The lab gains are real, but they are conditional — and the condition is usually the part that does not survive deployment. I have not deployed a production RIS, but I have measured passive and active reflectors in the same spirit, and the pattern matches what LunaWei outlined.
In a static, line-of-sight-controlled test, a RIS can deliver a large gain, because every element adds coherently. The textbook scaling is roughly N squared for the reflected power, so a few hundred elements can in principle give 10–20 dB over the blocked path. In practice I would budget for much less. One-bit phase control alone costs about 3–4 dB against ideal continuous phase, element insertion loss and polarisation mismatch cost another 1–3 dB, and any error in the channel estimate or in the element calibration eats the rest. A gain that was 15 dB in an anechoic chamber can be 4–6 dB in a real room, and close to zero if the user moves faster than the control loop can track.
That is the second problem: control overhead. The RIS has to know the channel to both ends before it can set its phases. For a static user in a fixed coverage hole, you can calibrate once and hold the pattern, and the overhead is negligible. For a moving user, you are re-estimating and re-programming the surface continuously, and the signalling and latency can cost more throughput than the reflection gains. That is why the convincing deployments I have seen are fixed coverage holes, indoor hotspots, and quasi-static links — exactly the cases LunaWei listed — rather than a general mobile-cell silver bullet.
The honest comparison is also against a passive alternative, not against nothing. A well-placed passive reflector, a repeater, or simply moving the access point often gives half the gain for a tenth of the complexity. I would only choose a RIS where the geometry is fixed, the baseline is a true blockage, and I can measure the gain as a distribution over a day — median and 10th percentile, not the best static point — against that passive baseline, including power, control, and failed-element replacement.
If you have a specific band, element count and scenario in mind, post them and I will sanity-check the expected gain against the N-squared ideal and the quantization and loss budget.