts, with segment splits
  • SQL across large behavioural datasets; Python for analysis
  • Shared metric definitions and pipelines with data science and data engineering

Mixed methods and AI-assisted

Behavioural data for the what; qualitative depth for the why.

  • Triangulating analytics, support and feedback channels with interviews
  • Qual-led hypotheses taken into experimentation
  • AI evaluation rubrics: blind human review plus LLM-as-judge across defined quality dimensions
  • LLM-assisted analysis of thousands of open-text responses, with human review of the output
  • Hands-on with Claude, including custom Projects, Skills and MCP integrations
  • Validating outputs, whether they come from a model, a dashboard or a query
04 · Sectors and audiences

Where the work has been, and who it was with.

Sectors
AI productsSocial platformsTrust and safety Advertising and B2B SaaSProperty software and CRM Live events and ticketingOnline marketplaces MotorsRetailFMCGFood and beverageMedia
Audiences
ConsumersAdvertisers and SMBs Estate and lettings agentsTrade and professional sellers Venues, promoters and artistsShoppers Parents and teenagersUsers in vulnerable moments
05 · How I approach a brief

Four rules I run every engagement by.

Start from the decision

Before choosing a method I pin down what the team needs to decide, by when, and what evidence would change their mind.

Fit the method to the question

Qual when the problem is still ambiguous, quant when it needs sizing or proving, and both when stakeholders disagree.

Make it usable

Findings land as recommendations tied to a metric and an owner, not a deck that sits in a folder.

Say what it cannot tell you

Every readout states its limits: sample, method and confidence, so the team knows how hard to lean on it.

06 · Ways to work together

Three ways to bring me in.

Fixed-scope project

A defined question answered brief to debrief, with a clear timeline, deliverables and fee agreed up front.

Embedded contract

Day-rate research inside your product team, running studies alongside your roadmap and rituals.

Audit and set-up

A review of the evidence you already hold, then tracking, repositories or team coaching so research keeps running after I leave.

Contact

Tell me the decision you are stuck on.

Oliver Barker, Director oliverbarker93@gmail.com +44 7966 703 068 linkedin.com/in/oliver-barker-96b75574 Download the capabilities statement (PDF) OTB Research Limited · London · Outside IR35 · VAT registered Start a conversation