Predictive Wi-Fi Surveys in 2026: What the Model Gets Right — and Wrong

none

Originally published March 2020 — rewritten and updated July 2026.

Back in 2020 I wrote about the fundamentals of predictive Wi-Fi analysis, back when we were all arguing about whether a heatmap on a screen counted as engineering. Six years, one citywide CBRS build, and a whole lot of festival stages later, the tools have gotten dramatically better — Ekahau AI Pro will practically design the network for you, and Hamina lives in the cloud and detects wall materials automatically. But here's the thing that hasn't changed: the model is only as honest as the person feeding it. So let's talk about what a predictive survey actually gets right in 2026, what it gets wrong, and what it flat-out cannot know.

What is a predictive Wi-Fi survey, actually?

A predictive survey (some folks say "desktop survey" or "virtual design") is RF math run against a floor plan. You import a drawing, trace the walls, assign each one an attenuation value in dB, drop virtual APs with real antenna patterns, and the software calculates propagation — signal strength, SNR, secondary coverage, channel overlap — for every point on the map. No plane ticket, no ladder, no badge escort through a warehouse at 2 a.m.

Modern tools have gotten legitimately good at this. Ekahau's AI Pro auto-places APs and optimizes channel plans. Hamina's automatic wall material detection reads your floor plan and takes a swing at identifying drywall versus concrete so you're not tracing 400 walls by hand. Both model 2.4, 5, and 6 GHz, and Hamina anchors wall attenuation at 5 GHz and curves it across the other bands using ITU material data. That's real progress from the days when we treated every band identically and hoped.

But a predictive survey is a hypothesis, not a measurement. Keep that sentence taped to your monitor.

What is the model actually good at?

Plenty. I don't want this to read as a hit piece on predictive design, because I start every single project with one — including the McAllen build, where we put 1,000 APs and 24 CBRS base stations across a city in 60 days. You do not pull that off by walking around with an AP on a pole first. You pull it off because the model let us answer the big questions fast:

• AP count and budget. The model gets you within striking distance of the real number, which is what procurement needs before anyone signs anything.
• Relative coverage. If the model says the northeast corner is weak, the northeast corner is probably weak. The absolute dBm might be off; the shape of the problem usually isn't.
• Channel architecture. Reuse patterns, channel width tradeoffs, co-channel contention — this is math, and computers are good at math.
• Antenna selection. Modeling a real directional pattern down a warehouse aisle or across a festival bowl tells you fast whether omnis are going to embarrass you.
• Iteration. Trying five designs on-site takes a week. Trying fifty in software takes an afternoon.

Which inputs make or break the prediction?

Wall attenuation values. This is the big one. The delta between "drywall" (~3 dB) and "concrete" (12–30+ dB depending on thickness and rebar) is the delta between a working design and a very expensive redesign. Auto-detection helps, but floor plans lie — that "drywall" partition might be old plaster on metal lath, which eats RF like brisket at a family reunion. Hamina's own docs tell you to measure your 3–5 most common wall types on-site: put an AP a few meters from the wall, measure both sides, subtract, average. Fifteen minutes of measurement beats a thousand assumptions.

Capacity assumptions. Coverage design died years ago; everything interesting is capacity design now. If you tell the model "40 users per AP, light browsing" and reality is 300 people livestreaming a headliner in 4K, your heatmap was a work of fiction. Airtime is the currency. Model the busy hour, not the average.

Client mix. The model defaults to a reasonably modern client. Your site has handheld scanners with one antenna and a 2016 chipset, VoIP badges that roam badly, and IoT sensors that only speak 2.4 GHz. Design for your worst important client — its receive sensitivity, its transmit power, its band support — or that client will design your support queue for you.

Ceiling height and mounting reality. The plan says 3 meters; the venue has a 9-meter deck and the AP is going on a truss next to a moving-head light and a fog machine. Ask me how I know.

What does the model get right, get wrong, and simply not know?

Gets right	Gets wrong (without good inputs)	Can't know at all
Approximate AP count and placement zones	Absolute RSSI when wall dB values are guessed	The wall that isn't on the floor plan
Relative coverage — where the weak spots are	Capacity, if user counts and app mix are wishful	Neighboring networks and non-Wi-Fi interference (radar, cordless gear, that one microwave)
Channel plans and co-channel overlap math	Attenuation of "surprise" materials: plaster-and-lath, low-E glass, aquariums, stacked inventory	How your actual clients roam — sticky clients don't read heatmaps
Antenna pattern behavior and downtilt	Multipath in metal-heavy spaces (warehouses, stadium bowls)	Crowd attenuation — 20,000 bags of salt water between AP and client
Band-by-band propagation differences (2.4/5/6 GHz)	Mounting constraints — where the AP can physically go vs. where you drew it	DFS hits, future construction, the forklift fleet, next year's client devices

When is predictive alone enough — and when do you have to validate on-site?

Predictive alone is defensible when: it's standard commercial construction (offices, schools, retail) with materials you've measured before; coverage requirements are ordinary data connectivity; the cost of being 10% wrong is "add two APs later"; or it's a cookie-cutter deployment where you've already validated the first three sites of the same prototype.

You validate on-site — with an AP-on-a-stick survey before deployment, a validation walk after — when: life-safety or business-critical apps ride the network (voice, healthcare, RTLS, warehouse automation); the construction is exotic or unknown (old buildings, metal buildings, cold storage, anything with "historic" in the name); it's high-density (stadiums, festivals, conference halls); it's outdoor or mesh; or the remediation cost is brutal — nobody wants to re-lift a scissor lift to a 40-foot deck because the model was 6 dB optimistic. AP-on-a-stick means the actual AP model, at the actual mounting height, at the actual transmit power, measured with a real survey walk. It's the only way to turn a hypothesis into a fact before you pull 300 cable runs.

What are my field rules after 20+ years of doing this?

These come from hurricanes, festivals, surf contests, and one very large Five Guys:

1. Measure your walls. Always. Five measurements, fifteen minutes, and your whole model snaps into focus.
2. Design at 5 GHz (and 6 GHz where clients support it), let 2.4 fall where it falls. If you design for 2.4 you'll build an interference machine.
3. People are attenuators. At Palm Tree and the World Surf League events, an empty-venue survey is a lie by showtime. Add crowd loss or survey during load-in with bodies in the space.
4. Design for the worst client that matters, not the best client you own. Your survey laptop hears everything. The ticket scanner does not.
5. The model ends where the mounting bracket begins. Walk the site with the electrician before you finalize placements.
6. Leave margin. I design to a tighter requirement than the SLA. RF is entropy with an antenna; things only get worse after day one.
7. Validate after deployment, every time. The post-install validation walk is cheap insurance and it's how you learn whether your wall values were right — which makes your next model better.

So is predictive design "good enough" in 2026?

The 2026 tools are the best we've ever had — genuinely. But AI-assisted design has made it easier than ever to produce a beautiful, confident, wrong heatmap in ten minutes. The model gets you 80% of the way faster than ever; the last 20% still lives on-site, in the walls, in the clients, and in the crowd. Predict boldly. Validate anyway.

Want to argue about wall attenuation? Catch me on the Waves podcast or hit me up on socials — I'm the Wireless Nerd, and yes, I will absolutely nerd out about this.

FAQ

How accurate is a predictive Wi-Fi survey?

With measured wall values and honest capacity inputs, a good predictive design typically lands within a few dB of reality in standard construction — close enough for placement and budget. With guessed inputs, errors of 10–20 dB through misjudged walls are common, and that's the difference between -65 and unusable.

Do I still need an AP-on-a-stick survey if I use Ekahau AI Pro or Hamina?

For ordinary offices with familiar construction, often no. For voice, healthcare, warehouses, high-density venues, outdoor, or any building whose materials you can't verify, yes — measure at least your common wall types, and do a full APoS pass where remediation would be expensive.

What inputs matter most in a predictive model?

In order: wall attenuation values, capacity assumptions (users and airtime per area), client capabilities (design for your worst important device), and true mounting heights/locations. Get those four right and the software will do the rest well.

Frequently asked

How accurate is a predictive Wi-Fi survey?

With measured wall attenuation values and honest capacity inputs, a good predictive design typically lands within a few dB of reality in standard construction — close enough for AP placement and budgeting. With guessed inputs, errors of 10–20 dB through misjudged walls are common, and that's the difference between a -65 dBm cell and an unusable one.

Do I still need an AP-on-a-stick survey if I use Ekahau AI Pro or Hamina?

For ordinary offices with familiar construction, often no — a well-fed predictive design is defensible. For voice, healthcare, warehouses, high-density venues, outdoor deployments, or any building whose materials you can't verify, yes: measure at least your 3–5 most common wall types on-site, and run a full AP-on-a-stick pass anywhere remediation would be expensive.

What inputs matter most in a predictive Wi-Fi model?

In order: wall attenuation values (measure them, don't guess), capacity assumptions (users and airtime demand at the busy hour, not the average), client capabilities (design for your worst important device, not your survey laptop), and true mounting heights and locations. Get those four right and the software handles the rest well.