The Digiday Podcast

Episode: How to build an AI-generated focus group
Date: November 4, 2025
Host(s): Kamiko McCoy & Tim Peterson
Guest: Tracy Averbon, GM, The Times & The Sunday Times

Episode Overview

This episode explores how media organizations are leveraging AI to create "synthetic research panels"—essentially, AI-generated focus groups trained on real audience data. Tracy Averbon, General Manager at The Times and The Sunday Times, discusses their approach, the technology partners involved, validation methods, and key challenges in balancing innovation with editorial integrity. The episode also touches on major media business developments like YouTube vs Disney carriage disputes, potential acquisitions in the streaming industry, and the surge of short-form video as a revenue driver.

Key Discussion Points and Insights

1. Industry Climate: Publishers, AI, & Revenue Pressures

• Tim returns from the Digiday Publishing Summit Europe, highlighting pervasive publisher anxieties around AI eroding referral traffic.
• Increasing marketing orientation at publisher events, with publishers using AI-driven ad tools (Google Performance Max, Meta Advantage+).
• “It was almost like everyone kind of just needed to get that stuff off their chest… AI is eating every publisher's lunch. What do you do about it?” — Tim Peterson [02:00]
• “Publishers need to be spending on paid search ads, they need to be spending on Meta… Costs are never great for media companies.” — Tim Peterson [02:30]

2. Synthetic Research Panels: The Times’ Approach to AI Focus Groups

What is a "Synthetic Research Panel"?

• AI models trained on real audience and behavioral data to mimic responses of target audiences for feedback, focus group, and product development purposes.
• Tracy Averbon prefers the term "synthetic research" over “synthetic audiences”.
• Allows quick, scalable, and often accurate audience insights without running full human panels.

Implementation & Validation at The Times

Development Timeline & Principles

• “We’ve been working on this for about eight, nine months... It’s not something you just switch on.” — Tracy Averbon [28:16]
• Built on a rich base of 642,000 paying subscribers with deep data on behavior and preferences.
• The challenge: The Times understands subscriber needs better than prospective/non-subscriber audiences; AI helps bridge this gap.

Partner Selection: Electric Twin

• Selected Electric Twin for incorporating behavioral and social science academic research atop typical LLM and market data. [30:35]
• "The difference... was that layer of behavioral data." — Tracy Averbon [30:35]

Accuracy & Trust in Synthetic Panels

• 10% “holdout” validation method: real panel vs. synthetic responses.
• Achieved ~92% accuracy (general research typically ~93%, due to human inconsistency).
• “General research is only 93% accurate. This is 92% accurate.” — Tracy Averbon [33:08]

Case Example: Naming a Podcast

• Both real and synthetic panels validated the name “The Business” for a new podcast, which subsequently did very well at launch.
• “We gave [the synthetic panel] about four different names… ‘The Business’ came really, really high with the demographic that we were really trying to target.” — Tracy Averbon [33:57]

3. Balancing Innovation with Editorial Integrity & Audience Trust

Varied Audience Attitudes Toward AI

• Loyal subscribers: Skeptical about AI, value human curation, wary of overpersonalization.
• Prospective/"news reading" audience: Prefer AI-powered features—summaries, explainers, more video, instant gratification.
• “Our loyal readers are very skeptical about AI… The prospect audiences were very different, they wanted summaries, AI explainers, more video.” — Tracy Averbon [35:56 & 36:00]
• Highlights need for careful balance: innovating for growth while protecting trust and value for core subscribers.
• “What is the level of personalization that's not jarring, that's helpful… but in the same way maintains integrity?” — Tracy Averbon [37:25]

Editorial Concerns and Industry Risks

• AI risks: hallucinations by LLMs, potential misattribution of errors to publishers.
• “A lot of these LLMs are stealing publisher traffic… and the publisher [is] also being blamed when they get it wrong.” — Tracy Averbon [37:54]
• Newer audiences are skeptical of humans (possible editorial bias), while loyalists are skeptical of AI.
• “You have your loyal audience who's skeptical of A.I.; you have the news reading audience… skeptical of humans.” — Host 1 [39:04]

Product Implications

• Not yet implementing AI-generated article summaries, but exploring features to provide deep context and discovery for readers.
• “How do we infuse [AI] in a way that's helpful?” — Tracy Averbon [41:15]

4. Synthetic Research in Organization Workflow

Integration & Usage

• Electric Twin embedded with The Times' data and insights teams for months, assuring model quality.
• Panels for both active subscribers and the broader UK news audience.
• SaaS license model: access generally limited to Data & Insights, Marketing teams, not organization-wide to prevent misuse/overreliance.
  • “Otherwise… you can weaponize data to make a point or an argument.” — Tracy Averbon [46:19]

Cost & Accessibility

• Vendors often provide 3-month proofs of concept for free; full SaaS contracts depend on seats/users, not per-query charges.

5. Media Industry News & Short-form Video Trends

YouTube vs. Disney Carriage Dispute (06:05–14:45)

• YouTube TV muscling into pay TV, flexing power with Disney/ESPN in deal negotiations.
  • “YouTube TV will become the top pay TV provider in the US in 2026.” — Tim Peterson [06:50]
• YouTube wants shorter deal terms; Disney wants long-term security; biggest loser in the short-term: subscribers facing channel blackouts.
  • “Ultimately, like subscribers are kind of always the losers in this case because you’re still paying for YouTube TV.” — Tim Peterson [13:56]

Netflix & the Warner Bros. Discovery Bidding War (15:17–19:37)

• Reuters reported Netflix is exploring a bid for Warner Bros. Discovery’s streaming and studio assets (not cable).
• Paramount, Amazon, and Apple also rumored as interested, but differences in what each wants from the business.
• “Ultimately that's all it's going to come down to, whoever makes the biggest offer.” — Tim Peterson [18:28]

Short-form Video as a Revenue Engine (20:03–25:29)

• Meta (Reels) $50B revenue run rate (caveated as projection). [21:47]
• YouTube Shorts now generates more revenue per watch hour than long-form video—signals major shift in platform priorities and potential squeeze on long-form creators.
  • “YouTube Shorts now generates more revenue per watch hour than long form videos, which is wild to me.” — Tim Peterson [22:25]
• Implications for "middle-class" creators who focused on YouTube long-form.
• “This puts YouTube long form creators very much in that crunched position.” — Tim Peterson [24:48]

Notable Quotes & Memorable Moments

• “The Times is using fake audiences for its marketing—not quite the case, at the same time, kind of the case.” — Host 1 [27:34]
• “We think about [AI audience panels] as synthetic research and a way to get to market quicker, get an answer quicker, take the guesswork out.” — Tracy Averbon [28:16]
• “General research is only 93% accurate. This is 92% accurate.” — Tracy Averbon [33:08]
• “How do we continue to serve our loyal readers… but also bring these new audiences, build products and services for them…without alienating our existing loyal subscriber base?” — Tracy Averbon [36:58]
• “You can weaponize data to make a point or an argument.” — Tracy Averbon [46:19]
• “Now we’re at the point where we’re really focused on our business strategy and then how we can use AI tools to supercharge that.” — Tracy Averbon [46:21]

Timestamps for Key Segments

• 00:10–03:20: Publishing Summit Europe recap; AI’s impact on publishers
• 03:20–05:22: Explaining "synthetic audiences"/research panels; The Times’ use case
• 27:27–49:32: In-depth interview with Tracy Averbon on synthetic research: accuracy, implementation, use cases, internal cultural balance, and technology partners
• 06:05–14:45: YouTube vs Disney carriage dispute deep dive
• 15:17–19:37: Potential Warner Bros. Discovery acquisition—industry implications
• 20:03–25:29: Earnings: Short-form video’s rise as a major revenue driver (Meta, YouTube)

Tone & Style

The conversation is candid, journalistic, and pragmatic, blending technical explanations with practical business realities and the human impact of change. There’s an undercurrent of skepticism toward hype and a focus on empirical validation.

Summary Takeaways

• AI-generated focus groups (“synthetic research panels”) are a practical, increasingly accurate tool for publishers to expedite decision-making and understand new (and existing) audiences without exhaustive surveys.
• Validation and responsible application are essential, both to avoid over-reliance and to preserve audience trust, brand integrity, and internal decision-making rigor.
• Media business models continue to be disrupted—by platform disputes (YouTube vs. Disney), accelerating consolidation (Netflix/Warner), and the dominance of short-form video for monetization.
• Product and editorial teams must balance serving traditional, skeptical subscribers and attracting newer, tech-savvy audiences—with AI both an opportunity and a challenge on all fronts.