The predictive-versus-on-site argument is usually conducted badly, because both sides have a commercial interest in the answer. Firms who only model will tell you site visits are an expensive ritual. Firms who only walk buildings will tell you a model is a guess in colour. Neither position survives contact with a real project.
What follows is what each method actually does, what each one genuinely cannot do, and how they get used together — including the cases where a predictive survey on its own is the right and sufficient answer.
What a predictive survey is
A predictive survey models radio propagation in software. You import a scaled floorplan, draw the walls and structures with attenuation values assigned to each material, place candidate access points with specific antenna patterns and heights, and the tool calculates the resulting signal, signal-to-noise and capacity across the space.
Done properly this is a considerable piece of work, and the phrase "desktop survey" undersells it. The quality of the output is almost entirely a function of how carefully the environment is built: a model with every wall drawn to its real construction, ceiling heights correct, mounting positions constrained to places that physically exist, and racking represented as attenuating structure will be far more useful than a fast model with generic walls.
What predictive does well
- Buildings that do not exist yet. The only option for new build and pre-fit-out, and the whole reason it is used at design stage.
- Comparing options cheaply. Access point count, mounting height, antenna type, channel width, band strategy — all can be tested against each other in an afternoon without anyone standing on a ladder.
- Producing a buildable specification. Positions, heights, orientation, cable routes and a port count, which is what the cabling and containment package has to be priced from.
- Budgeting and tendering. A defensible access point count and kit list early enough to be useful.
- Coverage across a whole estate. Modelling twenty similar sites is tractable; visiting twenty sites is a programme.
What predictive relies on, and where those assumptions break
Everything a predictive model produces rests on the attenuation values assigned to the materials in it. Those values come from published libraries and measured averages, and they are reasonable. They are also generic, and real buildings are specific.
| What the model says | What is actually there |
|---|---|
| Plasterboard partition | Single or double layer, timber or metal stud, sometimes with foil-backed insulation or a metallic vapour barrier that attenuates far more than the plasterboard itself. |
| Glass | Plain glass is nearly transparent to RF. Modern low-emissivity and solar-control glazing carries a metallic coating and can attenuate heavily — which is why some buildings have superb internal coverage and none of it reaches the car park. |
| Concrete wall | Thickness and reinforcement vary enormously. Dense rebar behaves closer to a mesh than to masonry. |
| Open floor | Racking, mezzanines, plant, stacked stock, partitioning added after the drawings were issued, and furniture and people that were never on any drawing. |
| Suspended ceiling | The void above it may be full of ductwork, cable tray and foil-faced insulation — all of which sit between a ceiling-mounted access point and the floor below. |
Beyond material values, there are four things a model structurally cannot know:
- As-built variation. Drawings are frequently out of date, and refurbishments rarely update them. Walls move, openings get filled, plant rooms appear.
- The neighbours. A model has no idea what the adjacent unit, the floor above or the building across the road is broadcasting, on which channels, at what power. In a multi-tenant building or a high street that can be the dominant factor in real-world performance, and it is invisible until somebody measures it.
- Non-Wi-Fi interference. Motors and drives, some lighting systems, microwave ovens, wireless cameras, cordless equipment, and — in parts of the 5 GHz band — radar, which will move your access points off a channel through dynamic frequency selection whether you like it or not. None of this appears in a propagation model, and much of it does not appear in Wi-Fi analysis tools either. It takes a spectrum analyser.
- Stock and use. What is in the racking, how full the building is, when the shutters are open, where the vehicles park.
What an on-site survey is, and its own limits
An on-site survey measures the building. In practice it covers several distinct activities that get bundled under one name, and it is worth separating them because they answer different questions:
- Passive survey. Walking the building listening to what is on the air — existing networks, signal levels, channel utilisation, the neighbours. Answers "what is happening here now".
- Active survey. Associating to a network and measuring what a real client actually achieves — throughput, retries, roaming behaviour on the move. Answers "what does a device experience".
- Access point on a stick. Temporarily positioning a real access point at candidate locations and measuring the resulting coverage. This is how you calibrate assumptions: it tells you what the racking, the walls and the stock are genuinely doing.
- Spectrum analysis. Looking at the RF energy in the band regardless of whether it is Wi-Fi. The only way to find the interference sources that a Wi-Fi-only tool reports as unexplained noise.
- Validation. Measuring a completed installation against the design it was supposed to deliver.
On-site is not a universal upgrade, and it has real limitations:
- It is a snapshot. Measured on a Tuesday in an empty building, it describes a Tuesday in an empty building. Warehouses fill; hotels fill; the neighbours change their network.
- It cannot survey what is not built. No amount of walking will tell you about a building at design stage.
- It measures what is there, not what should be. A survey of an existing bad network tells you the current network is bad. Turning that into a design still requires modelling.
- It costs time. Access, escorts, working at height, out-of-hours attendance, live operations. On a large or multi-site estate this is the constraint that shapes the whole approach.
When predictive on its own is genuinely enough
There are cases where insisting on a site visit adds cost and no information, and it is worth being straightforward about them:
- New build and pre-fit-out. There is nothing to measure. Predictive is not the compromise here, it is the method — and the design gets verified by validation after installation.
- Simple, uniform, well-documented spaces. A single-floor office fit-out with known partition construction, standard ceilings, ordinary glazing and a laptop-and-phone device fleet. The model has little room to be wrong, and the consequences of small errors are small.
- Budget and feasibility stage. When the question is "roughly what will this cost and is it viable", a model answers it well enough to make the decision, and a site visit can wait until the project is real.
- Rollouts to a validated template. Where one representative site has been surveyed and validated on site, and the remaining sites share its construction and use, modelling the rest against that calibrated template is sound engineering rather than a corner cut.
- Extending a known-good network. Where the existing design is understood and has been measured, an extension into similar space can often be modelled with confidence.
When it isn't
- Anywhere with racking or stock. The single biggest source of model error, and the effect changes through the year.
- Older or altered buildings. Where the drawings are unreliable or the construction is unknown, the model's inputs are guesses and its outputs inherit that.
- Dense multi-tenant environments. High streets, business parks, shared office buildings, blocks of flats. The neighbours may matter more than your own design, and only measurement reveals them.
- Critical or intolerant device fleets. Voice, scanners, automated vehicles, medical equipment, anything where a dropped session has an operational cost. The design margin needs to be verified, not assumed.
- Outdoor and mixed indoor-outdoor. Yards, quaysides, campuses, car parks. Terrain, vegetation, vehicles, weather and reflective surfaces are poorly represented in indoor propagation models.
- Any site with an unexplained existing problem. If something is already wrong, modelling will not find it. That is a measurement exercise.
Why most real projects use both
The two methods answer different questions, and a project of any size asks both.
- Model first. The predictive survey establishes the shape of the design — how many access points, roughly where, what antennas, what the cabling and containment package looks like — early enough to price and programme the work.
- Measure to calibrate. An on-site survey tests the model's assumptions where they carry the most risk: the racking, the odd walls, the neighbours, the interference, the mounting positions that turn out to be occupied by a duct. The model is corrected against measurement rather than replaced by it.
- Validate after installation. A validation survey measures the finished network against the design's stated thresholds. This is the step that closes the loop, and it is the only one that produces evidence rather than intent.
The proportions shift with the building. A new-build office might be mostly model with a short verification visit at the end. A live distribution centre with dense racking and a scanner fleet is weighted the other way. Both are legitimate; what is not legitimate is choosing the method for the supplier's convenience and presenting it as the only way.
If you are trying to work out which you need, the useful questions are: does the building exist, are the drawings trustworthy, is there anything in it that a model cannot see, and what is the cost of getting it wrong? Those four answers usually settle it. More on how surveys work.
Let's find out what your Wi-Fi is actually doing.
Book a survey and you get measured data, a design you can build to, and a number you can budget against — not an opinion.