USA TODAY
genAI answer engine embedded in editorial content, monetized at the session level
USA TODAY rolled out DeeperDive in September 2025, a Taboola genAI answer engine grounded on its own content, and announced in the second quarter of 2026 more than 50 million questions asked and 390,000 average daily interactions, with an effect described without figures on time on site, advertising revenue per session and subscription intent.
Key points
- USA TODAY rolled out DeeperDive, Taboola's genAI answer engine, across its whole site in September 2025.
- The engine answers from content produced by USA TODAY and the USA TODAY Network.
- More than 50 million questions asked since launch, 390,000 interactions a day on average.
- The business effect is announced as a direction only: no percentage published on time on site or revenue per session.
Objective
Search traffic is receding and monthly unique visitors are falling. The goal is no longer to win page views but to hold the reader who already arrived for longer, and to draw more revenue from each session: advertising served during the conversation, then conversion to a subscription.
The deployment
DeeperDive is a generative answer engine built by Taboola and installed directly in USA TODAY pages. The reader asks a question on the newspaper's own site instead of going to ask an outside engine. The answer is composed from content produced by USA TODAY journalists and the USA TODAY Network, with the relevant articles attached around it. A beta ran on part of the readership starting in June 2025, then the system was rolled out to the whole audience on September 15, 2025, on a stated base of more than 195 million monthly unique visitors. Gannett presented itself at the time as the first US publisher to embed this technology. The monetization logic fits in one sentence from the group: keep the conversation on your own inventory. A session spent in DeeperDive carries advertising served by the publisher, and the launch release announced higher ad CPMs, curated ads inside the experience, then tests of LLM ad formats with brands starting in the fourth quarter. The announced next stage of the partnership plans for agents able to offer the reader a purchase. Ten months later, on the second quarter 2026 earnings call (August 6, 2026), Michael Reed gave the volumes: more than 50 million questions asked since launch, and average daily activity above 390,000 interactions. On the business effect, he compared pages equipped with DeeperDive to pages without, and described it as a direction, not in points: more time on site, more advertising revenue per session, stronger subscription intent. No figure was published on those three effects. The group says it wants to extend usage beyond USA TODAY, to the whole network. The same call announced a separate layer: a partnership with Palantir to connect audience behavior, content engagement and first-party data, and to automate the matching of affiliate offers to content, where the work used to be manual. This component is presented in the future tense by management, it has produced no published result at this stage and should not be read as an achievement of DeeperDive.
Results Proof A
The figures come from the Q2 2026 earnings call held by the CEO and are corroborated by the official launch release and by trade press. Caveat to state upfront: the only hard figures are usage volumes (50 million questions, 390,000 interactions a day). The business effect is stated as a direction, without any percentage, and the group published neither a measured gap nor the protocol of the with/without page test. The quarter's financial results are the group's, they do not measure DeeperDive.
How it works
Inferred typical approachThe internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.
The stack in detail
- plateforme DeeperDive Taboola's genAI answer engine, integrated into USA TODAY pages and grounded on the editorial corpus of the USA TODAY Network. The underlying language model is not disclosed publicly.
- plateforme Palantir Partnership announced on August 6, 2026 on audience and first-party data. The exact product is not named publicly, and no result is published.
How it runs, concretely
For ops teams-
1The reader asks a question inside the page Reader
The entry point is built into USA TODAY pages rather than sitting on a dedicated page, so it captures the reader who already arrived through an article.
-
2The engine composes an answer from in-house content AI (Taboola DeeperDive)
The answer is generated from content produced by USA TODAY and the USA TODAY Network, with the relevant articles attached around it to give context.
-
3The session is monetized during the conversation Publisher ad sales
Advertising served inside the answer experience, with the stated goal of higher CPMs and curated ads, then a test of LLM ad formats with advertisers.
-
4The engaged reader is pushed toward subscription Subscription team
Longer engagement serves as the lead-in to paid conversion, subscription intent being one of the three effects management tracks.
-
5Management decides on a with / without comparison Management, data team
Steering rests on comparing equipped pages with unequipped pages, on time on site, advertising revenue per session and subscription intent.
The indexed editorial corpus and its freshness. If the day's articles are not available to the engine, the answer loses its edge over a general-purpose assistant and the reader goes elsewhere with the question. The second signal is the page with / page without measurement, without which the effect on revenue per session stays an impression.
How your customers perceive this type of use
Sourced studiesLes consommateurs n'acceptent pas les chatbots par defaut : 64% prefereraient que les entreprises n'utilisent pas d'IA dans leur service client (Gartner, 2024) et pres d'un utilisateur sur cinq du service client par IA n'en retire aucun benefice (Qualtrics, 2025). L'acceptation se construit sur trois conditions mesurees par Salesforce : savoir qu'on parle a une IA, pouvoir escalader vers un humain, comprendre la logique de l'agent.
Acceptance conditions
- Etre informe qu'on parle a une IA et non a un humain (pres de 75% le demandent, Salesforce 2024)
- Un chemin d'escalade clair vers un agent humain (45% plus enclins a utiliser l'agent IA, Salesforce 2024)
- Une logique de l'agent clairement expliquee (44% plus enclins, Salesforce 2024)
Red lines
- Rendre l'humain injoignable : c'est la premiere inquietude des consommateurs sur l'IA dans le service client (Gartner 2024) et 50% craignent que l'IA les coupe du contact humain (Qualtrics 2025)
- Remplacer le service client par l'IA sans alternative : 53% envisageraient de partir chez un concurrent (Gartner 2024)
Sources: Salesforce 2024 · Gartner 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- An editorial corpus owned outright, deep enough and fresh enough to answer better than a general-purpose assistant on the topics where you are legitimate
- Content indexing that keeps up with the newsroom's publishing pace, including same-day content
- Measurement at the session level, not the page view level: time on site, advertising revenue and subscription event tied to the same session
- The ability to serve equipped pages and unequipped pages in parallel, to hold an honest comparison
Org prerequisites
- A clear agreement between newsroom and product team on what the engine is allowed to rephrase, and on how the AI answer is flagged to the reader
- An ad sales operation able to sell the inventory of the conversational session, otherwise the extra engagement does not turn into revenue
- An accepted trade-off between keeping the reader inside the answer and sending them back to the articles
Possible stack
- An off-the-shelf publisher answer engine (Taboola DeeperDive or equivalent)
- Or an in-house build: search over the editorial corpus, answer generated by an LLM, systematic citation of the source articles
- Existing ad server, extended to the answer format
- Paywall and subscription offer tooling already in place
The plan, step by step
- Step 1Agree with the newsroom on what the engine can answer, on which topics, and how the answer is presented to the readerDeliverable: Written editorial rule, binding on the product
- Step 2Make the corpus queryable and keep it up to date at the pace of publicationDeliverable: Index of in-house content, fed continuously
- Step 3Open the engine to a fraction of the audience and keep a group of unequipped pagesDeliverable: Measurable beta with a point of comparison
- Step 4Instrument the session: time on site, impressions and advertising revenue, subscription eventDeliverable: Single session view that brings engagement and revenue together
- Step 5Sell the inventory of the conversation, including the formats specific to the answerDeliverable: Commercial offer for that placement
- Step 6Compare equipped pages and unequipped pages before rolling out, then extend to the group's other titlesDeliverable: A measured, quantified gap that decides the rollout
First step: Take the ten questions readers most often ask an outside assistant on your topics, check whether your own archive answers them better, and go further only if the answer is yes.
The same pattern, proven in other industries
Cars.com
Marketplace AI building blocks monetized on both sides: conversational search on the buyer side for conversion, and generation and AI tools packaged and sold to subscribing dealers.
Grab
AI assistant for merchant advice and activation
Ping An
AI conversational agent for service and sales support on the customer base
Sources
- S1 USA TODAY (TDAY) Q2 2026 Earnings Call Transcript Primary archive pending
- S2 USA TODAY Deploys Taboola's DeeperDive AI Answer Engine for all Audiences Primary archive pending
- S3 USA TODAY Co. loses 22 million monthly unique visitors in one quarter Secondary archive pending
- S4 USA Today Co. announces partnership with data analytics giant Palantir Established press archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.