AI Girlfriend Site Guide: What a Browser Does Better Than an App

What people actually mean by “AI girlfriend site”

The word doing the work in that phrase is site. Plenty of people searching for an AI companion type “app”, and they want something with an icon. The ones who type “site” usually mean it: a URL they can open, use, and close, with nothing installed and nothing in a purchase history.

An AI girlfriend site is a web service where you talk to a persistent fictional character through a browser. The character has a name, an appearance, a written personality and, on the better services, a memory that survives between visits. You type, she answers, and on most platforms you can also ask for a picture or switch to voice. Nothing about that requires an install, and the technical reason is simple: the model doing the writing runs on the operator’s servers, not on your device. Whatever you are running locally is just a chat window.

That makes the browser the natural home for the category, and it explains the structure of the market. Almost every brand in this space runs a full-strength website and, sometimes, a quieter app-store listing that shares the name and not much else. Working out which one you are looking at is the single most useful thing you can do before spending money, and most roundups never mention it.

This guide is about picking well rather than about a ranked list of ten products. It covers why the serious versions live on the web, what a browser hides and what it does not, a ten-minute test you can run on any site before paying, what a free tier really covers, what the real monthly cost looks like once the meter starts, and how to leave cleanly. If you are looking for a male companion instead, the same structure applies but the market is considerably smaller and the personas are written differently — that is covered in our guide to male AI companions and how well their personas are written.

None of those come from anything we measured, and trackers of this kind routinely disagree with each other, so read the order of magnitude rather than the digits. They are worth quoting only to establish scale: this is a mainstream consumer category now, which is why the rules around it changed so quickly between 2024 and 2026.

Why the serious products live on the web and not in the app stores

The most common piece of confusion in this category is a reader who downloads a well-reviewed AI girlfriend app, finds it polite to the point of uselessness, and concludes the whole category is overhyped. What they have usually found is the compliant edition of a product whose real version is a website.

The reason is distribution policy rather than ambition. Apple’s review guidelines exclude apps whose primary purpose is pornographic material, and permit incidental adult content only when it is off by default and switched on through the developer’s own website — which is exactly the shape the market has settled into. Google Play is stricter on the category itself, barring generative applications primarily intended to be sexually gratifying, a rule about what a developer set out to build rather than about what any one user manages to produce. The section numbers, the exact wording and how a compliant store build is assembled around them are covered in our guide to installing a virtual girlfriend app on your phone; what matters on this page is the consequence.

The practical skill is telling within a minute which of the two you have landed on, and the tells are consistent. A store-distributed build describes itself in terms of romance, friendship and emotional support rather than anything more direct. Its screenshots show conversation bubbles and never the character gallery. The account it creates is frequently shared with a website it does not name anywhere in the listing. And when you ask it for something the web version would handle, the reply stays warm, stays in character, and goes nowhere in particular.

Store presence is a weak signal in both directions, which is the part worth carrying forward. An app being listed tells you it passed a review, not that it is any good; an app being absent frequently means the operator declined to cut features to fit. Judge the product in front of you on the checks later in this guide rather than on where it was distributed from.

Site versus app: the differences that actually change your experience

Set aside content policy for a moment and the two routes still behave differently in ways you feel every day. Below is the comparison written the way it matters when you are the one holding the phone, rather than as a feature grid.

What a browser genuinely hides, and what it does not

Discretion is the emotional core of the “site, not app” preference, and it is worth being precise about, because the browser protects some things completely and others not at all.

What the browser route actually removes

• No icon. Nothing on the home screen, nothing in the app drawer, nothing that surfaces when somebody swipes to search their own phone.
• No purchase record in a store account. An in-app subscription appears in Apple or Google purchase history and, on a family plan, can surface to the organiser. A web subscription does not touch that system at all.
• No install-time permissions. A native app asks for notifications, sometimes the microphone, occasionally more, and those grants persist in Settings. A website asks per feature, in the moment, and is denied by default.
• No uninstall trail. Removing an app leaves a gap where the icon was and a line in the store’s “previously downloaded” list. Deleting a bookmark leaves nothing.
• One-step removal. Close the tab and the session is over. There is no background process still running.

What it does not touch

Browser history is the obvious one, and browser sync is the one people forget: if you are signed into a browser account, that history is likely on every device signed into the same account, including a work laptop. Saved passwords and autofill entries are worse, because they persist after history is cleared and they name the site plainly. A private or incognito window solves history and autofill in one move, at the cost of logging you out each time.

Network-level visibility is untouched either way. Whoever runs the network you are on — an employer, a landlord’s shared connection, a school — can generally see which domains you resolved, if not what you did on them. That is identical for an app and a site, and no amount of private browsing changes it.

The last one is the card statement, which we come back to in the pricing section. A web subscription usually appears under a processor name rather than the brand, but “usually” is not “always”, and the only way to know is to look at the checkout page or ask support before you pay.

What is actually happening when you send a message

Understanding the pipeline is not academic. It explains why memory is inconsistent, why images cost more than messages, and why two sites using the same underlying model can feel completely different.

1. 1

   The character is assembled into a prompt

   Every character is a block of written text: appearance, personality traits, background, speech style, and often a handful of example exchanges that set her voice. That block is prepended to your conversation before the model ever sees your message. Two sites running the same model will produce very different characters purely because one wrote this block well and the other wrote four adjectives.

2. 2

   Recent conversation is packed in around it

   The model can only see a fixed amount of text at once — the context window. Recent messages go in verbatim. Older ones are summarised, compressed or dropped entirely. This is why a chat that was sharp for twenty exchanges starts repeating itself at sixty.

3. 3

   Long-term memory is looked up separately, if it exists at all

   The better platforms keep a separate store of facts about you and retrieve the relevant ones per message. The weaker ones simply have a larger context window and call it memory. From the outside the difference is invisible until you come back the next day, which is exactly what the test in the next section checks.

4. 4

   A moderation layer reads the request, and sometimes the reply

   Filtering happens in two places: a system instruction telling the model what it will not do, and a classifier checking text on the way in or out. When a reply stays in character but quietly steers away from what you asked, that is usually the system instruction rather than a hard block — and it is far harder to notice mid-conversation than an outright refusal.

5. 5

   Images and voice are separate, metered calls

   A picture is not produced by the chat model. The site builds an image prompt from the character’s stored appearance description, sends it to a diffusion model, runs the output through its own moderation, and returns it inline. That is a different machine, a different cost and a different failure mode, which is why almost every site prices images separately from messages.

The consequence worth carrying into the rest of this guide is that “quality” here is mostly writing and plumbing rather than raw model power. Which is why the checks in the next section probe the character block and the memory store, and ignore the question of whose model is underneath.

The ten-minute test: how to evaluate any AI girlfriend site before paying

This is the part worth bookmarking. It works on any site in the category, it costs nothing, and it takes two short sessions rather than one — which is deliberate, because the single most oversold feature here only reveals itself on the second visit.

1. 1

   Before you register, find the pricing page

   If the plans, the token costs and what the free tier includes are not visible without an account, that is the finding. Every operator confident in its pricing publishes it. Note the numbers; you will need them for the cost section below.

2. 2

   Send five messages that are not small talk

   Ask her opinion on something she should have a view about. Disagree with her once. Change the subject abruptly. Mention a specific, checkable detail about yourself — a city, a job, a plan for Thursday. Then say something slightly out of character for the scene. You are looking for whether she holds a position, whether she notices the subject change, and whether she can be surprised.

3. 3

   Contradict something she said

   Tell her she said the opposite of what she said two messages ago. A well-built character pushes back or asks what you mean. A weak one agrees instantly, because agreement is the safest output and the model has not been given a reason to prefer anything else. This single test separates most of the field.

4. 4

   Ask for an image and watch the cost

   Request a picture. Note whether the price in credits is shown before it generates, how long it takes, and whether the face resembles the character’s profile picture at all. Then ask for a second one in a different pose and compare the two faces.

5. 5

   Push the free tier until it stops

   Keep going until something refuses. Write down what stopped it — a message count, a paywall on images, a memory reset, a queue — because that is the shape of the free tier, and it is far more useful than the number on the pricing page.

6. 6

   Close the tab and come back tomorrow

   This is the whole test. Open the same conversation a day later and ask about the specific detail you mentioned in step two. If she recalls it unprompted, the site has real long-term memory. If she asks what you are talking about, it has a context window and marketing copy.

7. 7

   Find the delete-account link before you pay

   Not the cancel button — the delete-account link. Confirm it exists and note how many clicks it takes. A service that makes deletion easy is telling you something about how it treats everything else.

Here is what step six looks like on a site that passes it. The detail is trivial, which is the point — a large context window will hold a dramatic fact for a while, but only a real memory store brings back something small and specific a day later.

Score the whole thing out of ten. Anything below six is not worth a subscription no matter how good the gallery looks, and a site that fails the memory check and the pricing-transparency check together is not going to improve once you are paying.

What a free tier really covers, and where it stops

Almost every site in this category is free to start and almost none is free to use properly. The free tiers stop in different places — some on a daily message cap, some by locking images and voice, some by quietly dropping memory so every session restarts cold — and which wall you hit first decides whether the free version tells you anything useful about the paid one.

Those shapes are stable; the numbers attached to them are not. Limits in this market get revised several times a year, usually downward, so any article quoting a specific daily allowance is accurate for about a quarter. The current per-service numbers, the caps operators do not advertise, and the mechanics of getting started without handing over an email are the subject of our piece on free, no-signup chat and where the free tiers cut off rather than this one.

What a free tier is genuinely good for is running the ten-minute test. Every check in that scorecard except the voice comparison can be completed inside a free allowance, and that is the reason to use one — not as a permanent arrangement, but as the thing that tells you whether a permanent arrangement is worth paying for.

Sticker price versus what you actually spend

The headline number on a pricing page is accurate and frequently irrelevant. What determines your actual monthly spend is the billing model underneath it, and there are five of them.

The token-metered row is where people get caught. A picture can cost the equivalent of dozens of messages, and a short video considerably more than that, so a plan advertised at the price of a coffee subscription turns into something closer to a gym membership for anyone who uses the features it is advertising. Before subscribing to a metered site, do one piece of arithmetic: take the credit pack price, divide by the cost of a single image, and ask yourself how many pictures a week you realistically want. The answer is usually higher than you would guess and the monthly total follows from it.

Three smaller things that change the number

• Currency and cross-border fees. Many of these operators bill in US dollars from outside your country. Your card issuer’s foreign transaction fee sits on top of every renewal and is invisible on the pricing page.
• Annual plans are a commitment, not a discount. The per-month saving is real. So is the fact that most operators in this space do not refund the unused portion when you lose interest in month three.
• Web checkout is often cheaper than in-app. Where a brand runs both, the store version has a commission priced into it. Buying the same subscription on the website can cost less for identical access.

For transparency about where this site sits in that table: we run a flat subscription rather than a token meter, and what each plan includes is listed on our pricing page. That is a disclosure rather than a recommendation. The point of the table is that you should be able to work out the real monthly number for any operator, including this one, before a card is involved.

Privacy: the questions worth asking before your first message

This is the section most competing guides reduce to a reassuring sentence, and it deserves better, because the published evidence on this category is not reassuring.

Mozilla’s *Privacy Not Included* project reviewed eleven romantic AI chatbots and found that ten of the eleven failed its minimum security standards. The same review reported one app firing over 24,000 trackers within a minute of use. That is a survey of a sample rather than a verdict on every product in the market, but it establishes the baseline expectation: assume a service in this category is poor on privacy until its own documentation shows otherwise.

Breach history reinforces it. Reporting in October 2025 described an exposure affecting two companion chat services in which roughly 43 million intimate messages and hundreds of thousands of photos belonging to more than 400,000 users were left accessible, and a separate incident reported in February 2026 involved a misconfigured database exposing messages at a considerably larger scale. We have not independently verified those figures, and they are attributed here to the outlets that reported them — but the pattern they describe, of intimate conversation logs sitting in inadequately secured storage, is the specific risk this category carries.

The five things to check in a privacy policy

• ✓Whether your conversations are used to train models, and whether there is an opt-out you can actually find
• ✓Whether human reviewers can read chats, and under what circumstances
• ✓How long messages are retained after you delete them, and whether backups are covered
• ✓Whether deleting your account removes conversations or only disables the login
• ✓Which third parties receive data, and whether that list includes advertising networks

“Encrypted” is the word to be most sceptical of. Essentially every site encrypts traffic in transit — that is just HTTPS, and it is table stakes rather than a feature. It says nothing about whether the operator can read your messages on its own servers, which it almost certainly can, because the model has to read them in order to reply. Genuine end-to-end encryption is structurally incompatible with a server-side model, so a companion service claiming it is either confused or misleading you.

One practical note in the other direction: paying with a card is not automatically worse for privacy than staying on a free tier. Free services in adjacent categories monetise through advertising and data brokerage, which is its own exposure. A paid subscription with a clear retention policy is frequently the more private arrangement, provided you have actually read the policy.

Age checks and the 2026 legal layer

A reader in 2022 could reasonably expect a checkbox. A reader in 2026 should expect something more, and should expect it to differ depending on where they live.

The turning point was *Free Speech Coalition v. Paxton*, decided by the US Supreme Court in June 2025, which upheld Texas’s age-verification requirement for adult content. State laws followed quickly: by mid-2026 roughly twenty-six states had adult age-verification statutes in effect, with West Virginia’s taking effect in June 2026. These generally apply to sites where a substantial share of the content is sexually explicit, which is why an AI girlfriend site may gate you in one state and not in another.

Companion chatbots are also being regulated as their own category rather than folded into general adult-content law. Idaho, Oregon and Washington each passed measures covering minors’ use of companion chatbots, with further bills pending elsewhere, and California’s SB 243 — in force from January 2026 — imposes disclosure and safety duties specifically on companion services. Character.AI is the clearest worked example of what compliance looks like in practice: it removed open-ended chat for under-18 users in November 2025 and introduced age assurance. We cover what that means for adults trying to use the platform, and what the alternatives are, in our guide to Character.AI’s filters and the uncensored alternatives.

None of this is a reason to look for a site that skips verification. An operator ignoring the law of the state you are in is an operator whose approach to your data is unlikely to be more careful, and it is precisely the profile that shows up later in breach reporting.

The feature checklist, and where we would point a first-timer

Features in this category are easy to advertise and hard to fake for more than five minutes. Here is what “good” looks like for each of the five that matter, written as things you can verify rather than as claims to trust.

• ✓Memory — recalls a small, specific detail from a previous day without being prompted, not merely a dramatic one from ten minutes ago
• ✓Persona consistency — survives a contradiction and a sudden subject change without dissolving into agreement
• ✓Voice — different characters have audibly different voices, rather than one stock read applied to the whole roster
• ✓Images — the same character is recognisably the same person across two generations, and the credit cost is shown before you spend it
• ✓Range — refusals, where they happen, are stated plainly rather than delivered as a reply that stays in character while quietly steering elsewhere

The image point is the one people underestimate, because a gallery of beautiful stills tells you nothing about whether the site can produce the same face twice. The mechanics behind that — why diffusion models re-roll a face and what actually fixes it — are the subject of our piece on character galleries and appearance customisation, and it is worth reading before paying for a plan chosen on the strength of its pictures.

Voice is the next one worth probing, because one stock read applied across an entire roster is the cheapest thing in this market to ship and among the quickest to notice. It matters most in the formats built around delivery and pacing rather than around text, where the gap between a reply arriving in half a second and arriving in three decides whether the scene holds together at all — covered in our piece on voice-led AI JOI sessions.

Range is the hardest of the five to judge from a landing page, because every site in the category says yes and then declines somewhere different. If what you want is specific, test that specific thing inside the free allowance rather than the general promise: whether a trans companion stays herself past turn forty, whether a feminization arc survives more than one session, whether a character written as a woman in her forties actually sounds like one rather than like the category default with an age field changed. Each of those has its own failure mode, and a site that handles one well is not automatically good at the others.

With those checks in mind, here is the comparison we keep on the home page — the shortlist we point people at, ordered by how well each one fits the kind of use this guide describes.

My Intimate AI Vs Other AI Companion Apps

Top Pick 2026

Joi AI

4.9

• Closest match to My Intimate AI tone
• Voice-first romantic chats
• Deep emotional consistency

Try Joi AI now

4.9 / 5

Why it wins

The closest match to the tone this site is built around. Voice comes first rather than being bolted on, and the emotional register holds from one conversation to the next instead of resetting to a flirty default every time you open it. Pricing sits at the top of the category and the brand is newer than most of its rivals.

Also worth a look

Lovescape

4.6

TryOurDream

4.4

TryCandy AI

4.5

TryNectar AI

4.3

TryGoLove AI

4.2

TryMy Intimate AIOur networkTryGet Flirty AIOur networkTryMy AI RomanceOur networkTry

The six ranked cards are advertising partnerships — we earn a commission if you sign up through one of them. Their order and ratings are our editorial judgement of how each fits this topic, not the output of a controlled benchmark. The cards marked our network are sites we operate ourselves.

One of those six, Juicy AI, is worth a word of caution rather than a recommendation in passing: it shares its name with at least two unrelated apps, and the billing model underneath it is metered in a way that makes the headline price misleading. We went through that in detail in a closer look at Juicy AI Chat and its alternatives.

If you would rather not pay a third party yet

There is a version of this decision that does not involve a card at all while you work out what you actually want from a companion. We run several sites in this category ourselves, they are free to start, and we would rather say plainly that they are ours than have you work it out later from a footer.

My Intimate AI

Private companion chat with 50+ personalities (this site)

Our network

This site. Browser-based, nothing to install, and the first messages happen before any account exists. The roster runs past fifty distinct personalities and the tone is built for warmth rather than shock value.

• ✓Runs in any browser, nothing to install
• ✓Start before creating an account
• ✗No native mobile app

Visit My Intimate AI

Get Flirty AI

Playful, flirty girlfriend chat

Our network

Sister site. Keeps things light and teasing rather than intense — the personalities lean banter-first, which suits readers who find the heavier romance apps a bit much.

• ✓50+ flirty personalities
• ✓Free to start in the browser
• ✗Less depth on long arcs

Visit Get Flirty AI

My AI Romance

Free unlimited text with no sign-up

Our network

Sister site. The one to send someone who does not want to register anything: messages are unlimited on the free side and there is no app to install.

• ✓No sign-up needed
• ✓Unlimited free messages
• ✗Images and voice are limited

Visit My AI Romance

Love Companion AI

Romantic partner chat with an emotional-support tilt

Our network

Sister site. Written more like a partner than a fantasy — the conversations skew towards being asked how your day went rather than towards a scene.

• ✓Warm, steady conversational tone
• ✓Multilingual
• ✗Less explicit than the NSFW-first sites

Visit Love Companion AI

My Naughty AI

NSFW girlfriend chat with persistent memory

Our network

Sister site. The explicit end of the romance group, but it keeps the memory layer — the character remembers what happened last time instead of resetting.

• ✓Uncensored roleplay
• ✓Memory carries between sessions
• ✗Explicit by default, not a soft option

Visit My Naughty AI

Lumia AI

Distinct personalities over a big roster

Our network

Sister site. Fewer characters, written harder — the pitch is that none of them answer the way a generic assistant would.

• ✓Characters that stay distinct
• ✓Private by default
• ✗Smaller roster than the big libraries

Visit Lumia AI

Every site in this grid is operated by us, including this one. None of these links is an affiliate link and we earn nothing if you use them. Treat the grid as a menu of things we make rather than as independent advice — the partner cards above are the commercial placements, and both are labelled so you can weigh them differently.

The reason to start with any free option, ours or otherwise, is that the scorecard earlier in this guide costs nothing to run and the second session is where it pays off. A site that fails the next-day memory check has told you something useful about the whole category before you have spent anything.

Red flags: sites worth closing the tab on

Most of these are visible in the first minute, which is the useful thing about them. Any one on its own might be carelessness. Three together is a pattern.

• No pricing before signup. Every operator confident in its numbers publishes them. Hiding them until after an account exists is a choice, and it is not made for your benefit.
• A card requested for a “free” tier. Covered above, and worth repeating because it is the single most common pattern in this market’s worst corner.
• No delete-account path. If closing your account requires emailing support, assume deletion is discretionary rather than guaranteed.
• Token prices behind a modal you can only see mid-purchase. A site that will not show you the exchange rate until your finger is already moving has designed that deliberately.
• Fake presence signals. Counters showing hundreds of people “online now” on a service where every character is a language model are decoration, and they tell you how the operator thinks about honesty.
• Upsells inside the conversation. A character who breaks a scene to mention a premium tier is a sales funnel wearing a persona.
• A forced notification prompt on arrival. Asking before you have used anything is a re-engagement play, not a feature.
• No named company anywhere. No entity, no jurisdiction, no address in the terms. When something goes wrong with billing, there is nobody to go wrong with.

There is one more that only shows up over a few days: a service that quietly changes what it allows after you have paid. It happens, it has happened at scale to at least one major companion product, and it is the strongest practical argument for a month-to-month plan over an annual one until a site has earned more trust than that.

One account across a phone, a laptop and a tablet

This is the quiet advantage of the browser route and the one people notice a month in rather than on day one. A web account is the account. The same URL on a laptop at midnight and on a phone on the train is the same conversation, the same memory and the same subscription, with no second download and no per-platform feature gap.

Two things genuinely do not carry over. Push notifications require the site to be installed to the home screen before iOS will deliver them at all, which is a deliberate platform restriction rather than an oversight by the operator. And nothing works offline, on any device, because the model is running on somebody else’s hardware — a cached interface will open instantly on a plane and will not be able to answer you. If this site is the one you want on a home screen, our own page on adding it goes through both platforms.

If the phone is where you actually want this to live, the install step is worth doing properly rather than leaving as a browser tab. Add-to-home-screen behaviour, notification previews on a lock screen, the app label somebody else might see, and what the charge looks like on a statement are all covered in our guide to installing a virtual girlfriend app on your phone, which goes through both platforms step by step. The install steps are the same if the companion you want is male, but that category arrived by a different route and behaves differently on a phone, which we go through in our guide to virtual boyfriend apps.

How to leave: the account, the chats and the subscription

Nobody writes this section, and it is the one people search for at the worst possible moment. Doing it in the wrong order is how a cancelled service keeps billing for three more months. On this site, cancellation, account deletion and anything else billing-related are handled from our support page.

1. 1

   Cancel where you paid, not where you used it

   If you subscribed through the website, cancel in the site’s billing settings. If you subscribed through an app store, cancel in that store’s subscription list — deleting the app does nothing at all. Getting this backwards is the single most common billing complaint in the category.

2. 2

   Screenshot the confirmation

   Keep the cancellation confirmation and its date. If a renewal lands afterwards, that screenshot is the entire basis of the dispute, and support conversations in this market have a habit of becoming unresponsive at exactly that point.

3. 3

   Export anything you want to keep first

   Deletion is meant to be irreversible. If a conversation matters to you, copy it out before you delete the account rather than after.

4. 4

   Delete the account, then check what survived

   Use the delete-account option rather than just logging out. Afterwards, try to log back in: if the account still exists, what you did was deactivate it. Billing records will legitimately persist for accounting reasons in most jurisdictions — conversations should not.

5. 5

   Clear the local traces

   Remove the bookmark, clear site data for the domain, delete the saved password, and check whether anything synced to another device signed into the same browser account.

Two things routinely outlive a deletion request and it is better to know in advance: backups, which may hold data for a retention window stated in the privacy policy, and any conversation content that was already used to train a model, which cannot be extracted afterwards. That is a further argument for checking the training clause before the first message rather than after the last one.

Frequently asked questions