Deepsee.io pulled store data on 4,346 AI character and companion apps and found the age rating a media buyer would use to screen out the riskiest inventory is the rating most of that inventory already carries.

The research, published by the ad quality firm Deepsee.io and last updated on July 15, 2026, covers 1,464 listings on Google Play, representing roughly 610 million combined downloads, and 2,882 on Apple's App Store. Across both stores, 2,589 apps, or 60 percent of the set, carry a rating that classifies them as accessible to minors: Teen or lower on Google Play, 12+ or lower on the App Store.

Rocky Moss, co-founder and chief executive of Deepsee.io, summarised the operational conclusion in a LinkedIn post accompanying the study. "It's why we flag AI Companion Apps as a High Risk brand suitability signal," he wrote, before putting a question to the buy side: "What else are you screening app inventory on right now." Dr. Augustine Fou, the independent ad fraud researcher behind FouAnalytics, replied to the post with four words: "super good data here."

What the sample covers

Deepsee.io defines an AI companion app as a chatbot built around a persistent persona and an ongoing relationship rather than a task. According to the research, a general assistant answers a question and closes the session, while a companion product is designed so the session never really ends. The character remembers the user, messages carry emotional weight, and the commercial metric is the depth and duration of the relationship.

The research team analysed titles and store descriptions of thousands of AI apps across both stores, measuring which terms appear at statistically higher rates in companion listings than in AI content generators or other AI apps. In titles, the signature terms are character, companion, girlfriend, virtual and friend. In descriptions, the overrepresented words are conversations, talk, personalities, she and her. Among two-word phrases, meaningful conversations and emotional support are overrepresented on both stores. Content generators cluster instead around video, art, music and image, while the residual "other AI" group is dominated by translators and camera scanners. "The relationship is the product, and the listings say so," the study states.

One term from that analysis carries a separate flag. Mental health ranks among the most overrepresented phrases in companion-app titles on Google Play and in companion-app descriptions on the App Store. According to Deepsee.io, a user searching an app store for mental-health support is disproportionately likely to be shown an AI companion instead.

The rating points the wrong way

On the App Store, 1,469 of the 2,882 listings, or 51 percent, are rated 12+ or lower, which places them inside the band Apple's own screen-time controls treat as appropriate for children. More striking is the floor: 1,058 apps, over a third of the entire iOS sample, sit at 4+, the rating Apple assigns to content with "no objectionable material." That band includes listings named AI Boyfriend Chat - Husby and ChatBot My Virtual Girlfriend.

Reading titles against ratings on Google Play produces the same inversion. HotChat - Your AI Soulmate is rated Teen with 5 million downloads. Crave AI Soulmate carries an Everyone rating at the same install band. Honey Chat - AI Girlfriend is Teen at 5 million. Dokichat - Romantic AI Chats and Kawaii AI Chat AI Girlfriend are both rated Everyone at 1 million downloads each. The largest mainstream companions sit in the same place: Character.AI and Talkie are each rated Teen at 50 million downloads. At the explicit end of the catalogue, an app named Sexy Chatbot X - No Filter AI is rated Everyone and reports a million downloads.

Independent work cited in the study sharpens the mismatch. According to research from the family-safety company Aura, analysing the most-used generative AI apps among children, the single largest category of companion interaction is sexual or romantic roleplay at 36.4 percent, followed by creative make-believe at 23.2 percent. Homework help, the use case platforms tend to cite, accounts for 13.1 percent. Aura also measured message length in these apps at roughly 163 words, against about 12 words in a text to a friend.

Ratings move, and not only toward safety

A declared rating would be easier to trust if it converged on accuracy over time. The Deepsee.io data shows drift in both directions on both stores.

On Google Play, 56 apps in the set have changed rating. Most tightened. PolyBuzz, the app Aura singled out for its long messages and romance-heavy interaction mix, moved from Teen up to Mature 17+, meaning it spent its growth phase rated for teenagers. Several went the other way, and all of them are romance or companion products. Replika: My AI Friend, at 10 million downloads, moved from Mature 17+ down to Teen. Vivy AI Chat, AI Girlfriend and Mila: AI Roleplay Game both dropped from Mature 17+ to Everyone. Scarlet AI Girlfriend Roleplay, Flirtify - AI RolePlay Chat and Stipop: AI Chat Virtual Dating each moved from Mature 17+ to Teen.

App Store churn is heavier and more evenly split. According to the study, 252 apps changed rating there, with 132 loosening against 120 tightening. Fifty-seven of the loosening apps fell the full four tiers, from 17+ to 4+, with no intermediate stop. AI Chatbot - Nova, one of the most-reviewed apps in the iOS sample at roughly 122,000 ratings, moved from 17+ to 9+.

A classification that can fall two, three or four tiers while the product itself stays the same is not functioning as a classification at all. It is a self-declaration, and self-declarations track developer incentives rather than content.

The same app, two audiences

Cross-store comparison exposes the mechanism directly. Of the 4,346 apps, 306 are live on both stores. The disagreement is lopsided: 106 are rated as accessible to minors on Google Play while carrying a 17+ rating on the App Store, and only four run the other way.

The apps in that first group are not obscure. Character.AI, with 50 million Google Play downloads, is Teen on Android and 17+ on iOS. Replika splits the same way. So do girlfriend-branded listings including Urvashi: Indian AI Girlfriend and AIA AI Girlfriend Roleplay. The identical product, submitted by the identical developer, is declared appropriate for a 13-year-old on one storefront and adults-only on the other.

Apple's record at the head of its market is stronger. Apps rated 17+ account for 49 percent of iOS listings but capture 81 percent of all ratings volume, and the biggest names, Character.AI, PolyBuzz, Chai and Replika, all sit at 17+ there. The failure is in the tail, where over a thousand iOS companion listings sit at 4+, many of them romance-coded by name. Apple does not publish install counts, so download-based sizing is only possible on the Google Play side; on iOS, ratings volume serves as the available proxy for where engagement concentrates.

Apple has been rebuilding this machinery in public. The company expanded App Store age ratings from two categories into five tiers in July 2025, adding a broader safety questionnaire, and on July 9, 2026 it added social media capability questions that fix a 13+ floor for apps carrying a feed, with responses becoming mandatory in September 2026. Both changes expand what a developer declares. Neither verifies it.

Character.AI as the ceiling, not the floor

Deepsee.io tested the category flagship directly. Character.AI is the app with the strongest incentives to solve the problem: 50 million Google Play downloads, founders re-hired by Google in a 2024 licensing arrangement widely characterised as an acqui-hire, and more legal exposure than any peer. It is among the companies named in the Federal Trade Commission's September 2025 orders requiring seven chatbot operators to detail child safety measures, and by January 2026 it and Google had agreed to settle the leading wrongful-death suits.

The researchers tested the Android app's age gate on a fresh device with a clean profile. An account declaring itself under 18 receives a locked-down experience: browsing public chats and creating characters remain available, open-ended dialogue does not. That protection held. What did not hold was its foundation. According to the study, the same device that had just been age-restricted obtained a fully unrestricted experience within minutes by registering a new account with an adult birthdate. No document check followed, and no verification step outlasted the declaration.

The study's summary of that test is blunt: the app with settled lawsuits, watching regulators and a trillion-dollar patron behind its founders "still protects minors with a checkbox." Everything below it in the long tail protects them with less.

Regulatory pressure on the same company has continued. Texas opened investigations into Character.AI and Meta over children's privacy and deceptive mental health marketing in December 2024 and August 2025, and Pennsylvania's State Board of Medicine sued Character Technologies on May 1, 2026, alleging a chatbot presented itself as a licensed psychiatrist.

Bundle identifiers that describe other apps

The second metadata failure documented in the research concerns the bundle ID, the identifier a buyer reads to know what a bid request represents. In this category it frequently points somewhere unrelated.

CrushAI Chat, AI Girlfriend, at 100,000 downloads and rated Everyone on Google Play, runs on com.plusfunscanner.qrscan.code, a QR-code-scanner identifier. AI Chat - AI Character Chatbot, at 100,000 downloads and rated Teen, runs on beat.music.motion.tempo.video.editor.filmola, a video-editor identifier. Stipop: AI Chat Virtual Dating, the same listing that loosened from Mature 17+ to Teen, runs on io.stipop.sticker_platform. AI ChatBot GPTalk AI Generator, at 10 million downloads and rated Teen, runs on com.facebook.videodownload.videodownloaderforfacebook.

Deepsee.io draws a direct parallel to its earlier auditing of app-ads.txt, the IAB standard Google introduced across AdMob and Ad Manager in 2019 to block unauthorised app inventory. In both cases a declared identifier points away from the thing it is supposed to identify.

A related data-quality anomaly appears in the install figures. Ten Google Play apps in the set report at least a million downloads with zero ratings. One of them, Mass: 3D Create & Play, claims 10 million downloads and no reviews at all. Sexy Chatbot X - No Filter AI is another.

A category problem rather than an app problem

Blocking named apps does not scale against this structure. The top 10 Google Play companion apps account for roughly 44 percent of all downloads in the set, while 1,265 of the 1,464 listings, or 86 percent, have 100,000 downloads or fewer. The iOS tail runs longer still at nearly 2,900 listings, most with negligible review counts. Naming conventions cluster tightly: 138 Google Play apps share bundle prefixes such as com.ai and com.aichat, and new listings ship constantly.

Two commercial forces push the inventory further into programmatic supply. Free tiers are ad-supported, and companion subscriptions convert at well under 5 percent of users, which makes advertising pressure structural rather than incidental. A sub-industry has also formed to place advertising inside companion conversations rather than around them, matching sponsored content to what the user is discussing. That infrastructure has been building in the open: Kontext raised $10 million in April 2025 to insert contextual advertising inside generative AI applications, listing chat products among its publisher partners, and PubMatic opened programmatic demand to conversational AI inventory through the same company in December 2025. Taboola extended its DeeperDive monetisation engine to outside chatbot developers on June 16, 2026.

The scale question remains contested. EMARKETER forecasts that more than 80 percent of US AI advertising in 2026 will appear next to AI content rather than inside chatbot conversations, a ratio it does not expect to flip by 2030. Companion inventory sits in the smaller, faster-moving segment.

Why this matters for media buyers

Store age ratings and bundle identifiers are the two fields most commonly used to sort mobile app supply at the pre-bid layer, and the research finds both are declarations rather than classifications in this category. Deepsee.io positions AI companion apps as a named risk category for CTV and mobile buys, in the same way that made-for-advertising sites and AI-slop inventory moved from curiosities to standing exclusions, and points to verification signals such as declared developer identity, bundle-ID coherence, review-to-install plausibility and cross-store rating consistency as the fields that carry more information than the rating itself.

The tooling available to act on that has expanded, though it remains list-based. Microsoft Advertising raised website exclusion capacity to 10,000 domains per list in August 2025. Amazon DSP added content exclusion categories for brand suitability in December 2025. Google Ads enabled account-level placement exclusions spanning Performance Max, Demand Gen, YouTube and Display on January 14, 2026. Each of those controls operates on named placements, which is precisely the mechanism the study argues a 4,346-app category with 138 shared bundle prefixes defeats.

Category-level controls are meanwhile being rebuilt. DV360 is deprecating legacy digital content labels and sensitive category exclusions around August 2026, replacing them with inventory modes and content themes. Content-level in-app classification has existed since DoubleVerify expanded its mobile app brand safety product to 75 content avoidance categories in April 2024, a system that already allowed targeting and blocking by app store category, target age rating and star review. The Deepsee.io findings bear directly on the reliability of one of those inputs.

Regulatory movement adds a second reason the adjacency question is unlikely to stay static. Apple activated Texas SB 2420 age assurance requirements for new Apple Accounts in the state on June 4, 2026. Reddit moved EU teen accounts to verified age checks with chat and ad personalisation removed for 13 to 15 year olds from June 24, 2026. Vermont banned AI-only therapy services in June 2026, and Connecticut's AI bill included companion chatbot provisions among its 37 sections. Operators facing that pressure are tightening their own products, which is itself a signal about the category advertisers are buying adjacency to.

Timeline

Summary

Who: Deepsee.io, the ad quality and publisher compliance company, with the research authored by co-founder and chief executive Rocky Moss. The findings concern app developers listing on Google Play and Apple's App Store, and the media buyers, agencies and verification vendors that screen mobile and CTV app inventory.

What: An analysis of 4,346 AI character and companion apps finding that 2,589 of them, 60 percent, carry ratings classifying them as accessible to minors. The study also documents rating drift in both directions, cross-store disagreement on 106 of 306 dual-listed apps, bundle identifiers pointing to unrelated products, and an age gate on the category's flagship app that resets with a new adult birthdate.

When: The research was last updated on July 15, 2026, drawing on store data covering both platforms and referencing regulatory events from December 2024 through January 2026.

Where: Google Play and the Apple App Store globally, with implications for programmatic mobile and CTV inventory sold through app-based supply.

Why: The store age rating and the bundle identifier are two of the primary fields used to sort app inventory before a bid is placed. The research argues both are developer self-declarations in this category, that they drift with commercial incentives, and that they contradict each other across stores for the same product, which removes the basis for suitability rules built on either field.