How Accurate Is a Wi-Fi Positioning System?

A vendor says one metre. A procurement document says three. The pilot in your own building says eight, and nobody can explain the gap. All three numbers can be honest, because they are answering different questions, and only one of them is about your building.

This article separates the physics from the marketing. It covers what received signal strength can and cannot do, what time-of-flight ranging changed, what IEEE 802.11az actually adds, and how to make a supplier state an accuracy figure in a form you can test.

How accurate is a Wi-Fi positioning system?

Between roughly 1 m and 5 m horizontally, depending on method and conditions. Signal-strength fingerprinting lands in the low single-digit metres at best. Time-of-flight ranging reaches metre-level in line of sight, and degrades to around 5 m when the path is blocked.

The fingerprinting figures come from the original RADAR system, whose median error of 2.94 m for fingerprinting against 4.3 m for lateration is recorded in a May 2022 review of fingerprint-based indoor positioning. The harsher number is from DorFin (arXiv, August 2013), which found that median errors reported by prior work were consistently above 5 m once tested in realistic settings. The Fine Timing Measurement figures are from a September 2025 survey of 802.11mc, 802.11az and 802.11bk research, which puts current FTM accuracy at 1 m to 2 m, lists 0.4 m for 802.11az in line of sight, and gives below 0.1 m as the 802.11bk target.

Notice how far apart the lab number and the building number sit for the same technique. That distance is the whole subject.

Why does signal strength stop at a few metres?

Because received signal strength measures power, and power tells you about everything between the radio and the phone, not just the distance. Walls, bodies, shelving and furniture all subtract from it, and none of them announce themselves.

A received signal strength indicator reading falls off roughly with the logarithm of distance, so each extra decibel of noise buys you progressively more metres of uncertainty the further out you go. At close range a few decibels of error is centimetres. At twenty metres the same few decibels is several metres.

Fingerprinting works around the model by abandoning it. Instead of converting power to distance, you record the pattern of signal strengths at thousands of known points, then match a live reading against the map. That is why RADAR beat lateration in the same building in 2000, and why the method is still in use.

It has two costs, and both show up after go-live rather than during the pilot. The survey has to be redone when access points are moved, replaced or re-channelled, and a building whose radio environment shifts quietly produces a position that drifts quietly. The second cost is that the fingerprint describes the building as it was on survey day, including where the stock, the partitions and the people were.

What does Fine Timing Measurement change?

It stops measuring power and starts measuring time. Fine Timing Measurement, introduced in the 802.11mc work and carried into the base standard, estimates range from the round-trip time of a sequence of frame exchanges between a client and an access point.

Time is a much better proxy for distance than power, because radio waves travel at a known speed. Light covers about 30 cm per nanosecond, from the defined value of the speed of light held by NIST at 299,792,458 m/s. Because the measurement is a round trip, a timing error of one nanosecond translates to roughly 15 cm of range error. Nanosecond-scale clock behaviour is therefore the entire ballgame, and the survey above says exactly that: metre-level accuracy requires timing resolution on the order of nanoseconds.

The second limit is bandwidth. A wider channel resolves two arrival times that are closer together, which is what lets a receiver tell the direct path from an echo off a filing cabinet.

Those are arithmetic, not measurements: path-length resolution is the speed of light divided by the channel width, and 802.11 channel widths run from 20 MHz to 320 MHz per the same survey. They explain a result that otherwise looks impossible. An 80 MHz channel resolves paths no finer than about 3.8 m, yet FTM on 80 MHz reports 1 m to 2 m accuracy, because averaging many exchanges drives down random timing noise. What averaging cannot fix is a reflected path mistaken for the direct one. That error is not random, and it does not average out.

This is the mechanism behind the single most useful measured result for anyone sizing a system. A November 2025 comparison of the 802.11mc and 802.11az FTM protocols reports metre-level ranging in line of sight on 80 MHz and 160 MHz channels, and about 5 m in non line of sight. The earlier MobiCom 2018 evaluation of FTM on an open platform by Ibrahim and colleagues at Rutgers WINLAB reached the same conclusion from the other direction: metre-level accuracy is attainable mainly in low-multipath settings, or with denser deployments that allow ranging at 80 MHz and above.

What does IEEE 802.11az add?

Higher accuracy from the same round-trip idea, plus security and scale. IEEE Std 802.11az-2022 was approved on 3 December 2022 and published on 3 March 2023 as Amendment 4, “Enhancements for Positioning”.

The amendment’s own scope language is modest: it enables absolute and relative position estimation with higher accuracy than the base standard, while reducing medium utilisation and power consumption. The survey is more specific, expecting 802.11az to improve the 1 m to 2 m figure to below 1 m, with MIMO used to attack the non line of sight case, support for 160 MHz channels, secure ranging, and position-aware link adaptation. The November 2025 measurement study found 802.11az outperforming 802.11mc even in multipath-heavy environments.

Two cautions before you build a business case on it. The amendment has already been rolled into a later revision of the base standard, which is normal and means the designation you should quote in a tender is the current 802.11 revision, not the amendment number. More importantly, ranging takes two radios. An access point that supports 802.11az ranges at 802.11az accuracy only with clients that also support it, and the client population in a public building is whatever walked through the door.

What accuracy does your building actually need?

Less than the datasheet implies, for most tasks, and far more for one of them. The useful question is not how many metres of error you can tolerate in the abstract, but which physical thing the error must stay smaller than.

Read that table against the one at the top. Fingerprinting at a 5 m median is adequate for a blue dot in an open concourse and useless in a hospital corridor layout where two parallel corridors sit 4 m apart, because the error quietly puts people in the wrong one and the route recalculates against a position that was never right. Asset finding is the row that Wi-Fi struggles with, and it is the row most often promised.

Why is a real building worse than a lab?

Because a lab has line of sight, no people and a stable radio environment, and the published gap between those conditions and real ones is roughly metre-level against 5 m. Four things account for most of the difference.

Multipath comes first. Every metal surface, glass partition and lift door produces a reflection, and a reflection arrives later than the direct path, so the system over-estimates range. A body between the phone and the access point attenuates the direct path, which makes a reflection the strongest arrival. Crowds therefore degrade accuracy exactly when a venue most wants it.

Geometry comes second, and gets the least attention. Three access points in a line give you good ranges and a bad fix, because the position solution is poorly conditioned along one axis. Coverage planning optimises for throughput, which is not the same shape as optimising for positioning.

Clock calibration comes third. Round-trip timing depends on each radio reporting its own turnaround delay accurately, and the Rutgers measurements found calibration offsets large enough to matter at metre scale. Device diversity comes fourth: the same software on two handset models can produce two different error distributions.

How do you test an accuracy claim instead of believing it?

Demand a percentile, a building and a ground-truth method, then test it yourself in the worst part of your site. An accuracy figure without a distribution attached is not a figure, and there is a published standard that says so.

ISO/IEC 18305:2016, Information technology, Real time locating systems, Test and evaluation of localization and tracking systems, exists precisely to stop this argument. It covers choosing performance measures and test scenarios, and presenting the results, with a focus on indoor environments. Citing it in a tender shifts the burden from the buyer to the supplier.

Regulators phrase it the same way. Under 47 CFR § 9.10, the FCC does not require wireless carriers to deliver “50 metre accuracy”. It requires a dispatchable location or an x/y location within 50 m for a stated and rising share of 911 calls, reaching 80 per cent at the six-year benchmark, and for vertical location it specifies plus or minus 3 m for 80 per cent of calls from a z-axis-capable device, with nationwide deployment required by 3 April 2025. Percentile first, metres second. Copy the structure.

Run the acceptance test where the building is hardest: the stairwell, the loading bay, the lift lobby, the atrium at lunchtime. An average across easy floor area hides the places people get lost.

When is Wi-Fi the wrong answer?

When you need sub-metre accuracy on tagged items, or when the building has no usable Wi-Fi to begin with. Wi-Fi positioning is strongest where dense enterprise Wi-Fi already exists and the thing being located is a phone someone is carrying.

Ultra-wideband wins the asset row on physics, not on branding: it occupies far more bandwidth than a Wi-Fi channel, and the resolution table above shows what bandwidth buys. Bluetooth Low Energy earns its place where mains power is absent and a coin cell has to last years. Most real deployments end up combining sources rather than choosing one, and a supplier who tells you Wi-Fi alone solves asset tracking to the shelf is describing a product, not a building.

If you are weighing these numbers against a specific site, the fastest way to get a real answer is to look at the floor plans, the access point layout and the task the system has to support. Book a scoping call and bring your worst floor, not your best one.

Frequently asked questions

Is Wi-Fi positioning accurate enough for turn-by-turn indoor navigation?

Usually yes, with a caveat about layout. Fine Timing Measurement at 1 m to 2 m handles most corridors and concourses. Signal-strength fingerprinting at a 5 m median struggles wherever parallel routes sit closer together than the error, because the route recalculates against a position in the wrong corridor. Check the gap between your parallel circulation routes before deciding.

Does upgrading to Wi-Fi 6 or Wi-Fi 7 improve positioning by itself?

Not by itself. Wider channels help, because path-length resolution is the speed of light divided by the channel width, so 320 MHz resolves about 0.94 m against 15 m at 20 MHz. But the gain only arrives if both the access point and the client support the ranging protocol and actually negotiate the wide channel. A new access point talking to an old handset ranges at the old handset’s capability.

How many access points do you need for accurate Wi-Fi positioning?

Enough for at least three good ranges at every point you care about, arranged so they are not collinear. Count geometry, not just density. Coverage designs optimise for throughput and often leave long corridors served by access points in a straight line, which produces accurate ranges and a position that slides along the corridor axis. Re-plan for positioning rather than assuming coverage is sufficient.

Why do published accuracy figures vary so much?

Because they measure different things. A median in an empty lab with line of sight and one handset model is not comparable to a 90th percentile in an occupied building across mixed devices. The published spread, metre-level in line of sight against about 5 m in non line of sight, is the same technology measured twice. Always ask which percentile, which building and which devices.

Can Wi-Fi positioning tell which floor someone is on?

Often, but it is a separate problem from horizontal accuracy and needs separate testing. Floor determination usually combines radio data with barometric pressure. For a sense of the bar regulators set, the FCC’s vertical requirement is plus or minus 3 m for 80 per cent of qualifying 911 calls, which is tight enough to distinguish most floor-to-floor heights and loose enough to fail on mezzanines.