Industries / Construction and property / Commercial and residential construction / Electrical contracting / Solar installation
Site survey and design
This work decides whether a roof or plot suits solar, measures its sunlight and shade, lays out the panels and wiring, and estimates how much power the system will make and what it will save.
5
Milestones achieved, from 1999 to 2024.
1
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1999
Free yield estimates online
The US National Renewable Energy Laboratory (NREL) put PVWatts online, a free calculator that estimates a system's yearly output from its location, size, tilt and direction, using methods from Sandia's PVFORM model of the 1980s. NREL's manual says the estimate can be off by up to 10% for a typical year's total and by up to 20% in any particular year, because real weather varies. Europe's Joint Research Centre built a free equivalent, PVGIS, described in a 2005 paper by Marcel Šúri, Thomas Huld and Ewan Dunlop, which still covers Europe, Africa and much of Asia and the Americas.
Credit
NREL; the European Commission's Joint Research Centre
2014
Shade measured without ladders
A solar access value (SAV) is the share of the year's sunlight that reaches a point on a roof after trees and buildings cast their shade. Installers used to climb the roof and read it with a handheld device such as Solmetric's SunEye. In a blind study, NREL found that the startup Solar Census, which works from aerial imagery and 3-D data, matched SunEye readings within ±3.4 SAVs. NREL later ran the same kind of test, 81 points on eight houses in Denver and Los Angeles, for OpenSolar in 2021 (within ±3 SAVs) and for Enphase's Solargraf in 2023 (within ±2.39 SAVs).
LessonA remote survey can stand in for a roof climb once an independent lab has tested it point by point.
Credit
Solar Census; NREL
2015
Sunlight mapped for every roof
On 17 August 2015 Google engineer Carl Elkin launched Project Sunroof, which uses the aerial mapping behind Google Earth to estimate the sunlight on a roof, allowing for its direction, shade from trees and buildings, and local weather, and then estimates the savings from solar. It started as a side project covering the San Francisco Bay Area, Fresno and the Boston area. Google later said Sunroof held height and color data for 43 million homes.
Credit
Carl Elkin and Google
2020
Plants miss their forecasts
A P50 forecast is the yearly output a solar plant should reach or beat half the time, and it is the base case lenders use. In its 2020 Solar Generation Index, kWh Analytics, which insures solar output, pooled data from 10 of the 15 largest US asset owners, about 30% of operating utility and commercial capacity, and found projects fell short of their P50 by 6.3% on average after allowing for weather, with the bottom quarter more than 10% short. The engineering firm DNV GL had separately found a 3% to 5% overestimate, and in May 2026 the Electric Power Research Institute (EPRI) still put the shortfall at 5% to 8%.
LessonDesign software's default loss assumptions were too hopeful, and the gap only showed once many owners' meter data were pooled.
Credit
kWh Analytics; DNV GL; EPRI
2024
Regulator flags oversold savings
On 7 August 2024 the US Consumer Financial Protection Bureau (CFPB) reported on solar sold with loans, often door to door from a tablet. It found sales pitches that told homeowners panels would pay for the loan and end their power bills, and quotes that subtracted a 30% federal tax credit from the price as if every buyer would receive it; the credit depends on the buyer's tax bill.
LessonA design's savings figure is only as good as the tariff, tax and output assumptions behind it, and buyers rarely see them.
Credit
CFPB
Watch for
Address to plan set
Design software makers say AI now turns a street address into a full design. Pegasus, a US maker of roof mounting hardware, said on 16 November 2026 that its GlideAI platform detects roof outlines and obstructions from an address and produces a proposal with a code-compliant mounting layout in under five minutes, plus draft plan sets, parts lists and permit documents. Aurora Solar says its Aurora AI builds a roof model and a code-compliant panel layout, including fire pathways (clear walkways firefighters need on the roof), in minutes. Neither company has published how often its designs pass plan review.
Achieved whenan outside party runs one of these tools on a stated number of addresses and reports how often its designs match the approved permit drawings or pass plan review unchanged.
Claimed by
Pegasus, November 2026; Aurora Solar, 2026
Open
01
Proposals checked against meters
A public set that pairs the yearly output promised in home solar sales proposals with what the same systems' meters recorded in their first full year, corrected for that year's weather, across many installers and at least three countries.
The briefCredit · unclaimed02
A shade test outside America
An open benchmark of remote shading tools on roofs outside the US: measured solar access values at marked points, compared with what Google's solar data, PVGIS, OpenSolar, Aurora and others predict for the same points.
The briefCredit · unclaimed03
AI designs against approved plans
A scored comparison of AI-made rooftop designs with the drawings that permit offices actually approved for the same houses: panel count, layout, setbacks and fire pathways.
The briefCredit · unclaimed
Open 01 · the brief
Proposals checked against meters
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set that pairs the yearly output promised in home solar sales proposals with what the same systems' meters recorded in their first full year, corrected for that year's weather, across many installers and at least three countries.
Why it is still open
The largest pooled checks cover utility and commercial plants, not homes, and the studies of homes point both ways. A TU/e Eindhoven study of 256 identical systems on one builder's houses at 19 Dutch sites found output about 6% below what simulation models predicted, and its author concluded that sellers tend to overestimate. Mahelet Fikru's 2019 study of US solar adopters found their actual bill savings averaged 20% more than estimated, though more widely spread. The CFPB found savings oversold in US loan sales. This research found no public set that compares many installers' own proposals with metered output.
Built with what exists
Homeowners' own proposal PDFs; data exports from inverter monitoring portals (Enphase, SolarEdge, SMA and others let owners download their history); PVGIS and PVWatts reruns with the actual year's satellite weather; an AI agent that pulls the promised kWh, system size and assumptions from each proposal and cites the page it took them from (see the Note "Findings that cite notes").
First step
Ask owners in two solar owners' forums for 30 proposals and 12 months of monitoring exports, and compute the first ratios.
Done looks like
500 homes in at least three countries, each with its proposal figure, metered first-year output and weather-corrected ratio, published openly with names and addresses removed, plus the median shortfall or surplus by country and by type of seller.
Open 02 · the brief
A shade test outside America
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open benchmark of remote shading tools on roofs outside the US: measured solar access values at marked points, compared with what Google's solar data, PVGIS, OpenSolar, Aurora and others predict for the same points.
Why it is still open
NREL has tested at least five remote shading tools against SunEye readings taken on the roof: Solar Census, Sighten and HelioScope in 2014 to 2018, OpenSolar in 2021 and Solargraf in 2023. Each test was arranged with the vendor, and all were in US cities: Denver, Los Angeles and, for HelioScope, Camden, New Jersey. No public test covers flat concrete roofs, dense terraced streets or tropical tree cover, where much new rooftop solar is now going in.
Built with what exists
A SunEye or Solar Pathfinder where one is available, or fisheye photos of the sky processed with open sky-masking code; GPS points; the public interfaces of the tools under test; NREL's equivalence test method from the reports above.
First step
Measure 20 points on flat roof terraces reached by a stairway, never on pitched roofs, at three buildings, and compare them with one tool's remote estimate.
Done looks like
300 measured points at 30 sites in at least three countries outside the US, released openly, with each tool's average error in SAVs and the share of points within ±3 SAVs.
Open 03 · the brief
AI designs against approved plans
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A scored comparison of AI-made rooftop designs with the drawings that permit offices actually approved for the same houses: panel count, layout, setbacks and fire pathways.
Why it is still open
AI design tools claim code-compliant layouts in minutes but publish no error rates against approved permits. Aurora's own study ran 4,000 US homes through Aurora AI and compared its roof models with models of the same homes drawn by people in Aurora's software, not with approved permit drawings. NREL's November 2025 review of US home projects found that the design phase, from contract signing to permit submission, was one of the two phases that grew longest between 2019 and 2023.
Built with what exists
Approved residential solar plan sets from city permit records; trial accounts on two or three AI design tools; a fixed list of yes-or-no code questions answered for each design by a model that takes many fixed-answer questions in one pass (see the Note "A thousand questions, one pass"), then checked by a licensed designer.
First step
Request 20 approved residential solar plan sets from one city's public permit records.
Done looks like
100 addresses, each with the approved plan and three tools' designs, and a published table of panel-count error, layout overlap and the share of designs a licensed reviewer would pass unchanged.