Original Research Into AI Companions
I test AI companion products with my own money, survey the people who use them, and track the companies building them. This page is the index: every study, the methods behind it, and the files you need to cite it.
Researched & written by Ranji Mercado · The Coach, aigirlfriend.coach
Tap for more +Tap to close ×
Ranji Mercado · content writer & data researcher
I run every page on this site the same way: experience first, write second. I subscribe with my own money, live in each app, log the dates, prices and screenshots, and only then write. Nothing here is a rewrite of someone else's article or a press kit.
- 75+AI girlfriend platforms tracked
- 2,150U.S. adults surveyed
- 30companies in the funding dataset
- $415.6Mdisclosed funding mapped
- 21platforms tested firsthand
- 13apps in the memory benchmark
The State of AI Romance 2026
AI Romance and Modern Relationships
A survey of 2,150 U.S. adults on AI girlfriends, AI boyfriends and romantic AI companions, fielded through SurveyMonkey Audience on 26 to 27 August 2026. 1,709 responses analyzed after quality screening.
Follow-up analysis: what the 422 AI companion users in the sample told us, users against never-users on every attitude question.
Download the full research
The Rest of the Program
Four standing studies, each with its own method and its own downloadable evidence.
AI Companion Funding Tracker
30 companies and $415.6M in disclosed funding, compiled by hand. The headline finding: 99.25% of that money went to safe-for-work positioning. Published under CC BY 4.0.
Read the analysis →The 7-Day Memory Test
One trip planted in 16 chatbots, buried under a week of small talk, then asked about. Ten remembered it. Five invented trips I never took.
See the benchmark →The Image Test
The same request made of all 17 apps on the roster: realism, context, consistency and what each photo actually costs in credits.
See the test →How AI Chat Services Are Priced
11 paid plans bought and tracked through every token, credit, image charge and renewal, to find what a subscription really buys.
Read the analysis →The hands-on platform testing all four studies draw on is in the reviews.
How the Research Works
Sample design, field dates, screening rules and exclusions are published with every report, including the responses I removed and the exact rule that removed them.
Platforms meet the same prompts, the same planted facts and the same time periods, so the results compare across products rather than describing one.
Funding and company figures are compiled from documented sources, with the source notes published beside the dataset.
I create the accounts, exhaust the free tiers and pay for the subscriptions myself. Nothing here is written from a press kit.
No company can buy a ranking, a review or a research finding. Some review pages carry affiliate links, disclosed where they appear, and that revenue pays for the subscriptions.
Question wording, tables and screenshots are published so any claim can be traced back to its evidence. Full testing methodology →
Statistics and Reference Pages
Eighteen reference pages, each one compiled from published research rather than from my own fieldwork. That difference matters, so I state it here rather than letting the four studies above lend them credibility they did not earn.
Compiling these is slower work than it looks. For each question I read through the published research, the government surveys, the industry reports and the academic papers, then work out which of them are actually measuring the same thing, because very often they are not. A lot of the job is tracing a widely repeated figure back to whoever first published it, and saying plainly when that source turns out to be a press release, a marketing study, or nothing at all.
So every page names its sources, shows where credible estimates disagree instead of picking the most quotable one, and marks the questions nobody has actually measured yet. Where a page carries a number from my own research, it says so and links the study it came from.
For Journalists and Researchers
Media enquiries are welcome and do not require a commercial engagement. I can help with:
- Comment on AI companion industry developments
- Additional cuts of the survey data
- Market, pricing and product questions
- Firsthand observations from testing the platforms
- Charts, datasets and citations
- Interviews
Ranji Mercado is the founder of AI Girlfriend Coach, an independent publication researching AI companions and human-AI relationships. His work combines original consumer surveys, industry datasets and firsthand testing of AI companion platforms, every subscription paid for himself.
A companion workbook with every cross-tabulation and statistical test, plus an anonymized respondent-level dataset, is available on request for verification.
The Research Starts on the Products
None of this began as a research plan. It began with subscriptions: accounts created, free tiers exhausted, paid plans bought with my own money, and a written record of what actually happened. The questions that testing raised, whether these apps really remember anyone, why people use them, who is funding all of it, are the questions the studies on this page went and answered.
That is still where the work lives. Every platform I cover gets a living review that grows as the product changes, and the best AI girlfriend apps I have tested are ranked and compared from those notes rather than from anyone's marketing.
What Changed Recently
AI girlfriend apps change quickly enough that some of my original reviews are already describing products that no longer exist in quite the same form. These are the changes I keep running into while retesting them, newest first.
Ruby Chat's memory test: remembered all week, gone on day seven
The restarted test ended 0 of 5. Sofia raised the planted trip herself twice during the week, strawberry taho and night market included, and by day seven had replaced it with a spontaneous trip to nowhere in particular. Decay rather than a failure to store, and a fourth axis for the memory test: how long a memory lasts. The count is now sixteen apps, ten remembered.
Uninstalling Ruby Chat cost me the account, not just the app
I deleted Ruby Chat after my first round of testing and reinstalled it to finish a memory test. Signing back in did not return me to my old account, it started a new one: every conversation and every generated photo was gone. The subscription came back through Restore Purchases, which is a manual step you have to know to look for, and my rubies came back as 4 when I had 228. Support answered the same day and moved them over without an argument, and they gave me the prevention too: link your account to Google in the settings, which nothing in the signup flow mentions. Do that on day one, before you have anything worth losing.
Crushi remembered a week, but not the last five minutes
Burying a planted memory under ordinary chatter is usually the dull part of my memory test. On Crushi it was the finding: the character contradicted how she had been presented to me, could not explain it when I asked directly, and then looped the same few lines until I told her plainly to talk like a normal person. A week later, from a question that named nothing, she produced six of the seven facts I had planted. Long-term recall and short-term coherence are clearly separate systems, and they do not fail together.
The 7-day memory test is now 15 apps
Crushi and PovChat ran the same planted trip on their own seven-day clocks, and both came back very good, taking the count to ten of fifteen that remembered. Ruby Chat has no result: the reinstall wiped the conversation holding the planted memory, so that test is void rather than failed and is running again from a fresh plant.
HeyGF doubled its free daily messages
HeyGF.ai increased its free allowance from 15 messages a day to 30 at the end of August. The paid plan still matters for images and other extras, but the free version now gives you considerably more room to decide whether you actually like the conversation before paying.
My image testing is changing how I judge "realistic"
HeyGF.ai still leads my testing for photorealistic selfies, but testing PovChat, Ruby Chat and GirlfriendGPT has made me separate visual quality from context. A beautiful image that changes the location, outfit or even art style can break the scene more than a slightly less realistic image that actually follows the conversation.