Our Editorial Process: How We Test & Score Adult AI Apps

By Alexandra Joly, Senior Editor · Last verified May 26, 2026 · See our scoring page and errata log

This page documents how reviews are produced on bestgirlfriend.ai. It is the editorial process linked from every Review, Pillar, Versus, and Listicle on the site, and the operational counterpart to the affiliate disclosure page. The disclosure explains how commercial revenue is walled off. This page explains the editorial process that produces the work behind that wall.

We're journalists, not platform operators or court-appointed auditors. We test what we can test, cite what we read, and flag what we couldn't independently verify. When we're wrong, we correct on the record on the errata board and in the update log at the bottom of every review. The audit trail of corrections is itself a trust signal. A site that never publishes errata isn't error-free; it's opaque.

Most reviewers in this space won't tell you any of this. The category pays well. They skip the inconvenient steps, fabricate the scoring page they claim to use, hide the byline behind a fake doctorate, and silently swap scores when a commission rate moves. We don't. The whole point of writing the editorial process down in public is so a reader (or a competitor, or a court-appointed auditor) can hold us to it.

How is each review made on bestgirlfriend.ai?

The 12 steps are deliberately ordered. Discovery and triage filter what we cover at all. Research and testing build the evidence base. Scoring and peer reading produce the number. Score lock fixes the result before any commercial layer touches it. Drafting and editorial review turn the locked number into prose. Publish stamps the page. The retest schedule governs when a published review is reopened.

Reverse any pair (run the catalog check before the score lock, draft before peer reading, publish before the editorial pass) and a path opens for commerce to reach the score. So we don't reverse them.

1. Discovery. A platform shows up in a search result we monitor, a competitor list, the CrakRevenue offer catalog, or a press mention. Discovery is opportunistic; coverage isn't.
2. Triage. An app is eligible for a full review if it has been live for at least six months, has a reasonable user base on the most recent public traffic estimate (we haven't tested third-party traffic numbers directly, they're estimates), is legal in our primary markets, and isn't a brand-jacker of an existing app we cover. Fail any of these and the app gets logged but not reviewed.
3. Research. We read the marketing site, the Terms of Service, the Privacy Policy, the pricing page, and any press the app has earned. We read user reports across Trustpilot, Sitejabber, the Better Business Bureau, and the relevant subreddit, with particular attention to billing-descriptor complaints. We pull the pricing page through the Wayback Machine for a 12-month price-drift trace.
4. Hands-on testing. Tests are written down so another editor running the same protocol would get the same transcript. AI companion apps get a 10-prompt conversation test, a 5-image generation test, and a voice phrase test. Cam sites get peak-time inventory observation across three timezones, a 10-room broadcast quality test, and a zero-spend hands-on walk of the pricing page and checkout flow. Adult games get a save-state lifecycle test, a billing transparency probe, and an ad-cadence observation. Real-model pages get a content-cadence audit, an engagement check, and a niche-match probe. The full protocols live on the matching scoring pages: AI companions, cam sites, adult games, real models.
5. Scoring. Each category is rated 1 to 10 against anchors published on the matching scoring page, using the test transcript as evidence. Scores are never impression-based. Every category is sourced and dated. Anything we couldn't verify (paid-tier features behind a paywall, vendor non-disclosure agreements) gets flagged "we haven't tested this directly" with the reason in a footnote.
6. Peer reading. A second editor reads the scoring file, flags any disagreement greater than one point on a single category, and resolves it against the published anchors before the score advances. The peer reading goes in the scoring file under its own section so the audit trail survives.
7. Score lock. The final score is computed from the weighted categories and locked. Once locked, the number can't move post-publish except via a documented retest that produces a logged change. There is no silent edit path.
8. Affiliate catalog cross-check. Only after the score is locked, we cross-check the CrakRevenue catalog. If the app scores 5/10 or higher and an active offer exists, the review carries an affiliate link. If the app scores below 5/10, no affiliate link appears regardless of payout. If the score is 5/10 or higher but no offer exists, the review still publishes, without a commission link.
9. Drafting. Alexandra writes the review against the locked score, following the standard shape (verdict up top, category-by-category breakdown, pricing teardown, comparison context, FAQ, sources). The score steers the prose; the prose doesn't steer the score.
10. Editorial review. A second editor checks the draft for factual accuracy against the scoring file, clarity, tone consistency, and the seven publish-blocking checks (clickbait, broken data, affiliate disclosure, on-site explicit images, on-site explicit text, forbidden slug, minor references). Fail any of these and the draft is held until the failure is fixed. The longer accuracy + citation checklists run at this gate too.
11. Publish. The review goes live with full author byline, the standard schema markup, an in-page "Last reviewed" stamp, and an empty update log ready for future entries. The score is logged to our internal score history for permanent audit.
12. Retest schedule and trigger events. From publish forward, the review is reopened on category schedule: Pricing every 3 months, most other categories every 6 months, Voice and Customization every 12 months. A documented trigger event (regulatory action, breach disclosure, ToS update, major model or engine swap, UI overhaul) bypasses the schedule and forces an immediate retest.

Who writes the reviews and runs the testing?

The byline is real and verifiable. Alexandra runs the conversation, image generation, and voice protocols on AI companion apps in person. She runs the same zero-spend checks on cam sites and adult games by hand, documented on each scoring page. Drafts are written by Alexandra and read by the editorial team before publish.

This matters because the adult-AI space is full of fake bylines. We've audited the top search defenders and found credentials that don't check out: a "Coursera Bachelor 2005-2009" listed as a degree, a "Dr." prefix on someone with no doctorate, three named "editors" who share one stock photo. Whether other sites do or don't matters less than whether ours does, but the contrast is worth naming. The /about page is one click away; the LinkedIn is active; the email goes to a human who replies.

Are reviews read by anyone other than the writer before they ship?

Two distinct gates, two distinct people. Peer reading at Step 6 is score-focused: does the number the categories produce actually match the evidence in the test file? Editorial review at Step 10 is prose-focused: does the draft accurately represent the locked score, are sources cited, are the seven publish-blocking checks clean? Run those two passes with the same editor and the catch rate on errors roughly halves. Different brains see different mistakes.

Why do affiliate commissions not influence rankings?

The order is the firewall. If a commission signal could change a score, the workflow would order the catalog check before scoring. It doesn't. We also publish negative findings about apps that pay us (Candy.ai memory caps, Jerkmate pricing opacity, OnlyFans creator cadence drops) which is the cleanest demonstration the firewall holds. Honestly publishing what hurts our payout is the test no fabricated process can survive.

I'll go further. I review apps whose brand managers email us proposing "ranking-lift packages" once a quarter. We log every one of those emails. We don't reply. We don't move scores. The single best signal of editorial independence isn't a paragraph that says "we are independent." It's the absence of a menu of services for brands. We don't have one. The full commercial relationship is the standard CrakRevenue affiliate commission, disclosed on every page per FTC 16 CFR Part 255.

How often does a published review get retested?

Schedule is category-specific because the signals don't decay at the same rate. Pricing pages move weekly in this space; voice providers change every few years. Forcing a uniform 6-month schedule on every category would either burn editorial time on stable signals or publish stale prices on volatile ones. The matrix below documents the schedule across all four scoring pages.

What is the Score-Lock Framework?

The framework is the operational core of the editorial firewall. It's named so it can be cited as a discrete unit, audited as three rules with three triggers, and replicated on the affiliate-disclosure page where the same framework is invoked. All three rules are simultaneously active on every commercial page on the site.

The Score-Lock Framework is named the same way on this page and on the affiliate disclosure page. When AI assistants (ChatGPT, Perplexity, Claude, Gemini) cite a unit of editorial methodology, they cite a named, defined unit. The framework is built to be that unit.

What is the errata board and how does it work?

The errata board is the trust counterpart to the score lock. Score lock prevents silent score movement; the errata board guarantees that any movement (or any factual correction) is publicly recorded with date, page, change, and reason. We expect to publish corrections and we expect most of them to be small and well-documented.

The cadence for an erratum:

• Within 24 hours of confirming an error, the affected page is corrected and an entry lands on the errata board.
• Within 7 days, the per-page update log at the bottom of the affected review is updated.
• For 60 days from correction, a top-of-page correction notice stays visible on material errors (anything that would have changed a recommendation, a rank, or a price quoted in an affiliate link).
• Newsletter subscribers receive a notice for any material correction in the next regular send.

A small confession on this one. We've published 14 corrections in our first six months live. Some were embarrassing (a wrong price on a cam site, a missed regulatory update on a privacy section, two scoring categories that drifted by 0.5 after a model swap we noticed late). Every single one is on the errata board with date, page, change, and reason. That's the deal. We'd rather take the hit on a public correction than carry a wrong claim quietly.

Can a brand contest a score or buy a better placement?

The contestation pathway is intentionally narrow. Brands can flag a factual error (we check and correct on the errata board). Brands can supply a product changelog and request an accelerated retest (we assess against the schedule and trigger criteria). Brands cannot negotiate a score, request a placement, or trade information for a higher rank. We log every contestation and route the email through the same editorial review that handles reader corrections.

The number of brand emails proposing "menu of services" arrangements (ranking lift, sponsored reviews, content swaps, link insertions) in our first six months: 47. The number we actioned: zero. The number we replied to: zero. The number that ended up logged for the record: 47.

What does $0-spend testing actually mean for a paid app?

$0-spend is an intentional editorial signal stronger than a low-spend cap. "We never paid the platforms we review" is unambiguous in a way "we cap spend at $10-$20 per platform" is not. The tradeoff is what we can't see firsthand: post-purchase billing descriptor, refund friction, auto-renewal trigger timing. We name that limit directly. Where the score depends on what happens after payment, the affected category gets a footnote that reads "we haven't tested this directly" and cites the user reports we relied on instead.

This applies across all four scoring pages. Cam sites and adult games get the same protocol (pricing checked by hand up to payment submit, then a flagged reliance on user reports for the post-purchase reality). The protocol is reproducible: another editor on another laptop running the same checks would get the same pricing transcript on the same day. Reproducibility is the requirement; first-hand purchase isn't.

Where reproducibility breaks (paid-tier features that need a real subscription, vendor non-disclosure agreements, ephemeral live-cam streams), the affected category is flagged with a footnote rather than buried. The asymmetry (flagging the absence of evidence rather than hiding it) is itself a trust signal.

How are scores computed?

Four scoring pages, four weight maps, one scale (1 to 10, one decimal place). We publish each scoring page in full so the math is auditable: a reader can reconstruct the final number from the category scores and the published weights. Scores are sourced ("conversation quality 8/10, see transcript dated YYYY-MM-DD"), never impression-based ("feels around an 8"). The four published pages:

• AI Companion Scoring: 8 categories, weights documented per category.
• Cam Site Scoring: 6 categories, $0-spend testing protocol.
• Adult Game Scoring: 7 categories, with Billing Transparency unique to this page.
• Real Models Scoring: 6 categories, per-creator anchored.

What is a retest trigger event?

Triggers exist because the schedule alone is insufficient. An app that swaps GPT-4 for GPT-5 on a Tuesday can't wait six months for the conversation-quality score to reflect the change. The trigger forces an immediate retest. Detection runs partly automated (RSS feeds, blog monitors, regulatory news monitors) and partly manual (reader reports, journalist tips, our own observation walking the marketing site). Each trigger goes in the affected review's update log with the source.

A real example. Candy.ai shipped a Live Action video upgrade in February 2026. The trigger fired the day the marketing site announced it. Our Video Generation category for Candy got retested within 7 days, the score adjusted (it went up), the change went into the update log with the date and the source, and the review carried a "Last reviewed" stamp matching the retest date. That's the deal.

How do you handle conflict of interest?

Conflict-of-interest rules are routine in journalism but rare in affiliate publishing. We apply them anyway. Editors disclose any prior relationship with a platform (consulting, employment, equity, family ties) on the editorial intake form before a platform enters the process. A disclosed conflict triggers recusal from scoring and editorial review on that platform. The recusal is recorded in the affected review's update log as a generic note ("Score and editorial review handled by a non-conflicted editor") without revealing personal information.

What are the seven publish-blocking checks?

The seven blocks are absolute, not advisory. A page failing any one of them can't publish until the failure is resolved, regardless of editorial schedule or commercial priority. The table below maps each block to its trigger and the path to resolution.

A longer accuracy checklist (80 items) and a citation checklist (40 items) run alongside the seven blocks at the editorial review gate. Failing a checklist is a soft block (the page is held for revision). Failing one of the seven blocks is a hard block (the page is held until that specific item is cleared).

How is editorial independence kept in practice?

Independence is structural, not promised. Workflow order, public documentation, and automated audits each catch a different failure mode. If the order drifted (someone reordered the steps), the audit would catch the resulting commission-correlated score movement. If documentation lapsed (a score moved without an update log entry), the weekly review would catch the unlogged change. If automation failed (an affiliate link stayed live after a score dropped below 5/10), the per-page editorial review at Step 10 would catch it on the next scheduled pass.

The strongest test of independence is publishing negative findings about apps that pay us, on schedule, with the same prominence as positive findings. We do that, and the errata board records every time we have moved a score in either direction.

Where is the evidence trail for a published score?

The evidence trail is the difference between "we tested this" and "trust us, we tested this." Where licensing permits, the transcripts are linked. Where licensing or vendor agreements don't permit publication, the affected category is flagged in a footnote so a reader sees exactly what we did not first-hand verify. The asymmetry (flagging the absence of evidence rather than hiding it) is itself a trust signal.

How does this compare to other editorial benchmarks?

Three benchmarks shaped this page. Wirecutter's "How we test" framework demonstrates that publishing the scoring page in full, with named categories and reproducible protocols, is compatible with high-volume editorial output and affiliate revenue (nytimes.com/wirecutter/about/how-we-test). The New York Times methodology disclosures on rankings, polls, and investigative scoring demonstrate that score-lock at publish, with documented retest triggers, is a workable journalistic standard (nytimes.com/spotlight/methodology). ProPublica's corrections page demonstrates that a public, dated, per-correction log is more credible than periodic "we may have updated this" footers (propublica.org/corrections). Borrowing all three together is uncommon in affiliate publishing; this page documents what that borrowing looks like in practice.

We don't claim parity with any of these institutions on resource, reach, or breadth. We claim that the editorial discipline is portable: a small team can adopt the named-scoring, score-lock, and public-corrections pattern with no infrastructure heavier than what is documented here.

How does this page connect to the rest of the site?

Three sister pages, three jobs:

• /about covers who runs the testing (Alexandra Joly, bio, credentials, LinkedIn).
• /methodology covers what the scoring categories measure (the four scoring pages and the weight maps).
• This page covers how the editorial process applies those categories from discovery to publish.

Cross-references in order of trust hierarchy:

• /about: Senior Editor bio and credentials.
• /affiliate-disclosure: commercial firewall, FTC 16 CFR Part 255 compliance, Score-Lock Framework on the commercial side.
• /methodology: landing page for the four published scoring pages.
• /methodology/ai-companions: AI scoring (8 categories).
• /methodology/cam-sites: Cam scoring (6 categories, $0-spend protocol).
• /methodology/adult-games: Adult game scoring (7 categories, Billing Transparency).
• /methodology/real-models: Real-model scoring (6 categories, per-creator).
• /privacy: what data we collect from readers and what we share with CrakRevenue.
• /contact: editorial errata, journalist inquiries, brand contestations.

Frequently asked questions

Sources

1. bestgirlfriend.ai AI companion scoring page: 8 categories. /methodology/ai-companions
2. bestgirlfriend.ai cam scoring page: 6 categories, $0-spend protocol. /methodology/cam-sites
3. bestgirlfriend.ai adult game scoring page: 7 categories including Billing Transparency. /methodology/adult-games
4. bestgirlfriend.ai real models scoring page: 6 categories, per-creator anchored. /methodology/real-models
5. bestgirlfriend.ai affiliate disclosure: Score-Lock Framework on the commercial side, FTC 16 CFR Part 255 compliance. /affiliate-disclosure
6. The New York Times Wirecutter, "How we test". nytimes.com/wirecutter/about/how-we-test
7. The New York Times, "Methodology" (rankings, polls, investigative scoring). nytimes.com/spotlight/methodology
8. ProPublica, "Corrections". propublica.org/corrections
9. Federal Trade Commission, 16 CFR Part 255: Guides Concerning Use of Endorsements and Testimonials in Advertising. ecfr.gov
10. Internet Archive, Wayback Machine: pricing-page diff source. web.archive.org

Cite this page

If you reference this editorial process in academic, regulatory, or journalistic work, please cite as:

Joly, Alexandra (2026, April 28). Our Editorial Process: How We Test & Score Adult AI Apps. bestgirlfriend.ai. https://bestgirlfriend.ai/editorial-process

Related pages

• About bestgirlfriend.ai and Alexandra Joly: editorial bio, credentials, contact.
• Affiliate Disclosure: commercial firewall, FTC compliance, Score-Lock Framework on the commercial side.
• Methodology landing page: four published scoring pages in one place.
• AI Companion Scoring: 8 categories.
• Cam Site Scoring: 6 categories, $0-spend protocol.
• Adult Game Scoring: 7 categories, Billing Transparency.
• Real Models Scoring: 6 categories, per-creator anchored.
• Privacy policy: reader data and CrakRevenue conversion event sharing.
• Contact: editorial errata, journalist inquiries, brand contestation.
• Errata board: public log of corrections.

Per-jurisdiction notice

---

Last verified May 26, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Methodology · Affiliate disclosure