What problem does it solve?
Readers increasingly ask AI chatbots and search engines for news instead of visiting a news site. The answers are built from publishers' reporting, but the publisher loses the visit, the subscriber relationship and the chance to show its depth. Archives are often hard to use even for the publisher: TIME told Digital Content Next that its content sat in five different databases going back to the 1920s, some of it only as PDFs of magazines. Paying readers can struggle too: the Financial Times found that FT Professional subscribers were looking for specific sources or data points and struggled to find them quickly.
General AI assistants also get the news wrong. The BBC found significant issues in more than half of the answers four assistants gave to questions about the news, and Apple suspended its AI summaries of news notifications after repeated mistakes. Publishers that answer with AI take on the same risk under their own brand, so the design has to protect accuracy and attribution first.
- The Reuters Institute Digital News Report 2026 finds that weekly use of AI chatbots such as ChatGPT, Perplexity and Google Gemini for news rose year on year from 7% to 10%.Overview and key findings of the 2026 Digital News Report (2026)
- BBC research found that 51% of all AI answers to questions about the news were judged to have significant issues of some form.Groundbreaking BBC research shows issues with over half the answers from Artificial Intelligence (AI) assistants (2025)
How does it work?
- Index the journalism. Published articles, and where rights allow the digitized archive, are indexed with their dates, bylines, sections and corrections. New articles are indexed as they publish.
- Retrieve and rank. A reader's question is matched against the index with semantic and keyword search, favoring recent and relevant reporting.
- Answer only above a threshold. If no article scores above a relevance threshold, the tool says it cannot answer rather than filling the gap from general knowledge.
- Answer with sources. The answer summarizes the retrieved reporting and links every point to the articles it came from, with dates, so readers can check and read on. Some tools also translate answers or read them aloud, as TIME's agent does in 13 languages, which opens the reporting to readers who prefer another language or audio.
- Label and learn. The tool is labeled as AI generated and experimental where needed, readers can flag bad answers, and editors review flagged answers and the most asked questions.
- Audience
- Customer facing
- Autonomy
- Autonomous
- Adoption
- Early adopters
- Channels
- Web chat, Mobile app
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Customer experience, Revenue growth, Inclusion and access.
Indicative value
A subscription news publisher with 100,000 paying subscribers
USD 5000 to USD 135,000
Subscription revenue retained per year
How this is calculated
Formula: subscribers * usageShare * churnPointsAvoided * subscriptionValue. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Paying subscribers subscribers, subscribers | 100,000 | 100,000 | The reference publisher. |
| Share of subscribers who use the answer engine regularly usageShare, fraction of subscribers | 0.05 | 0.15 | Editorial assumption, replace with your own usage data. |
| Reduction in annual churn among regular users churnPointsAvoided, fraction of users (percentage points as a fraction) | 0.01 | 0.03 | Editorial assumption. The Financial Times says Ask FT supports retention of key accounts but publishes no figure, and the return rate Digital Content Next reports for users of TIME's agent is not a controlled comparison. |
| Annual revenue per subscriber subscriptionValue, USD per subscriber per year | 100 | 300 | Editorial assumption, replace with your own average revenue per subscriber. |
What it leaves out: A retention effect only, and hard to separate from the fact that engaged readers are the ones who use the tool. It leaves out model and search costs, which rise with usage, editorial time for review, any conversion or advertising value, and the reputational cost of a wrong answer.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
TIME
United States · Media and entertainment · 2025
TIME first tested a conversational AI toolbar on its Person of the Year coverage in December 2024, and chose not to let that version use content from outside TIME. In November 2025 it launched the TIME AI Agent, built with Scale AI, which lets readers ask questions, get summaries in text or audio, translate into 13 languages, and search more than a century of TIME reporting with semantic and hybrid search. TIME says attribution and citation are preserved in every interaction and that the system was red team tested. Digital Content Next reports that the agent is now trained on more than three quarters of a million pieces of TIME content, some digitized from PDFs of old magazines, that new articles are indexed in near real time, and that users who engage with the agent are more likely to return and spend more time on the site than those who do not.
No outcome disclosed.
Financial Times
United Kingdom · Media and entertainment · 2024
Ask FT is the Financial Times' first customer facing generative AI product: a search tool that answers questions using FT content only, with references to the articles used. It was tested internally by editorial and product teams, piloted with a few hundred FT Professional subscribers, then offered to larger client accounts, and became available to all FT Professional customers in April 2025. Users, mostly in finance, consulting and law, use it to prepare meetings and reports and tend to open the cited sources. The FT says the links to full articles encourage deeper reading that supports renewal of key accounts, and that the tool calls out when it lacks sufficient information for a credible answer.
No outcome disclosed.
The Washington Post
United States · Media and entertainment · 2024
In November 2024 The Washington Post launched "Ask The Post AI", an experimental generative AI tool that answers readers' questions with summary answers and curated results drawn from articles its newsroom has published since 2016, ranked by relevance. To protect the integrity of the reporting, the tool does not serve an answer when it does not readily find a relevant article above a set threshold, even if one exists. It followed earlier experiments such as Climate Answers and AI article takeaways. In July 2026 Arc XP, the Post's technology business, launched a comparable answer layer, Ask The News, for other publishers.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Structured article data with dates, authors, sections, updates and corrections
- Clear rights to use archive, wire and syndicated content in AI answers
- An editorial policy for AI generated answers and how corrections apply to them
Systems to integrate
- Content management system and publishing pipeline for near real time indexing
- Subscription and paywall system, to meter or gate answers
- Analytics to measure return visits, reading and retention of users
Complexity: Medium
Retrieval over clean article data is well understood. The hard parts are archive digitization and rights, a relevance threshold that refuses often enough, source linking, cost control as usage grows, and editorial ownership of a product that speaks in the newsroom's name.
- 1
Start where the reporting is deepest
Launch on a bounded topic or franchise with a rich archive, as TIME did with its Person of the Year coverage before opening its wider archive. The Washington Post also ran smaller AI experiments, including Climate Answers, before Ask The Post AI.
- 2
Set the refusal threshold first
Decide the relevance score below which the tool declines to answer, and test it on questions the publication has not covered. Refusing is better than answering from outside the reporting.
- 3
Link every answer to articles
Show the source articles with dates next to each answer so readers can verify and read on, which also drives article consumption.
- 4
Put editors in the loop
Give editors a daily view of flagged answers and top questions, a way to correct or block answers, and a rule for how corrections to articles flow into answers.
- 5
Roll out in phases
Test internally, then with a small group of subscribers, then widely, as the Financial Times did, measuring answer quality, repeat use and cost per answer at each step.
Guardrails
- Answers only from the publisher's own indexed journalism, never from general model knowledge
- A relevance threshold below which the tool declines to answer
- Source links with dates for every answer
- Clear labeling as AI generated, with a way for readers to report errors
- Handling of sensitive topics (elections, health, breaking news) with stricter thresholds or editor written answers
KPIs to instrument
- Share of questions answered versus declined
- Accuracy of answers on a weekly editor reviewed sample
- Click through from answers to articles
- Return visits and retention of users versus comparable non users
- Cost per answer
Human in the loop
Editors own the product's scope and policy, review flagged and high traffic answers, correct or block answers that fall short of editorial standards, and decide which topics are excluded, especially fast moving breaking news.
Common failure modes
- Confident answers from thin coverage
- The tool stretches a few articles into an answer the reporting does not support. Tune the threshold and review declined and borderline questions.
- Stale or corrected facts
- Old articles contradict newer reporting or corrections. Rank by date, index corrections and show publication dates.
- Costs that scale with success
- Model costs rise with every answer. Cache frequent answers and balance speed, accuracy and cost, as the Financial Times notes.
- Intrusive design
- An AI box that crowds out the journalism annoys readers. Keep it optional and secondary to the article.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(1) requires that readers know they are interacting with AI. Article 50(4) requires deployers to disclose AI generated text published to inform the public on matters of public interest, unless the content has undergone human review or editorial control and a person holds editorial responsibility. Whether answers generated on demand for a single reader count as text published to inform the public is open to interpretation, but they are rarely reviewed before readers see them, so the conservative choice is to label them.
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Requires disclosure of AI generated text published to inform the public on matters of public interest, with an exception for content under human editorial control.
Controls to put in place
- AI labeling on the answer box and on every answer
- Editorial policy for AI answers, owned by a named editor
- Rights register for the content the tool may use
- Weekly accuracy review on a sample, with results reported to the editor
- Logging of questions, retrieved sources and answers for corrections and complaints
When it went wrong elsewhere
- Apple Intelligence: iPhone AI news alerts halted after errors. Apple suspended AI generated summaries of news notifications, which appeared to come from within news organizations' apps, after repeated mistakes summarizing headlines. Not a publisher's own tool, but the failure mode an archive answer engine must prevent.
Frequently asked questions
- Which publishers run their own AI answer engines?
- The Washington Post launched Ask The Post AI in November 2024 over its reporting since 2016. The Financial Times made Ask FT available to all FT Professional customers in April 2025, and TIME has shown its AI agent on almost every piece of content since November 2025.
- How do publishers stop the AI from making things up?
- By answering only from their own reporting and refusing below a relevance threshold. The Washington Post does not serve an answer when the tool does not readily find a relevant article, and Ask FT is built to call out when it lacks sufficient information.
- Do AI answer engines help publishers commercially?
- Early signs point to engagement and retention rather than direct revenue. The Financial Times says Ask FT drives deeper reading that supports renewal of key accounts, and Digital Content Next reports that users of TIME's agent are more likely to return and spend more time on the site, although that compares users with non users rather than a controlled test.
How to cite this page
Blits.ai AI Use Case Library, "AI answer engine for readers built on a publisher's own journalism", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/publisher-archive-answer-engine. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published