How to evaluate an AI news product: a five-axis framework

Why five axes

Two AI news products can look superficially similar — both have AI, both use news sources, both ship summaries — while making completely different structural choices. The choices that matter are orthogonal: each one can vary independently. Combining the five into a single "quality" score loses too much information. The five-axis framework keeps the dimensions separate so you can see what you're actually choosing between.

Axis 1: Selection model

The question: how does the product decide which content to surface to a specific user?

The three positions on this axis:

• Algorithmic personalization: the product surfaces what a model thinks the user is most likely to engage with, given the user's past behavior. Examples: Artifact, Particle's recommendations, Apple News's "For You" tab, Google News's personalized section.
• Editorial curation: humans (or human-edited algorithms) pick the day's most important stories. Examples: 1440, Morning Brew, the front page of any traditional outlet.
• User-named topics: the user explicitly declares what they care about, and the product covers those topics indefinitely. Example: Sentinel.

Why it matters:the selection model determines whose definition of "relevant" gets applied to your news consumption. Algorithmic personalization optimizes for engagement, which is uncorrelated with what you actually need to know. Editorial curation optimizes for what the editor thinks is universally important, which is uncorrelated with what you specifically need to know. User-named topics let you specify the function directly.

The 10-minute test:open the product. Try to find a specific named entity that's in the news but probably not trending — a mid-cap company, a regional regulator, a court case you know about. Can you set it as a standing topic and get future coverage on it? If yes, the product supports user-named topics. If you have to search every time, it's reactive search, not topic-driven coverage. If you have no way to specify the topic at all, it's algorithmic or editorial-only.

Axis 2: Verification depth

The question: what does the product do, structurally, to ensure that what it shows you is accurate?

The four positions on this axis, from weakest to strongest:

• None: the product surfaces or summarizes whatever sources matched a query, with no verification step. Risk: confident hallucination of false claims. Example: many early AI news products in 2023.
• Publisher-trust: the product trusts the publisher's reporting standards and surfaces their content unmodified. Risk: single-source failures propagate. Examples: most aggregators (Google News, Apple News, Yahoo News).
• Multi-source surfacing: the product shows multiple outlets reporting on the same story side-by-side, so the user can triangulate. Risk: surfaces inconsistencies but doesn't resolve them. Examples: Ground News, Particle's Multi-Source Summary, Google News's "Full Coverage."
• Cross-referenced claim verification: the product extracts the load-bearing claims from a candidate piece, checks each against multiple independent outlets, and only files what holds up. Risk: slower than the rumor cycle. Example: Sentinel.

Why it matters: verification depth is the most-undercommunicated axis in AI news marketing. Products at different verification depths look identical at first glance and produce very different output quality on contested or fast-moving stories.

The 10-minute test:find a story where there's a known disagreement between outlets — say, an earnings number that some reports got wrong before correction. Does the product show the disagreement explicitly? Does it tell you which version it's presenting? If it shows you only one version with no source-chain visibility, verification depth is shallow.

Axis 3: Persistence

The question: does the product remember your interests across sessions, or does each session start from scratch?

The three positions:

• None: every query is independent. Example: ChatGPT, Perplexity (in their default modes).
• Soft personalization: the product remembers what you've clicked on, but you don't explicitly direct it. Example: Artifact, Apple News.
• Standing assignments: the product takes explicit topic assignments and works them indefinitely. Example: Sentinel.

Why it matters:persistence determines whether the product is reactive or proactive. Without persistence, you have to remember to check on each of your topics. With it, the product remembers for you. For information needs that span weeks (a court case, a regulator's rulemaking, a competitor's product cycle), persistence is the difference between catching the news and missing it.

The 10-minute test:set up coverage on a specific topic. Close the app. Wait 24 hours. Open the app. Did the product surface anything on the topic on its own? If yes, it has standing assignments. If you have to re-query, it doesn't.

Axis 4: Output format

The question: what unit does the product hand you?

The four positions:

• Article: the product surfaces a full article from a single publisher. Example: traditional aggregators.
• Headline: the product surfaces a title plus a few lines of summary, optimized for browsing. Example: Apple News's default view.
• Summary: the product synthesizes content from multiple sources into a short AI-generated brief. Example: Particle, Perplexity Discover.
• Dispatch: the product files a state-change-level update — "X happened on topic Y, confirmed by N outlets" — with the source chain attached. Example: Sentinel.

Why it matters:the output format determines what cognitive work the user has to do per item. Articles require reading-and-extracting. Summaries require trusting the summarizer. Dispatches require reading the state change and acting on it. The right format depends on what you're doing — reading deeply for pleasure (articles), grazing (summaries), or tracking specific things to act on (dispatches).

The 10-minute test:look at one item the product surfaces. How long does it take you to extract the "what changed" from it? Five seconds is dispatch-grade. Thirty seconds is summary-grade. Two minutes is article-grade.

Axis 5: Engagement model

The question:is the product's revenue model aligned with surfacing what's relevant to you, or with maximizing your time-on-app?

The three positions:

• Engagement-optimized (ad-funded): the product makes more money the more time you spend in it. Examples: ad-funded aggregators, social-media news feeds.
• Mixed: the product has both subscription and ad revenue, with the ad layer creating an engagement incentive even when subscriptions exist. Example: most legacy outlet apps.
• Subscription-only, time-saving incentive: the product makes more money the more value you get with less time. There's no incentive to make the experience longer. Example: Sentinel.

Why it matters:the engagement model is the deepest determinant of the product's long-term behavior. Even if a product's current design is good, if its revenue depends on engagement, it will trend toward optimization for engagement over time. The only durable defense against engagement-optimization is the absence of an engagement-revenue line.

The 10-minute test:count the ads. Look at the "recommended for you" surface area on the product's main view. Engagement-optimized products will have many ads and large recommendation surfaces. Subscription-only products will have neither.

How the major products score

A condensed read of the framework against major AI news products in 2026:

How to choose, by use case

• Ambient news habit, low cost: Apple News (free tier) or Particle. Algorithmic + publisher-trust + headline + free.
• Triangulating ideological framing: Ground News. Multi-source with explicit bias annotation.
• Ad-hoc questions: ChatGPT or Perplexity. Reactive search.
• Tracking specific named entities: Sentinel. User-named topics + cross-referenced + standing + dispatch + subscription. The framework points to this as the only product that scores well across all five axes for entity-tracking use cases.

What the framework is for

Use the framework when a friend asks you which AI news app to use. Use it when you're evaluating a new product launch. Use it when you read the next round of "best AI news apps" listicles, which almost universally score on the wrong axes.

The selection model is the most-undercovered axis in the trade press. The engagement model is the most-undercovered axis in the consumer press. Both are doing more work than the others to determine what kind of reader the product will create over time. Watch them.

Related reading on Sentinel: The structural failure of the algorithmic news feed in 2026 · The future of news consumption: a 2030 forecast · The best AI news apps in 2026.