How we test garden apps

The 16 tests, and which have run

Each test is a number, a count or a checklist we can show you. If we can't measure something, we call it opinion. T0 is the one test only our app takes, because it checks our own claims. Every app we compare head to head gets T1 to T15, ours included. Apps we only profile get the scripted ones, the tests a script can run.

Hands-on means a person uses the app, so those tests need a test account for each app. “From November” is when our first round of testing starts, and Wave 2 is our second round, January to April. Before we test an app by hand, we read its terms. Where they forbid benchmarking, we use public information only.

Two limits. We test for US gardeners, so an app built for another country is judged by US needs. And our hands-on tests use one tester, one garden and one region. Wave 2 adds 5 gardeners per app.

The standard test garden

Every app gets the same garden, so the inputs match. Tests T3, T6 and T7 use it.

It's a 4 × 8 ft bed (48 × 96 in), a 3 × 6 ft bed (36 × 72 in) and two 18-inch pots, in ZIP 80202 (Denver). It grows Sungold tomato, a sweet pepper, basil, lettuce, carrots, bush beans, zucchini and fall-planted garlic.

The frost test, written down before it runs

Two regions had fewer than three qualifying stations: Northeast (2) and West (1). Each gave all it had, and 6 more places (the 3 missing, plus 3 extra) were drawn at random from every qualifying station left.

Every planting date an app gives you hangs off two days: your last spring frost and your first fall frost. This test checks how close each app's two dates are to the weather record. We check 30 places and score every app the same way, ours included.

A test you can pick after seeing the results is a test you can rig. So the rules come first. We fix the answer key, the draw and the scoring now, and the 30 places once they're drawn. All of it comes before anyone reads a single app's dates, GardenTrack's included.

Nothing below changes after the first app is read. If a rule turns out to be wrong, we publish the results under both rules.

The answer key comes from stations outside NOAA's station list

Our first plan used our own frost data as the answer key. We dropped that, because it grades every app against our own model. We'd win before the test began.

NOAA's official station numbers have a quieter problem (we publish a cleaned copy of them). If an app just looks up the nearest NOAA station, then testing it at a NOAA station only hands the answer back to us. It would score perfectly, and tell you nothing about a garden that sits between stations, which is nearly every garden.

Our own model is no cleaner. It's built on NASA's Daymet grid, and Daymet is built from NOAA's daily station records, called GHCN-Daily. Here's how we compute frost dates.

So the answer key comes from weather stations that are not in GHCN-Daily. An app built from GHCN-Daily can't have used them, and neither can our model. State and federal networks run their own: university farm-weather networks, state weather networks and the Bureau of Reclamation's AgriMet. A station qualifies only if all six of these are true:

1. It's not in NOAA's GHCN-Daily station list. No GHCN-Daily station sits within 0.6 miles of it, even one listed under a different name and network. We check against NOAA's station file, ghcnd-stations.txt.
2. Its network lets a company publish results from its data, with credit. We quote the terms in the study. A network that limits commercial use is left out.
3. Its daily low temperatures download without an account.
4. It has at least 20 complete years between 1991 and 2024. A spring counts as complete with no more than 9 days missing from January 1 to July 31. A fall counts as complete with the same limit from August 1 to December 31.
5. It freezes (32°F or below) in at least 90% of its years. Where frost barely happens, its usual date barely exists, and our nearest-station study uses the same cut.
6. A ZIP code's center sits within 3 miles of it. That ZIP is the place we type into each app, and each ZIP gets one station.

The dates are counted, not fitted. For each year, we find two days: the last day from January 1 to July 31 at or below the threshold, and the first day from August 1 on that is at or below it. The 50% date is the middle one (the median) across those years. We work out 28°F, 32°F and 36°F, each at 10%, 50% and 90%, so we can match any app's own definition.

The 30 places are drawn, not chosen

• 27 places, three per climate region. NOAA splits the country into nine regions: Northeast, Upper Midwest, Ohio Valley, Southeast, Northern Rockies and Plains, South, Southwest, Northwest and West.
• 3 more, from every qualifying station left. That makes 30.
• The draw is random, with a fixed seed: 20261012. A seed is a starting number. The draw runs over qualifying stations sorted by network and station ID, so anyone with the station list can re-run it and get the same 30.
• A short region gives all it has. If a region has fewer than three qualifying stations, we take them all and draw the rest from the other regions. The study names the short region.
• The list is frozen once it's drawn. For each place: the station, its network, the ZIP, the distance from the ZIP's center, the years used and every answer-key date. It goes on this page before we read any app.

How we read each app

• We type the ZIP. If an app takes only a town, we type the ZIP's town. If it uses only your phone's location, that place counts as “no answer” for that app.
• We write down what the app says its date means. Which temperature, and which odds (“average”, “10% chance of a later frost”). If the app doesn't say, “not stated” goes on the record as a finding.
• We keep the evidence. For each reading we record the screen, the app version and the day we read it. Every app gets read within the same 7 days.
• GardenTrack gets the same ZIP and the same rules. It also takes a street address. That reading, at the station's own spot, is a separate row and never the headline.
• We don't change GardenTrack's frost model between the draw and the results. We record which version we used. And we don't look up GardenTrack's dates for these 30 ZIPs before the run.

How we score it

• Error is days. It's the gap between the app's date and the answer key, matched to the app's own definition. If an app says 32°F at 50%, we score it against that. If it says nothing, we score it against 32°F at 50%, the common meaning of “average last frost.”
• Spring and fall are scored apart. For each app we report six things: the median error (the typical miss), the error that 9 in 10 places stay under, the worst miss, the share of places within 7 days, the share within 14 days, and how many of the 30 places it answered.
• The result is a table, not a score. We don't weight anything or hand out stars.
• Every place stays in. The only exclusion is “no answer.” We count it as a gap in coverage, never as an error. Nothing is dropped after the run.

What this test can't tell you

• The answer-key stations are new to every app, ours included. That's the design. If a station turns out to be in GHCN-Daily after the run, we report it and show the results with and without it.
• Frost history isn't a forecast. The test measures how well an app knows a place's usual dates, not whether it predicts this year.
• A ZIP code is a few miles wide. Two gardens in one ZIP can differ. Every app gets the same limit, and the street-address row shows what an address adds.

The schedule

• October 12, 2026: these rules, written down.
• By October 26, 2026: the answer key computed. The 30 places were drawn and frozen on October 13.
• November 2026: the run, once we have a test account for each app.
• Early December 2026: the study. The method, the frozen list, every reading with its screen, and the raw data, free to reuse with credit (CC BY 4.0).

How facts stay true

These are the rules for every app page we publish.

• Every fact has a source and a date. The source is a link, and the date is the day we read it.
• We re-check within 60 days. Past 60 days old, a fact raises a warning. Past 120 days, it fails our publishing checks, so it comes off the page.
• A refresh means reading the source again. Changing the date alone doesn't count.
• App Store facts are re-read by script once a month. Plans and makers' pages are re-read by hand and dated again.
• “Can't” needs proof. We say an app can't do something only if we've seen it for ourselves, by hand. A feature missing from a home page isn't a feature the app lacks.
• Our own app gets the same treatment. Our record of GardenTrack's features was filled on October 14 from a hands-on walk of the running app, never from memory or our own docs. A feature we looked for and did not find is recorded as not found, with where we looked.
• We name apps in text and never show their logos.

Plant data follows its own rules. They're on How we research.

Why no scores yet

Our first round has no overall score. No app is best at everything, ours included. So we give verdicts by use, like “best for…”, and put the measurements under each verdict. You can see what it rests on.

Scores come in Wave 2, from January. We'll compute them from the measurements, and we'll publish the weights before we compute anything.

Every comparison we publish has a “where it beats us” section. It's never empty.

If we got something wrong

Anyone can send a correction. That includes the makers of the apps we test. Email hello@gardentrack.app with the page and what's wrong. We fix errors and log every fix here, with its date.

• October 14, 2026. Our own app (T0) checked by hand in the running app. Its feature record is filled from that run.
• October 14, 2026. First hands-on run, on Seedtime's web app. Test statuses updated.
• October 13, 2026. The 30 places drawn and listed.
• October 12, 2026. Page published. Corrections so far: none.