I don’t work for Match, Inc or Bumble, so I cannot share any internal secrets. But the apps are managed by giant publicly-traded corporations and lots of information has been intentionally or accidentally shared over the years. The missing parts we can infer via a basic understanding of modern ML and modern corporations.
Corporate incentives are the most underrated mental model: these apps cannot exist without female users and ~all of their revenue comes from men paying to compete for women’s attention. Priority one for the apps is thus keeping women on the platform. Priority two is keeping paying men on the platform. Everyone else are an afterthought downstream of those two. Every ranking decision that might seem weird starts making total sense once you keep this frame of reference in mind.
Every app first does filtering by explicitly defined user preferences: location radius, age range, gender preference, and filters like height, education, religion, drinking, smoking, kids and pets. If either of you falls outside the other’s hard constraints, you are invisible to each other.
The apps almost never ignore the explicit filters. Paid boosts can loosen some soft preferences (Hinge Boost says it may show you to people “a bit outside” your specifications) but only ever-so-slightly. The apps are all publicly nudging users to loosen distance in particular - but those nudges are explicit and won’t change anything without user consent.
What to do: review your filters honestly. Think about what’s on your profile that might be filtering you out unnecessarily to others. Remember that filters for optional fields like ‘smoking’ will also remove anyone who just didn’t fill them out in the first place. If you are in anything but the largest metros, a tight filter stack can shrink your eligible pool to hundreds or even dozens of profiles.
A great profile that has not opened the app in two weeks is bad inventory. The app can show it, but the person on the other side is less likely to get a match, a message or a date from it.
Tinder says this directly: it prioritizes potential matches who are active, especially when both are active around the same time. Hinge’s Your Turn Limits blocks new likes when a user has too many unanswered conversations, turning response behavior into a constraint on future matching. Bumble’s 24-hour match expiration is even more explicit and nudges everyone to check the app daily.
Peak activity across all three apps is Sunday 8-10pm, with Monday and Thursday evenings close behind. Friday and Saturday evenings are least active.
What to do: open the app every evening, at least briefly, with extra emphasis on Sunday 8-10pm. Boost or Spotlight in the evening, likewise prioritizing Sundays.
A user can say they want X in prompts/tags/interests and consistently swipe on Y in practice. The app algorithms have long ago learned not to take soft preferences too literally.
Explicit filters (age, height, distance, smoking, etc) don’t get overriden. as mentioned before.
Your prompt and bio text gets processed with NLP to extract topics, interests and lifestyle signals that become features for retrieval and ranking. This is how the system finds women whose stated interests and photo embeddings cluster near the profiles you have historically liked. It is also how the system retrieves you for women whose revealed preferences align.
Your swipe history is the strongest signal. If you write “looking for a serious relationship” but right-swipe exclusively on women with party photos, the ranker trusts the swipes and will prioritize party profiles for you eventually.
What to do: fill your bio and prompts with specific, distinctive content. Generic text produces a weak feature vector and the system retrieves you for a less well-targeted audience. Specific text (a named hobby, a concrete opinion, a city you lived in) produces richer features that pull you into the right retrieval buckets. Five specific > ten generic ones. Then try to swipe in a consistent manner to ensure your ML-derived embeddings aren’t constantly shifting around.
The problem the apps are solving is not “will he like her?” It is “will he like her, will she like him back and will the match turn into a successful interaction?”
Match Group patents like US9733811B2 describe mutual-selection and communication-likelihood scoring. The academic literature on reciprocal recommenders (RECON, 2010) reported improving top-10 success rate from 23% to 42% over one-sided ranking on real dating data. Every major recommender system likely uses a variant of this.
The ranker scores candidate pairs with some variant of P(A likes B) × P(B likes A) × P(useful interaction). The last term is based on data from prior matches: length of conversations, whether contacts were exchanged, whether the users unmatched each other, and We Met feedback for Hinge. Mass right-swiping hurts your calculated ranking under this formula: the ranker sees a flood of your likes going unreciprocated and adjusts predictions downward for future matches.
What to do: stay in the 10-20% right-swipe range, never any higher. Avoid swiping on profiles you won’t actually interact with later. Try to have meaningful conversations that ideally end with exchanged contact information.
The most attractive women on these apps receive thousands of likes per week. They cannot read every profile. What they actually do is look at the primary photo, make a veto decision in under a second and only tap into the profile if the photo passes.
This has nothing to do with the algorithm. No ranker optimization, no boost, no paid feature, no clever bio fixes this. If your primary photo does not clear the veto, the five perfect photos behind it won’t matter because nobody will see them. The algorithm then sees you failing to convert impressions into likes and ranks you lower.
Note that a bad photo might still be getting dozens of matches if you’re sufficiently attractive, but these matches would be less attractive than if your photo was actually a good one. Don’t get blinded by raw match counts, as they are meaningless.
What to do: get the primary photo right before anything else. See my previous article for a detailed breakdown of how to get this done. Or see my commercial photography services if you want someone else to do it for you.
The apps run computer vision and NLP on your content to make the user experience as best as possible. These are purely internal models, separate from the customer-facing “let us pick your best photo” features discussed in the next section.
Fraud and safety. Lewd content detection (Bumble open-sourced its Private Detector), stolen-image detection, underage-subject detection, catfish-pattern detection. Bumble’s Deception Detector automatically actions 95% of accounts flagged as spam or scams per Bumble’s own data (BusinessWire, Feb 2024). Photo analysis extracts a FaceVector that gets used to enforce bans across Match Inc’s apps.
Quality and attractiveness scoring. The ranker wants a prior on how your photo is likely to perform before it has swipe data. Standard quality features (face clarity, composition, eye contact, smile, solo vs group, indoor vs outdoor, framing) feed directly into the ranker as input signals. Some sort of metric for attractiveness is likely computed as well to provide an initial best-effort estimate for who should see your profile.
Taste matching. Photo embeddings let the system learn what kinds of photos each user responds to. If you consistently right-swipe women with outdoor adventure shots or in bookstores, the retrieval step finds more profiles whose photo embeddings cluster with that pattern. The same machinery works on text: bio and prompt embeddings get matched to what you have engaged with in the past.
What to do: what works for the algorithm also works for the users and vice versa. Optimize your photos using my guide and the ML algorithm will also happily rank you higher when you sign up initially.
Tinder, Hinge and Bumble all offer features that claim to pick your best photos for you. These are separate systems from the internal ranker discussed above and they share a set of problems worth understanding before trusting them.
Sparse data. Male profiles get left-swiped most of the time. Right-swipes are rare events. Any algorithm drawing on them is working with noisy, small samples.
Engagement signals are a trap. Where swipe data is sparse, these features fall back on taps, zooms and comment engagement. A photo that gets hovered or zoomed on (blurry, oddly cropped, contains text that needs reading) can score well without converting into a match.
No voter-style adjustment. A right-swipe from a woman who right-swipes 2% of the time is worth far more than one from a woman who right-swipes 30% of the time. These features count them equally.
Self-comparison, not market comparison. These features tell you which of your photos beats your other photos. They do not tell you how any of them stack up against the competition in your city.
AI-based solutions are even worse. Tinder’s Photo Insights model claims to be able to pick the ‘best’ photos from your camera roll, but these models inevitably have small token budgets and apply general photo-quality heuristics, so they’re only good for eliminating the worst candidates.
What to do: Treat these recommendations as one data point and never rely on them to pick the primary photo. Pick the primary based on human feedback from women in your target age range, which is easiest done via Photofeeler. Smart Photos can then reorder the remaining pics if you’re not sure about the best order to use them in.
Tinder confirmed an Elo-style desirability score to Fast Company in 2016, then said Elo was “old news” in 2019. Hinge CEO Justin McLeod told Fortune in January 2024 “we don’t really have an attractiveness score.” Bumble has never confirmed its folk-named “Beehive Score.” But that just means that the apps no longer rely on a single chess-like “hotness” number to rank users.
When the apps say they “no longer have an attractiveness score”, it just means that instead of a single scalar, the ranker’s learned feature is a multi-dimensional vector incorporating the rate of incoming likes, the response rate to your messages, the mutual-match rate… and any other popularity signals crammed into the model by ML engineers.
The apps’ top priority is keeping ensuring that their female userbase is happy. The ranker tries to show women the best men it can find, as judged by a combination of her taste cluster and the men’s ranker-implied desirability relative to her own. In other words, the good old ELO is still alive under a fancier name.
As of today here are the products offered by the three apps:
Tinder Boost puts you among the top profiles nearby for 30 minutes. Super Boost extends the window at significantly higher cost.
Tinder Super Like attaches a label the recipient sees and claims 3x match rate.
Tinder Priority Likes and Skip The Line (on Platinum and Select plans) pin you at the top of the recipient’s queue.
Hinge Roses move to the top of the recipient’s Likes You screen and are the mechanism for liking Standouts profiles.
HingeX Priority Likes keep your like near the top of the recipient’s list (Hinge Subscriber Experience).
Bumble Spotlight pushes your profile ahead for 30 minutes.
Bumble SuperSwipe moves you to the front of the recipient’s queue with a visible tag.
Bumble Premium+ says it makes likes be seen sooner and prioritizes the subscriber in feeds (Bumble paid features).
Before you buy anything, remember an important principle: Boosts/Superlikes amplify what you have. A weak profile seen by 10x more women is still a weak profile. Do this after optimizing your photos and prompts, not as a standalone crutch.
New accounts get elevated visibility for roughly 24-48 hours. The apps need this window because they have no behavioral history on you and need a few hundred swipe outcomes to calibrate where you belong in the ranker output.
What to do: set up your profile and then don’t touch it within the first 48 hours to wait for the algorithm to calibrate. Likewise always let at least 48 hours pass after making any profile changes before deciding whether to keep them.
New profiles are your best shot at high-quality matches for two reasons. First, new profiles have not yet been liked by every man in the city, so the competition for attention is lower. Second, the user is fresh to the app and hasn’t yet been burned out or started talking to twenty other candidates.
But the apps cannot just show you only new profiles. New female profiles are already flooded with likes as it is. The rational design is to throttle exposure to new women, spreading their early impressions across enough men to calibrate the ranker.
This means your default stack is mostly populated with women who have been on the app for a while. The ranker is trying to serve you its best guess at who you will match with, weighted toward women the system has decent data on, not weighted toward the newest signups.
What to do: always try to swipe till the end of your stack, so that the algorithm is forced to show you new profiles more often. If not possible due to the size of your metro area, consider making your filters more strict until the remaining stack is of a reasonable size.
Big week-over-week drops with no profile change can be any of:
You ran out of women in your area who pass your hard filters. In anything smaller than a major metro, this is the single most common cause. The apps cycle through your eligible pool and when the pool gets thin, your stack fills with repeats or more distant profiles and your new-match rate collapses. Widen distance by 10-20 miles and see what happens. If distance is capped by geography, lower your minimum age or relax a dealbreaker you do not actually care about.
The ranker re-scored you after a profile change. Changing your primary photo or rewriting your bio resets part of the system’s prior. Give it at least 48 hours before concluding the new version is worse.
Negative signal accumulated. Rapid mass-right-swipes, a few reports, a failed photo-verification check, unmatching too many women in a short period, or ghosted conversations will all pull you down the ranker.
Shadowban. Covered in the next section. Usually appears as near-zero matches plus Boost not working.
The app kicks you off. Login fails or a block screen appears. The apps are trying to be aggressive on this for reports, fake profiles, spam behavior, off-platform promotion and terms-of-service violations.
The economics favor aggressive banning: the cost of losing one legitimate user is much lower than the cost of one bad actor staying and degrading everyone else’s experience. Appeals rarely succeed. Most banned users reset or give up.
Hinge’s spokesperson confirmed to Gizmodo that the company sometimes handles bad actors by preventing other users from searching for them, seeing them in the feed or receiving their messages. Neither Tinder nor Bumble has confirmed shadowbanning explicitly but community evidence is consistent across all three: sudden collapse to near-zero matches, frozen Likes You queue, messages that don’t deliver, Boosts that produce nothing.
Why shadow bans can be better: a hard ban makes the bad actor create a new account (expensive to detect), while a shadowban lets them keep paying, keeps them from knowing they’ve been throttled and avoids the appeal volume a hard ban generates. Nothing personal, just business.
Typical triggers, compiled from community reports and the apps’ own disclosures:
Multiple reports from other users in a short window. One or two rarely triggers anything. Three to five often does.
Failed or suspicious photo verification. Adding a photo that doesn’t match your previously verified face is one of the fastest paths to demotion.
Rapid location changes that look like spoofing.
Mass right-swiping. Above roughly 80% right-swipe rate, the anti-bot classifier will almost certainly you.
Unmatching too many women too fast. The apps read repeated unmatches as the behavior of a bot or of a man using the app for something other than dating (spam, business, catfishing). If you are going to unmatch, do it gradually, not in a single cleanup session.
Copy-paste openers across many matches.
Profile-edit thrashing. Changing photos and bio many times per day.
Before concluding you’re shadowbanned, rule out the more common explanation: you’re just ranked low or your eligible pool is exhausted. A true shadowban looks different: near-zero matches that don’t recover across multiple Boosts, messages that fail to deliver to matches, Likes You queue frozen for weeks, Boost spend producing no uplift at all. Low ranking can recover on its own over time, but a shadowban is unlikely to ever do on its own.
Can be a good practice for a major profile overhaul. You do not need a new phone number, device, email or any of the fingerprint-breaking stuff below. The reset should give you a fresh set of embeddings and clear out your list of sent likes. Worst case scenario you’ll be back to where you were before the reset happened.
The apps store identity signals specifically to block banned users from coming back and retain that data long after you delete. What they might be tracking:
Phone number
Email address (Gmail aliases with “+” or dots get caught)
Apple ID / Google account ID
Device fingerprint (hardware model, OS, app version, screen, timezone, language, fonts, carrier - probably good enough to overcome the Apple tracking opt out)
IP address / network
Payment method (card hash, Apple Pay ID, Google Pay ID)
Perceptual photo hashes (so light edits still collide with original pics)
Social graph (linked Instagram, Spotify, Facebook, imported contacts)
FaceVector from previously uploaded pics, aka a facial recognition vector to tag any future photos of the same person
The apps almost certainly also run face recognition on every uploaded profile photo, not only verification selfies. If they do, no combination of new phone, new device and new IP will save you while your real face is in the photos. There’s no public confirmation of this, but it’s the obvious extension of the existing FaceVector system. Given that many people do successfully reset their profile a shadowban, I suspect that a full facial-recognition based ban is only reserved for the worst offenders.
Hinge retains identifiable data for roughly 90 days after deletion, other apps are likely to be in the same ballpark. If nothing from the list below worked, wait for 90 days and then try again.
New non-VOIP phone number. VoIP services (Google Voice, Hushed, Burner) are defacto banned. Instead use a $5/month eSIM from Tello, which runs on T-Mobile’s network. Cheapest option as of 2026 for a proper fresh number.
New email.
New Apple ID / Google account.
Different physical device. Factory reset of the old phone might not be sufficient.
Different IP at signup. Use your cell plan’s data instead of your home Wi-fi.
New payment method.
New photos. Not edits of the old ones, as image comparison algorithms are extremely good these days and very cheap to run.
Don’t link your Spotify and don’t import your contacts.
The three apps converge on the same underlying machinery, so most of what works is common across them. App-specific sections below cover only what the UI or official product descriptions actually differ on.
Profile basics:
Aim for 6 photos on all three apps. Tinder allows up to 9, Hinge requires 6, Bumble allows up to 6. Six is best for all three.
Get the primary photo right before anything else. See the photos section above. It’s worth more than the rest of your profile combined.
Include one full-body shot and at least one action photo (you doing something concrete). Action photos give conversational hooks and readable interest features to the ranker.
Write specific, distinctive bio and prompt content. “I love travel” is too generic. “Lived on four continents, best meal was a street cart in Bangkok” helps with your embeddings and is a much better conversation starter.
Verify. Tinder claimed 67% more matches in pilot data. Hinge claims verified users go on 200%+ more dates. There’s no downside to doing it.
Get a paid account. Corporations don’t care about male users on free accounts. Pay for at least the cheapest paid tier account.
Behavior:
Open the app briefly every evening, especially on Sunday 8-10pm. You can skip Fridays and Saturdays to take a break.
Right-swipe in the 10-20% range. Higher than that and you will be harming your calculated ranking score. Much higher and you’re risking a shadowban.
Respond to matches quickly. Stale conversations are a negative ranker signal everywhere.
Exchange contact info inside the app where natural. The apps can tell this happened and read it as “match converted to real conversation.” That’s a positive ranker signal and your future exposure improves.
Don’t unmatch lots of profiles in one day. A few at a time is fine. Ten in a row looks like bot behavior and is a known shadowban trigger.
Don’t edit your profile multiple times per day. Once every few days max.
Paid features:
Boost / Spotlight only in the evenings. Preferentially pick Sunday.
Use SuperLikes sparingly. Send them to genuinely outstanding profiles, don’t spam them blindly.
Don’t boost a weak profile. Fix photos first. If a boosted weak profile gets seen by 10x more women, it also gets 10x more left swipes, which will further erode your profile’s ranking.
Lock your primary photo manually. Don’t let Smart Photos pick it. Smart Photos works from sparse, biased, attribution-problematic swipe data and often surfaces the wrong primary. It can still reorder the rest if you leave it on.
Photo Selector is pure on-device AI with no swipe data. Use it for a first pass through your camera roll, not as final authority.
Comment when you like. Hinge’s own data says commented likes have roughly 47% higher date conversion than bare photo likes. It’s the strongest free lever on any of the three apps.
Prefer liking text prompts over photos. The Hinge UI shows the recipient exactly what you liked in the Likes You feed. A text prompt plus your comment stands out visually more than a tap on a photo, reads as more intentional and provides something concrete to respond to.
Your Turn Limits will throttle you if you sit on unanswered conversations. Don’t accumulate inbox debt.
Use the ‘We Met’ feature. It almost certainly helps your ranking when both sides report a positive encounter.
Answer Opening Moves quickly. The ML algorithm likely penalizes those who wait till the last minute.
Don’t let matches expire. A corollary to the previous rule: this will likely demote your profile.