Editorial policy and independence

Admission rules for model-behavior claims: source quality, numerical provenance, AI-assisted drafting, specialist-review boundaries, and stale-page records.

RankRift names no human or specialist reviewer and publishes no measured study of its own. The rules below are the standard every page here is held to, and the one to apply to any source you cite.

Read the measurement methodology · Read the disclosure

In short

This page defines a reader-checkable evidence bar. A behavior claim must cite dated provider documentation or a reproducible run; otherwise it remains a labelled hypothesis. A measured figure needs its prompt set, model identifiers, run dates, sample size, variance, and raw evidence. A review claim needs the reviewer's identity, scope, date, and disposition.

Who this is for

Practitioners deciding whether a number from this site is safe to put in front of their leadership
Journalists and researchers who need to know what a RankRift claim is actually backed by before citing it
Vendors and brands who want to see the standard applied before their product or category is discussed
Readers who suspect AI-assisted content and want an explicit statement of how it is used and reviewed here
Technical readers checking whether a site arguing for retrievability actually allows the retrievers in

When to use this

Before quoting a RankRift figure in a report, a pitch, or an article
When you want to know whether a specific sentence on this site is measured, documented, or inferred
When a page here looks out of date and you want to know what our refresh obligation is
When you believe something on this site is wrong and want to know what standard we will be held to

Accuracy: what we are willing to state as fact

Every claim about model behavior falls into one of three buckets: provider-documented with link and access date; observed in a reproducible run with its method; or uncertain and labelled accordingly.

That last option is the one most of this category skips. "ChatGPT does X" is almost never a defensible sentence. "As measured on these dates against these model versions, X occurred in this share of runs" is. The second is longer and duller, and it is the one that survives a model update.

AI answers vary by model, phrasing, region, personalization, and date. Anything written here that ignores that is a defect.

Provider-documented behavior is linked and dated; if the documentation changes, the claim is re-checked, not left standing
Measured behavior is attributed to the named run, with the method public and raw data available
Everything else is labelled as inference, hypothesis, or open question — including on pages where a confident answer would read better
No invented statistics: no adoption percentages, market sizes, traffic shares, or benchmark numbers we cannot reproduce
Illustrative examples are labelled illustrative, every time, and are never dressed up as a customer or a case study

Sourcing: what counts as evidence

Primary sources come first: provider documentation and changelogs, reproducible published runs, and peer-reviewed or preprint research. A reader must be able to inspect the cited source directly.

We do not treat a vendor's marketing chart as evidence. This is the most common failure in AI-visibility writing: a striking percentage from a platform's launch post gets quoted, then re-quoted, until it circulates as a fact with no method behind it. If we reference vendor-published material at all, it is labelled as vendor-published and unverified, and we do not build an argument on top of it.

The same rule applies to competitor facts. We do not assert another product's features, model coverage, or accuracy unless we have verified them hands-on, and we describe pricing as a model with a link to the vendor rather than restating an amount that may already be stale.

Preferred: provider docs and changelogs, reproducible runs, and academic retrieval or generative-search research
Labelled and never load-bearing: vendor marketing material, unsourced industry statistics, secondhand summaries of studies we cannot read
Never: screenshots implying a study that was not run, or another organisation's chart presented as our testing
Competitor coverage is editorial and unaffiliated, with the disclosure stated on the page

Reproducibility: no number without its provenance

This is the non-negotiable rule: a published measurement carries its prompt set, model identifiers, run dates, sample size, variance, and raw evidence. Without them, omit the result rather than presenting an empty or implied study.

Label an under-powered result as such. Do not round a noisy result into a clean claim.

A method change requires a version bump and changelog entry linked from every affected result.

Required with every figure: prompt set reference, model versions, run dates, sample size, variance
Raw responses must remain available under the study's stated access and retention rules
A conforming research dataset uses an open format, explicit licence, and citation string
If a figure cannot be sourced, restructure the page around what can be defended or omit the figure
Measurement uses official model APIs only — never scraped interfaces, rate-limit evasion, or automated consumer accounts

AI-assisted writing and review records

AI assistance may be used for drafting, structuring, and editing. It cannot supply a fact, source, result, credential, or review record. Every claim must be checkable against a cited source, every link must resolve, and every number must be recomputable. RankRift names no human or specialist reviewer, so no page here presents a review record.

There are hard limits on what generation may produce. No invented examples presented as real. No fabricated case studies, customer stories, testimonials, or client names. No invented statistics, sample sizes, or study results. No invented credentials, awards, or counts of any kind. Where an example helps, it is labelled illustrative and constructed to be obviously so.

Specialist review is relevant only when a page makes a claim requiring that expertise. Record the specialist's identity, relevant scope, date, and requested changes; otherwise keep statistical, legal, analytics, privacy, and ethics material within the general source-checkable boundaries below.

Statistics boundary — show assumptions and formulas; obtain qualified review before asserting a novel statistical conclusion
Legal boundary — keep risk material general, cite the responsible official source, and direct case-specific questions to qualified counsel
Analytics boundary — distinguish observation from causation and state confounders
Privacy boundary — describe only documented processing and identify the responsible operator
Ethics boundary — do not turn community participation or reviews into manipulation instructions
Brand-naming boundary — publish the prompt set, conditions, uncertainty, and dispute route with any study naming brands

Bylines are either real people with verifiable relevant experience, or a clearly identified house byline that does not imply credentials nobody has.

The claims we never make

Some sentences are barred here regardless of how well they would perform. They are listed rather than implied, so that a reader who finds one can hold us to this page.

Any guarantee of placement, recommendation, ranking, traffic, or revenue, and any implication that we can control what a model says
Any claim that RankRift is the best or the leading option in its own comparison
Fake maturity — invented customers, testimonials, client logos, awards, team credentials, or user counts
Manipulation instructions. Astroturfing, review fraud, parasite spam, and prompt injection are covered defensively — what they are, how they get detected, what they cost — never how to run them
"Search is dead" framing, and its mirror, "AI answers don't matter." Both are positions taken for attention rather than accuracy
A measured figure presented as a quality judgement. Naming a brand in research is a factual statement about model output, nothing more

Dates, re-verification, and stale pages

Model behavior changes without warning, and a technically accurate page can become misleading with no edit at all. Dating is therefore part of the claim, not metadata.

A page that publishes a measured figure or a volatile model or tool fact needs a last-verified date and a recheck interval chosen for that source.

A page that fails re-verification is corrected, or visibly marked as stale with what we no longer stand behind. It is never quietly left standing while a date implies someone checked.

A last-verified date and recheck interval on any measured figure or volatile model or tool fact
Refresh intervals follow source volatility, not a fixed calendar
A reference row we cannot verify is marked unverified rather than guessed at
Corrections are appended in place with a date and a description of what changed

AI crawler access

A prose statement about crawler access is not evidence of it, so this page states rules rather than a configuration. A crawler policy should document each allowed or blocked crawler, its verified identity, purpose, control mechanism, and last-checked date, and the deployed configuration and server logs must agree with it.

Private user output, prompt sets, results, and cost-incurring endpoints require separate access decisions and must not inherit a public-content default.

A machine-readable summary such as llms.txt can be evaluated as an experiment only when its exact file, date, intended consumer, and observation method are linked.

Whether you should allow AI crawlers is a different question with a different answer for every site. The readiness pillar has the decision framework rather than a slogan.

How corrections work

A correction is appended to the affected page with a date and a description of what changed, and material corrections go into an append-only log. The correction protocol defines the required fields.

A disputed measured result should be re-run under the published method, with the outcome recorded whether it changes or confirms the original. The corrections page defines the full process.

Frequently asked questions

Is this site written by AI?

AI assistance may be used for drafting and structuring, but it cannot create facts, results, sources, credentials, or review records. RankRift names no human reviewer, so no page here claims a human review. Constructed examples remain explicitly illustrative.

How can I tell whether a claim here is measured or just inferred?

By how it is written. A measured claim names the models and versions, run dates, sample size, variance, raw evidence, and method. A documented claim links to dated provider documentation. Anything else is phrased as inference or an open question. A confident sentence without that support fails this specification.

Why won't you cite a vendor's published statistics?

Because in this category those figures usually arrive without a prompt set, a sample size, or a model version, which makes them unauditable. Quoting them would launder a marketing claim into a fact, and the whole argument of this site is that unauditable numbers are the problem. If vendor material is referenced at all, it is labelled vendor-published and unverified, and no conclusion here rests on it. The same standard applies to our own figures.