How AIOTruth decides

1. How candidates are discovered

Three surfaces run daily. Reddit and YouTube are where a development is usually discussed first, and the platforms' own feeds are where it is actually documented.

• Reddit. Public feeds from 13 communities, batched into 4 requests and paced inside Reddit's rate limit. The feeds carry titles, bodies, and timestamps. They do not carry vote counts, so AIOTruth has no engagement signal from Reddit and does not pretend to.
• YouTube. 6 standing searches over the last 7 days, enriched with real view and comment counts.
• Official sources. 12 platform blogs, changelogs, standards repositories, and research feeds.

Reddit engagement and a YouTube demonstration are discovery signals. Neither is proof that a claim is correct.

2. What is in scope

A candidate belongs on AIOTruth only when it materially relates to one of these questions:

1. Does this change how brands are found through AI?
2. Does this change how AI systems understand an entity?
3. Does this affect citation, recommendation, or representation?
4. Does this help businesses publish clearer evidence?
5. Does this improve measurement of AI visibility?
6. Does this reveal meaningful model or platform behavior?
7. Does this provide a repeatable AIO workflow?
8. Does this challenge a common AIO claim with evidence?

Everything is sorted into exactly three categories, and an item that fits none of them is rejected rather than forced into one.

• Tools & Apps. Software a practitioner can actually open: AI visibility tools, citation trackers, entity analysers, structured data tools, extensions, agents, and workflow applications.
• Models & Platforms. The systems that decide what gets found: ChatGPT, Claude, Gemini, Perplexity, AI Overviews, Copilot, browser agents, answer engines, model updates, platform documentation, and agent standards.
• Content & Methods. How the work gets done: experiments, tests, reproducible workflows, case studies, prompts, structured content approaches, citation strategies, and entity-building methods.

3. How sources are verified

Each candidate's links are fetched and read. A page earns a place in the record only if it was actually reached and it actually discusses the item: a page that returns 200 but never mentions the subject is not evidence, and treating it as evidence is the easiest way for an automated publisher to become dishonest.

Reached sources are ranked on a published hierarchy. The strongest tier reached drives the evidence rating.

An unreachable link is recorded as unreachable rather than deleted. When a source later disappears, the AIOTruth summary stays, the source is marked unavailable, and the remaining sources are re-scored.

4. How evidence status is assigned

Every published item carries exactly one evidence label. The label describes the strength of the support, and it is separate from the score on purpose: a well-evidenced item can still be unimportant, and an important item can still be thinly evidenced.

An Unverified item can never be Best Today, can never be Best This Week, and never displays a Value Score. It may appear only inside the clearly labeled Claims Under Review section.

5. How the Value Score is calculated

The AIOTruth Value Score answers one question: does this deserve attention. It does not score truth.

It is deterministic. Every point comes from a countable feature of the candidate: matched beat terms, matched scope questions, reachable artifacts such as documentation or a repository, verified source tiers, the number of independent domains, measured figures in the material, and similarity against the existing archive. No model opinion enters the number. Run the same item through twice and it scores the same twice.

Each dimension scores 1 to 10, then:

Value Score = relevance x 0.3 + usefulness x 0.3 + evidence x 0.25 + originality x 0.15

rounded to one decimal place. Every item page shows the working that produced its own four numbers.

The thresholds

A threshold is never lowered to produce output. A week where nothing clears 7.5 publishes nothing, and the roundup says so. Cadence is a ceiling, not a quota.

6. One development, one item

Several threads and videos about the same release collapse into a single item that cites the primary source and keeps the rest as references. Matching runs on canonical URL, shared primary source, product or model name within the same week, title similarity, and near-identical body text, which is the syndicated press release tell.

7. How limitations are disclosed

Every item page carries a Limitations block listing what could not be verified: unreachable sources, single-domain sourcing, the absence of official documentation, tests AIOTruth did not reproduce, and any commercial relationship. The block is generated from the verification record, so it cannot be quietly omitted.

8. How corrections are handled

AIOTruth corrects material errors openly. A correction records the original statement, the corrected statement, the date, the reason, and the supporting source, and the affected page links to it. A material factual conclusion is never silently changed. See Corrections.

9. How commercial conflicts are disclosed

AIOTruth is owned by Digilu, which also owns other properties in this field and works with clients. When an item involves a Digilu-owned product or a Digilu client, three things happen automatically: the relationship is recorded, the disclosure appears on the item page, and the item is held for human review instead of publishing itself. A Digilu product is never ranked Best of the Week automatically. The rubric applied is the same one on this page.

10. How AI is used, and what needs a human

Automation assists discovery, organization, summarization, and scoring. Publication requires the evidence and quality gates defined on this page.

Concretely:

• Deterministic code does the discovery, the fetching, the source tiering, the evidence label, the score, the duplicate collapsing, and the selection.
• A language model writes only four prose blocks on an item page: the summary, why it matters, the judgment, and the action. Those blocks pass a gate that rejects any number not present in the checked material, any hype phrasing, and any claim that AIOTruth tested something it did not. Copy failing the gate is replaced with plain text built from the verified facts.
• A human must approve anything involving accusations, fraud or security allegations, legal disputes, privacy concerns, a disputed claim carrying reputational risk, a claim resting only on anonymous posts, or any Digilu-owned product or paying client.

11. What cannot be verified

AIOTruth does not independently reproduce most demonstrations, and it says so on the pages where that matters. It cannot verify private results, unpublished data, or anything behind a paywall. It cannot confirm a causal claim about why an AI system cited one source over another: no one outside those companies can, and any publication claiming otherwise is guessing. Where a claim rests on a single account, the item says that in its limitations rather than rounding it up to a fact.

12. Source material

AIOTruth summarizes in its own words and links to the original. It does not republish articles, reproduce paywalled content, or rehost video. YouTube material is embedded from YouTube and credited to the channel. Reddit threads are linked, not reproduced, and an anonymous comment is never presented as established fact.