How the apps are checked
What this page covers
This page explains, step by step, how the reviews on AI Partner Guides are made, what the labels on them mean and how the scores are worked out. It is written so that you can judge how far to trust any page on the site. Where the method has limits, this page says so plainly, because a method with hidden limits is worse than one with honest ones.
The short version: every review rests on two things. The first is Chris Furrey's own use of the app at some point between spring 2023 and September 2026, written down in his notes at the time: when he used it, which plan he paid for, what he spent and what happened. The second is a dated re-check of each app's current prices, plans, terms and screens on 23-24 September 2026, so that the facts in a review describe the app as it is now and not only as it was when he used it.
Of the 91 apps covered, 73 were used on a paid plan, 17 on the free tier only, and one is a desk review. 90 of them have a score.
Step one: the author's own use
Every review starts from Chris Furrey's notes. He has used AI companion apps since spring 2023, pays for the plans with his own card and keeps every charge in a subscriptions spreadsheet. For each app the notes record the period he used it, the plan he paid for, what he spent in all, including coin or token packs bought on top of a plan, and short dated entries about what happened: what the memory held on to, where the pricing surprised him, what the pictures looked like and how cancelling went.
Those notes are printed on the review as a dated log. Everything a review says in the first person has to come from that log. A review never adds a moment, a conversation or a charge that is not in it, and it never stretches a short stay into a long one.
The periods differ a lot. Some apps were used for a few days, most for about a month, and a few for more than a year. Each review says when the use took place, so you can tell whether a first-hand line describes last month or two years ago.
Step two: the dated re-check
Apps change faster than notes age. A plan that cost one amount in 2024 may cost another now, a free tier may have shrunk, and terms may have been rewritten. So each app's current prices, plans, free allowance, terms, privacy policy and main screens were re-checked on 23-24 September 2026, and every fact from that re-check carries its date in the record.
The re-check reads the app's own sources: its pricing page, its plan screens, its help centre, its terms of service, its privacy policy and its listing in the Apple App Store or Google Play. Prices shown in another currency are kept in that currency and never converted, because conversion rates and bank fees would make the number look more precise than it is.
Where the notes and the re-check disagree, for example a price the author paid in 2024 and the price the app asks today, the review keeps both and says which is which: the older figure with its date, the current one as today's price. Seven of the apps have closed since he used them; their reviews say what happened, and their prices stop printing. When two official sources disagree, both are kept with their dates, and the page uses the newer or the more careful of the two until the difference is explained.
Step three: other reviewers, cross-checked
A few facts are covered neither by the author's notes nor by the re-check, such as a feature he never tried or how an app's support has handled other people's refunds. For those, a review may use what established review sites report. One report is not enough; a point goes into the record only when it is consistent across sources or clearly attributed to one of them.
On the page, such a point is always written as what other reviewers report, never as the author's own experience.
This is the part of the method where readers should be most careful, and the pages are written to make that easy: you can always tell a first-hand line, a checked fact and a reported one apart by the wording.
- Apps tested
- 90hands-on, 73 of them on a paid plan
- Money spent
- $2,083subscriptions, paid with my own card
- Apps in the database
- 79every price with a check date
- Latest test
- FunFun AI
Computed from the test database when this page was built. The same numbers on every page.
Step four: screenshots
The screenshots on the reviews come from the dated re-check on 23-24 September 2026: interfaces, plan screens, settings and terms. Before a screenshot is published, it passes two independent checks: no people in the picture, nothing suggestive, and private data such as an email address or a user name blurred.
How a record becomes a page
Take the free tier of one app. The record says what the app publishes about it, what the re-check found and what the author's notes say about living with it, each with its date. The review prints today's allowance in the box at the top, explains in the text what it means in practice, and says who it suits: a free start that ends after a handful of messages is a demo, one that resets daily is somewhere you can stay.
The same fact then appears, in the same words and with the same date, on every list that includes the app.
Each review names the practical downside a reader is most likely to meet, whether that is image costs on top of the plan, a free tier that stops without warning or a token meter that turned a cheap plan into an expensive month. Reviews and lists print it in the text, never hidden in a footnote.
The three labels
Every review carries one of three labels, printed from the record.
| Label | What it means | Apps with it now |
|---|---|---|
| Hands-on test, paid plan | The author paid for a plan and used it; his dated log is on the page | 73 |
| Hands-on test, free tier | The author used the app on its free tier only; his dated log is on the page | 17 |
| Desk review - not paid for yet | Built from the app's own pages and the dated re-check, with no use of the app behind it | 1 (Dreemy AI) |
A label can only move up when a new dated log exists, and it moves down if a log turns out to be wrong. A partner programme, a request from an app or a confident paragraph cannot change it. The paid label does not mean that every plan of the app was used: the log names the plan the author paid for, and the other plans are described from the re-check.
The scoring rubric
Every app with a hands-on label has a score out of ten. The score is Chris Furrey's own judgement, formed from his use of the app and from the dated re-check, on seven criteria with fixed weights. The weights are the same for every app. The overall score is the weighted mean of the seven.
Chris Furrey can also add a personal author's bonus, always shown as its own line in the score box.
| Criterion | Weight | What it asks |
|---|---|---|
| Conversation | 25% | Does the character stay coherent, in role and engaging? |
| Memory | 15% | Does it remember what you told it, days later? |
| Pricing value | 15% | What does regular use really cost, including coins or tokens? |
| Privacy | 15% | What happens to your chats, and can you delete them? |
| Customisation | 10% | How far can you shape the character? |
| Media | 10% | Images, voice and video: available, and at what cost? |
| Reliability | 10% | Does it work consistently, and is the company stable? |
Conversation and memory can only be judged from time with the app, so they rest on the author's own use. What the plans cost today and what the privacy policy says today come from the re-check. The breakdown by criterion is printed next to each score.
Apps used only on the free tier are scored on the same seven criteria, and their label tells you that the score rests on free use.
The one desk review, Dreemy AI, has no score and shows "Not rated yet" in place of a number, because a score needs time with the app. A score is never given on the strength of marketing pages alone.
How lists are ordered
Where the site lists all of its reviews together, rated apps are ordered by their score. Every best-of list opens with the author's three top picks, SpicyChat, Nomi AI and Candy AI, in that order, even where another app fits the topic better. The other apps are chosen by how well they fit the topic, and apps with a partner programme can be placed higher.
The affiliate disclosure explains that openly, and every entry prints the same checked facts and score whether the app pays a commission or not.
The 30-day rule for prices
Prices are the facts most likely to change and the ones that cost readers money when they are wrong. Every price and every deal on the site carries the date it was checked, and the build that publishes the site refuses to print any price older than thirty days. An outdated price therefore disappears instead of lingering.
To stay on the page, every price has to be checked again at least once a month. A price the author paid years ago can still appear in a review, but only with its date, as what he paid then, and never as today's price.
Deals follow the same rule. A coupon or discount appears only with the date it was seen and only while that date is fresh. There are no countdown timers and no "ends tonight" banners, because the site cannot know when an app will end an offer.
Checks the build runs on every page
Before any page is published, an automatic build checks it. Every price printed in a table must exist in the records. No price may be older than thirty days. Every internal link must lead to a real page, and every sales button's label must match where it goes.
The wording must follow the site's content rules. And no page may claim that the author used, paid for or tested something unless a dated log backs it; a sentence like that stops the page from being published until it is rewritten. A page that fails any of these checks does not go live.
What the method cannot tell you
One person's use of an app is one experience, not a survey. Memory, filters and reply quality can vary by character, by model, by plan and by week, and an app that behaved well for Chris Furrey in 2024 may behave differently now. The re-check shows what an app publishes and what its screens look like today; it cannot show how its chat feels today.
It cannot promise that your account will get the same free allowance either. Where his use was long ago, the review says so, and on questions that neither his notes nor the re-check answer, the pages quote other reviewers, clearly labelled, or say that the answer is not known yet.
What the method can do well is tell you what each app costs today, what it allows, how it bills, what it does with your chats and how to leave it, next to what it was really like to use it and, for most apps, to pay for it. Those are the facts behind most regrets in this category.
Why the notes and the re-check are kept apart
It would be easy to fill the site with stories of long evenings spent with each app. Many sites in this category do, and many of those stories cannot be true, because the same invented moments turn up on site after site. The stories on AI Partner Guides come from one place: the author's own notes, dated and printed on each review as a log. Nothing is added to them to make a page more colourful.
Keeping the notes and the re-check apart matters just as much. A price he paid in 2024 is not today's price, and a free tier he used in 2023 may not exist any more. Each review says which facts come from his use and when, and which come from the re-check of 23-24 September 2026, so you never have to guess whether a sentence describes the app as it was or as it is.
When facts change
When an app changes a price, a plan or a policy, the record is updated with the new date and source, and the site is rebuilt so every page that uses the fact changes at once. If the site had been printing something wrong, the fix is also logged on the corrections page. If you notice a change first, please write through the contact page; a link to the app's own page or a dated screenshot is the fastest way to confirm it.
Where this method is going
When Chris Furrey goes back to an app, the new period is added to its log with its dates, and the score can change with it, printed with its new breakdown. The one desk review will get a hands-on label and a score only when a dated log exists behind them. Nothing on the site will claim more than the logs support, however long the site has been running. The editorial policy sets out the wider rules this method follows.