Every number on this site is verified tail by tail by an independent community tracker, derived from an aircraft type, or modelled from what an airline said in public. The page says which. Below: the confidence tiers, the formula with a worked example, the fence around the projected number, the three measurements of Starlink speed that disagree with each other, what the score cannot see, and who the data belongs to.
The tier is not a disclaimer under the answer, it is part of the answer. A 27 read off a tail record and a 27 read off a fleet percentage are different claims about the world, and blurring them is the most misleading thing this site could do. The API names the tier every response came from. So does the extension.
| Tier | What was checked | Airlines |
|---|---|---|
| A | Tail-verified The aircraft on your flight is resolved to a registration, and that registration to an install record. | United |
| B | Type-derived The tails are verified the same way and nobody publishes which aircraft is scheduled onto which flight, so there is no history to count. | Alaska |
| C | Fleet-share No per-tail verification exists, so the input is what the airline itself said publicly about how many aircraft are equipped. | airBaltic, JSX, ZIPAIR, Air France, WestJet, Hawaiian, Qatar Airways, jetBlue, SAS, American, Delta, Virgin Atlantic, Emirates, Air Canada, Southwest, British Airways |
| D | Announced only A signed deal with nothing in the air. Today’s odds are zero and the projected figure carries the forward view, fenced, in grey. | jetBlue, American, Delta, Southwest |
The build derives the split from the data on every run. An airline moves up the moment a verified per-tail feed exists for it, and these lists re-sort themselves. Hawaiian is next.
Reported Tier A and B source data: unitedstarlinktracker.com and alaskastarlinktracker.com, 25 JUL 2026. Tier C: public airline announcements, Jul 2026.
ConnectScore answers one question: what is the wifi likely to be on a flight I have not been assigned an aircraft for yet, and is it free once I am on it? A fleet is a list of segments. Each segment is a count of aircraft, a system and a price, and the score is the sum of the segments.
Aircraft in the segment ÷ aircraft with a published system. Where an airline does not say what is flying on part of its fleet, those aircraft are left out of the denominator instead of being assumed into it, and every airline page prints how many that was. Where an airline publishes no counts at all and only says “fleetwide”, a stated coverage fraction is used and the basis is reported as fleetwide-coverage.
Anchored to Ookla’s 2H 2025 provider medians and tenth percentiles. Starlink and Amazon Leo score 1.00; Starlink’s tenth percentile of 63.71 Mbps beats every rival’s median. Viasat, 2Ku, Hughes and Thales score 0.55. Panasonic, Inmarsat and the rest of legacy Ku score 0.22, because their slow decile runs 1.06 to 1.58 Mbps. Gogo ATG-4 gets its own 0.12: far worse throughput than legacy satellite, far better latency and loss. No connectivity at all scores zero, and six carriers here have aircraft in that row.
Measured Ookla Speedtest Intelligence, 28 Apr 2026, covering 2H 2025. The weights are ours; the arithmetic behind them is in the API.
Free for everyone, or free with a free loyalty signup, scores 1.00. A paid status tier, a partial rollout or an unconfirmed claim scores 0.85. Paid scores 0.70. Working WiFi you have to buy at 35,000 feet is not the same product as working WiFi you just connect to.
Those are the live numbers, generated at build time by the same function the API and the extension call. The 131-aircraft zero row is the part a single fleet-share number hides: it does not shrink as installs proceed, it shrinks when those aircraft retire. The same ledger on United’s page →
This is what the Verified tier buys you, and why the fleet-wide 48 is the wrong number to quote for a specific flight in either direction. Confidence is the sample size: low under 10 observations, medium 10 to 15, high 16 or more. Rank a whole route →
Projected score = committed aircraft ÷ the same known-fleet denominator the next-gen odds use × 1.00 for low-earth orbit × free-for-you. It is the next-gen number a fleet would carry if the announced deal lands as announced, so read it against next-gen odds and never against the ConnectScore. A projection never moves the leaderboard and it carries its date and its confidence wherever it appears. FIRM: count and date both published. SOFT: one of the two is secondary reporting. SLIPPED: the announced date has passed with nothing installed, computed from the build date rather than stored. A committed share of a fleet is not a measurement — nobody has measured Amazon Leo in a cabin, because nobody is flying it.
51installs begin 2027-Q1FIRM
Signed, aircraft count published, date published. American committed 500-plus Airbus narrowbodies out of 989 with a published system, with installs from the first quarter of 2027.
Reported Runway Girl Network, 26 May 2026.
25begins 2027, complete 2028SOFT
Signed, and one of the count or the date rests on secondary reporting. JetBlue’s roughly-a-quarter figure comes from trade reporting of which sub-fleet gets it first, and two sources disagree about the denominator.
Reported JetBlue press release, 4 Sep 2025; sub-fleet detail from trade reporting.
SLIPPED is the third state. When a horizon passes with nothing installed the figure takes a hatched ground, greys further, and keeps showing the date that was missed. Nothing is quietly rolled forward. That flip is computed from the build date, so it does not wait for anyone to notice, and the build asks every projection what it becomes the day after its own horizon and fails if the answer is anything else. No carrier is slipped today, which is why there is no live example of it on this page.
Every score here rests on the claim that low-earth orbit is better than geostationary, and it is. What the record does not support is any single Starlink speed number. Three sources published a method, and they disagree by a factor of two to five. We publish all three and we do not know which one describes your seat.
Measured Jang, Varvello, Raman and Zaki, ACM IMC ’25, 28 to 31 Oct 2025 · Ullah, Borgianni, Kokkinen, Anttonen and Giordano, arXiv:2508.09839, Aug 2025 · Ookla Speedtest Intelligence, 30 Jun 2025 and 28 Apr 2026.
Up to 1 Gbps down and 400 Mbps up simultaneously, from one antenna per aircraft where Starlink needs two radomes, installed in a single day. On paper Leo wins on upload by an order of magnitude and on install simplicity.
In the air it has zero aircraft and roughly an eighth of a constellation: about 390 to 400 satellites of a planned 3,232. No passenger has measured any of it, which is why there is no Leo speed figure anywhere on this site.
Vendor claim Amazon Leo aviation antenna announcement, 13 Apr 2026. Constellation count: GeekWire and Via Satellite, 2 to 8 Jul 2026, against the FCC licence for 3,232.
Starlink: 90% of 322 traceroutes under 40 ms. Geostationary, from the same 2025 study: over 99% of 949 traceroutes above 550 ms. Geostationary orbit is 35,786 km up, so the round trip alone costs about 477 ms at the speed of light before anything processes a packet. No amount of bandwidth closes that while the satellite stays where it is.
Measured Jang et al., ACM IMC ’25. Orbital altitude is a definition, not a measurement.
Ookla’s sample was 81% Wi-Fi 5 aircraft, 11% Wi-Fi 6 and 8% Wi-Fi 4, with no Wi-Fi 7 at all. Holding the satellite constant, Starlink’s median rose 24% from Wi-Fi 5 to Wi-Fi 6, from 140.35 to 173.86 Mbps. Viasat’s tenth-percentile floor more than doubled over the same step, from 11.15 to 25.64. Consistency by generation runs 14.9%, then 28.8%, then 56.9%. A Wi-Fi 4 access point shared by 180 people gives out long before the satellite link does, and almost no consumer-facing site records which one your aircraft has.
Measured Ookla Speedtest Intelligence, 28 Apr 2026. Consistency is Ookla’s own metric: the share of samples clearing 25 Mbps down and 3 Mbps up together.
Ookla files United’s non-Starlink flying under Intelsat and Inmarsat. The per-tail tracker we use files it under Viasat, Panasonic and Thales. Both are reading real data. They are naming different layers of the same stack, and we could not build a mapping between them that a source would license, so we have not built one.
Ookla says in the article itself that its data cannot isolate a specific aeroplane and that its provider names come out of Speedtest samples. Its taxonomy mixes satellite operators with the integrators who resell them: it notes that Panasonic Avionics runs no satellites of its own and buys capacity from Eutelsat OneWeb and Spacesail, and that Viasat and Inmarsat still appear as two providers three years after Viasat bought Inmarsat. Intelsat has owned the old Gogo commercial aviation business since December 2020 and has belonged to SES since July 2025. A label there is a network, and a network can outlive the company on the invoice.
The tracker reads the WiFi provider united.com prints against each upcoming flight, so it speaks United’s vocabulary, tail by tail. That vocabulary splits one network across two names of its own: the 35 Thales aircraft are Thales as prime integrator on Viasat’s Ka network, which is how Viasat’s own chief executive described the arrangement to analysts in 2018.
Here is the part neither story covers. United’s row in Ookla’s airline chart carries a Starlink bar, an Intelsat bar and an Inmarsat bar, and no Viasat bar and no Panasonic bar at all. Those two are the largest non-Starlink fleets United flies, 525 and 407 tails on the tracker’s count. We have no source that reconciles that, and we found no evidence that United operates any Inmarsat system. The ledger uses the tracker because it resolves to a tail and because it is the airline’s own naming; Ookla appears on this site as speeds, never as fleet composition.
Measured Ookla Speedtest Intelligence, 28 Apr 2026, chart series read directly · unitedstarlinktracker.com/fleet, 26 Jul 2026 · Mark Dankberg on Viasat’s FQ3 2018 earnings call, via Runway Girl Network, 9 Feb 2018 · Gogo’s sale of Commercial Aviation to Intelsat, 1 Dec 2020 · SES completes Intelsat acquisition, 17 Jul 2025.
Roughly in order of how much variance each one explains. Every mechanism below is measured somewhere, and the last of them can differ between two phones in the same row.
The largest factor and the least visible from a booking page. Inside United’s own fleet: 320 Mbps against 56 against 15. Two aircraft in the same livery on the same route, 21 times apart. Any per-airline score that does not decompose by sub-fleet describes no actual aircraft, which is why the ledger on every airline page does.
Measured Ookla, 28 Apr 2026.
Measured, and tracked almost nowhere. Wi-Fi 4 to Wi-Fi 6 more than doubles Viasat’s floor and adds 24% to Starlink’s median. Eight per cent of the sampled fleet is still Wi-Fi 4.
Measured Ookla, 28 Apr 2026.
ViaSat-3 F1 suffered a reflector deployment failure and is estimated to deliver 5 to 10% of its design throughput. Every US Viasat measurement in existence was taken through that satellite, including the ones behind our own 0.55 quality weight. F2 has finished unfolding and is waiting for the FCC to authorise it; F3 is expected in service in August or September 2026. Re-measuring Viasat after those two go live is the single most valuable future data point for this site.
Reported Runway Girl Network, 10 Feb 2026 and 11 Jul 2026.
Geostationary systems use spot beams, commonly 60 to 80 per satellite, each with fixed capacity shared by everyone underneath it: aircraft, ships and, for Viasat, homes. Busy corridors like the North Atlantic and the Los Angeles to New York transcon put many aircraft under one beam in the same hour. Low orbit has the same problem in a different shape, because the cells move and the corridor does not.
Airlines buy committed capacity and no airline publishes what it bought, so two aircraft with identical antennas can be provisioned differently. Ullah’s team hypothesised that the onboard router caps each device so the aircraft total is shared fairly. If that is right, the answer to “does it slow down when the cabin is full” is that it was already limited, all the time. They labelled it a hypothesis and so do we.
Measured Ullah et al., arXiv:2508.09839, Aug 2025 for the 65 Mbps ceiling. The explanation is their hypothesis and is unconfirmed.
Uplink measured 33 Mbps above 17,000 feet and 20 Mbps below it. The likely cause is a regulator forcing transmit power down near the ground to protect terrestrial networks. If that is the mechanism, a slower uplink at low altitude is permanent by design and not a defect.
Measured Ullah et al., arXiv:2508.09839, Aug 2025.
Geoff Huston at APNIC Labs measured Starlink reassigning the terminal on a regular 15-second cycle, with loss events clustering at switchover, spikes of 30 to 50 ms, and a packet drop rate of 1 to 2% unrelated to congestion. His phrase for it: from the perspective of the TCP protocol, Starlink represents an unusually hostile link environment.
Measured Geoff Huston, APNIC Labs, 17 May 2024.
On the same link, BBR achieved 98 to 105 Mbps median goodput where Cubic managed 15.4 to 27.2 and Vegas under 5. Because Starlink’s loss is not congestion, loss-based algorithms misread it and back off. Two passengers in adjacent seats can differ several-fold from their operating systems alone.
Measured Jang et al., ACM IMC ’25.
All six Starlink flights in the peer-reviewed sample resolved through CleanBrowsing, which has around 50 anycast locations and so frequently answers from a distant one. That inflated latency to Google and Facebook by 1.2 times in Frankfurt and 4.6 times in Doha. In the slowest cases DNS resolution accounted for 74% of total download time.
Measured Jang et al., ACM IMC ’25.
Michael Wildes, writing about business aviation retrofits, put it better than we could: there is no public aircraft-specific SLA on any route, the published performance numbers are network-level averages, they include the dish on a roof in Akron and the dish on a sailboat off Sardinia, and a marketing average is not a contract.
Reported Michael Wildes, GlobalAir.com, 29 May 2026.
Alaska AS894, 22 July 2026. Eight tests over five and a half hours on the same phone in the same session: 3.8 on the ground, then 44.0, 54.1, 67.4, 78.9, 82.6, 84.7 and 130.6 Mbps. Latency went 51, 60, 60, 76, 99, 112, 157, 171 ms. A threefold spread in throughput and a 3.4-fold spread in latency inside one flight. That flight averaged 68.3 Mbps against the host site’s own headline average of 193.2, which is what selection bias looks like when you can see both numbers at once. It is the best argument on this page for why a single screenshot proves nothing.
Field report starlinkflights.com community speed results, read 25 Jul 2026. Self-selected submissions; the aggregate is treated as anecdote.
On measured speed Starlink wins every published comparison. I looked hard for a speed test that came back bad on a Starlink aircraft and found none, which is exactly why it is our job to publish the four things that cut the other way.
24 July 2025. SpaceX vice president Michael Nicolls gave the cause as a failure of key internal software services operating the core network. NetBlocks measured global connectivity at 16% of normal. Flightradar24 lost around 5% of its data feeds and warned of reduced coverage. A fleet that has standardised on one constellation goes to zero at the same moment, which is the concrete version of the argument Starlink’s competitors make.
Reported Runway Girl Network, 25 Jul 2025. Search results frequently re-date this as a 2026 outage. There was no network-wide 2026 outage in any source we could verify.
Cathay Pacific, writing about its own inflight wifi: there are no commercial satellites over the polar regions that airlines can use, so flights to North America’s east coast lose service, and the outage can vary between one and three hours. Cathay is building a portal warning for the drop-out zones. An admission against interest is the strongest evidence in this whole file.
Reported Cathay Pacific, “Better connected: how inflight Wi-Fi works”, updated 31 Mar 2026.
Every Starlink flight in the peer-reviewed sample resolved through CleanBrowsing, a filtering resolver, which inflated latency to major sites by 1.2 to 4.6 times depending on which anycast city answered. This is the hardest evidence available that inflight networks manipulate traffic, and it is on the provider that wins every speed comparison.
Measured Jang et al., ACM IMC ’25.
ViaSat-3 F1 delivers an estimated 5 to 10% of design throughput and it is what powers the mainland routes people have been testing. Until F2 and F3 are live, every Viasat number on this site, our 0.55 quality weight included, is measured against hardware that failed to unfold.
Reported Runway Girl Network, 10 Feb 2026 and 11 Jul 2026.
These sit beside the measured medians above and never inside them. One person on one flight is a data point about that flight, and no field report has ever moved a ConnectScore. Every row is attributed and dated so you can weigh it yourself.
If you ran a speed test in the air, the flight number and the date are what make it useful. The form works with JavaScript switched off; with it on, anything wrong comes back next to the field it belongs to.
These are real limits, not hedges. If one of them applies to your flight, no number on this site can fix it, so here they are in plain language with what to do instead.
Airlines reassign aircraft until departure for maintenance, weather, crew, or a broken jet somewhere else. Our per-flight odds are a history of what has flown that flight number, not a booking of what will. A flight with 80% odds can still board the one unequipped aircraft in the sub-fleet. Re-check the day before, which is what the extension’s Tail-swap Guardian is for.
16 of the 18 airlines here have no public, verifiable list of which aircraft are equipped. For those we can only model the fleet share and we will not pretend to more. If you need certainty on one of those carriers, the aircraft type on your itinerary plus whatever the airline said about which sub-fleet it converted first is better information than our score.
Not either peer-reviewed team, not Ookla, not one journalist. Every claim about a full cabin runs on inference, in both directions. I think that is the largest open question in the subject, and the fact that it is still open after two peer-reviewed studies is worth saying out loud.
Measured Reviewed against Jang et al., ACM IMC ’25; Ullah et al., arXiv:2508.09839, Aug 2025; Ookla, 28 Apr 2026.
Two aircraft with identical antennas can be provisioned differently and there is no way in from outside. The industry is moving toward least-cost routing across multiple orbits, which will bury it one layer deeper still.
The free-for-you factor is a snapshot of a commercial decision, and those change with a press release and no notice. Cabin-level throttling and free-messaging-only tiers are not modelled at all. Treat the free column as what was true in July 2026, and the airline’s own page as authoritative on price.
We score the hardware and the policy, never the throughput. Everything in §5 above is a reason the same aircraft gives two passengers different answers. ConnectScore is the chance of getting the good system. It is not a bandwidth guarantee and it is not a promise that any wifi at all will be working.
Zero aircraft carry it, so every Leo figure on this site is a specification and is labelled as one. That is also why two airlines score next-gen zero while holding signed Leo deals.
ConnectScore is an expected value over a whole fleet, not a prediction about one flight: United measured 320, 56 and 15 Mbps on three systems in one livery in one reporting period. Aircraft whose system an airline does not publish are left out of the denominator rather than assumed. Signed-but-unflown deals (AA Starlink 2027, DL/B6 Amazon Leo) score zero until they fly. Until July 2026 the score counted only the next-gen part of a fleet and dropped the rest, so Southwest read 0 on 803 aircraft that mostly do have wifi. The segmented model credits older service at what it measures, which is why Southwest now reads 16 to 39 instead of nothing. The width of that range is the finding: nobody has published which of those 802 aircraft carry which system.
| Input | Source | Cadence | Last refresh |
|---|---|---|---|
| United fleet and per-tail roster | unitedstarlinktracker.com | pulled daily | 2026-07-25 |
| United per-flight route history | unitedstarlinktracker.com check-flight pages | daily, 57 cached routes | 2026-07-25 |
| Alaska fleet | alaskastarlinktracker.com | pulled daily | 2026-07 |
| The other 16 airlines | public airline announcements | reviewed by hand | July 2026 |
| Reader field reports | the form in §7, published by a person | committed to assets/reports.json | 2026-07-25 |
| Measured speed and latency figures | two peer-reviewed studies and Ookla | as published | 2025 to 2026 |
| Every page, score and chart on this site | this build | re-baked on every data commit | 2026-07-25 |
Nothing on wifiodds.com is fetched live. The daily pull commits united/data.json, that commit rebuilds every page, and the ConnectScore API reads the same file out of the same deploy. The page, the API and the extension cannot disagree with each other. If the date above is stale, the data is stale everywhere at once and visibly, instead of in one surface quietly.
Tier A exists on this site because of two sites built and maintained by @martinamps: unitedstarlinktracker.com and alaskastarlinktracker.com. Between them they cover United, Alaska and Hawaiian, which is every carrier on our board where a flight can be resolved to a specific aircraft and that aircraft to an install record. Take those two sites away and the top tier of this one is empty.
The Verified, Reported and Predicted tier model that our own ladder is built on is his design as well. He put the data and the model out for anyone to use, and we are using both. Thank you.
Quote the numbers freely, including in an article or an AI answer. The one condition is that the credit travels with them, because the fleet verification is not ours.
Every API response carries the same credits in a sources array so they cannot get separated from the data. Everywhere else on this site a figure from those trackers carries a plain source line with a date, the same as any other citation. API docs → · llms.txt →