Three running routes in San Francisco, measured

Only one of the three San Francisco routes closes a loop. All three are published with measured figures, not descriptions: Crissy Field to the Golden Gate at 3.96 km, the Embarcadero Waterfront at 4.29 km, and the Lake Merced Loop at 7.01 km.

Three routes that follow the water

All three sit at the city's edge. None of them crosses the city. Two follow the waterfront and one circles a lake, which is what they have in common and the reason the word flat gets attached to them in a city known for its climbs. What the route data actually carries is narrower than that word suggests.

Elevation is not in the route data.

No gradient and no climb figure is published for any curated route, so nothing below describes one. Four things are measured: distance, surface, whether the route returns to its start, and how many corners it contains.

Crissy Field to the Golden Gate

3.96 km, mixed surface, nine corners. It does not return to its start. Run it out and back and the day comes to 7.92 km, which is a useful number to know before you commit to the far end of it.

The park authority publishes its own figure for the same ground. The National Park Service page for the promenade gives "Distance: 2.3 mi (3.7 km)" and a trail surface of "Compacted Natural Material, Boardwalk, Asphalt" (East Beach, Crissy Field Promenade, nps.gov, fetched 2 August 2026), while the Golden Gate National Parks Conservancy lists the same trail as "3.44 miles round trip" (Crissy Field Promenade Trail, fetched 2 August 2026).

Three numbers for one stretch of shoreline is normal.

They differ because they start and stop in different places, and because one of them is a round trip. Pick the one whose endpoints match what you are actually going to run, and ignore the others.

The Embarcadero waterfront

4.29 km, paved throughout, eight corners. It does not return to its start either. As an out and back it is 8.58 km, which lands close to a common weekday distance without any planning.

Eight corners in 4.29 km is one roughly every 540 m.

For a path that follows a shoreline that is a lot of direction changes, and each one is a point where the app has something to say and where you have to pick a line through pedestrians and crossings. If you are running splits, expect a corner near a kilometre marker instead of explaining one afterwards.

The practical cost of a point to point route is the return leg. Either you run back the way you came, which doubles the distance and gives you the same corners in the other order, or you plan a way home before you start. There is no third option that the route itself solves for you.

The Lake Merced loop

7.01 km, paved, fifteen corners.

It is the only one of the three that finishes where it starts. That single property changes how you use it. There is no turnaround decision, no return leg to plan, and a second lap is a known quantity, not a new route.

Fifteen corners over 7 km is roughly one every 470 m, closer together than on the Embarcadero. A perimeter path around water does that: it follows the shape of the shoreline instead of a straight line, and every bend that turns far enough gets counted.

No official perimeter figure appears here. The San Francisco Recreation and Parks page for the lake could not be reached to quote it, and the third party listings that fill that gap disagree with one another, so repeating one of them would pass a guess off as an authority. The 7.01 km above is the measured length of one published line, which is a specific shape, not a statement about the park.

Mixed surface, and what it means underfoot

Surface in the route data takes two values, paved or mixed. Crissy Field is the only one of the three marked mixed. The Embarcadero and Lake Merced are both paved.

The catch?

Mixed means exactly what it says and no more: the ground is not uniform along the route. It does not tell you which part changes, how long that part is, or what the alternative surface is. For Crissy Field the park's own page fills that gap. It lists compacted natural material, boardwalk and asphalt on the same trail (nps.gov, fetched 2 August 2026).

• A change of footing shows up in your splits. The kilometre containing it can read slower without you having worked any less, which is a measurement effect and not a fitness one. The same route after rain is the same distance and a different run.
• Compacted natural surfaces hold water.
• Mixed is a flag, not a description. Treat it as a prompt to look at a photograph or a park page before you decide what to put on your feet.

Corners, and what you hear

Nine, eight and fifteen.

Those counts are not an editorial judgement about which bends feel like corners. A bend only becomes a turn if the direction of travel changes by at least 55 degrees measured 35 m before and after it, and the same 55 degrees must still hold when the check is repeated over a wider 56 m window. That second, wider check is what stops a short sidestep around a building from being announced as a corner.

Turn instructions are computed from each route's own shape by Runflake's own turn detector before the route is published, not written by hand and not taken from a navigation provider. Turns are then classified into six kinds: 55 to 75 degrees are bear left or bear right, 75 to 160 degrees are turn left or turn right, and above 160 degrees is a U-turn.

What that sounds like on the Embarcadero or around Lake Merced is short and repetitive by design. A corner is announced at most twice: once 120 m out, which is the pre-announce distance on every curated route, and once again 25 m before the corner, with the second call dropped if the first already landed inside 50 m. Spoken distances are rounded down, so you hear "In 100 meters, turn right." and never an odd figure, and at the corner itself the distance goes and the line is just "Turn right." Those turn lines are read by the phone's own text to speech voice, not a recorded one, because the distances fall outside the recorded number range.

There is one silence worth knowing about in advance. At the start of a course run no turn is announced until 60 m of forward progress has been seen, and until then the screen says "Working out which way you are running." On a route whose first corner comes early, that gap is the design working, not guidance failing.

Corner density is the number to compare when you are choosing between cities, not between routes. Fifteen corners in 7 km asks for different attention than the two corners in the 9.89 km of Central Park's full loop, and what turn cues do on city routes goes through how that plays out.

Downtown buildings and what they do to your track

The Embarcadero runs along the downtown grid. Any route that passes close to tall buildings meets a well documented mechanism. As one navigation paper puts it, "in the urban canyons of dense cities, buildings block and reflect the signals, causing multipath errors" (Ollander, Bode and Baum, Simulation-based Analysis of Multipath Delay Distributions in Urban Canyons, European Navigation Conference 2020).

For a route run the consequence is that the phone's idea of where you are gets less certain in some places than others, and a fixed tolerance would be wrong in both directions. So the off-route corridor is not a fixed distance. It is 2.5 times the phone's own reported accuracy, with a floor of 45 m and a ceiling of 80 m. When the phone reports itself as less certain, the corridor widens with it, up to that ceiling.

Two things follow that you will actually see. A moment outside the corridor is not enough to be called off route: that takes 12 continuous seconds outside it, or 40 m of route progress while outside, which is why a brief wobble between two buildings passes without comment. When it is called, the run keeps recording and the app says "You are {N} {unit} off the route. Keep running. Guidance resumes when you are back on it." It does not reroute and it does not point you back, so getting to the line again is yours to work out. And when the signal drops out mid route, Runflake stops judging the course rather than declaring you off it, and says so: "No usable signal, so the course is not being checked. Your run keeps recording."

None of this is unique to San Francisco, and the same mechanics apply under a bridge, in a station approach or between two office blocks anywhere. What happens to a GPS track in tunnels and urban canyons covers the general case.

Where the route data comes from

Route shapes come from OpenStreetMap. They are credited to OpenStreetMap contributors under the ODbL licence. OpenStreetMap states the terms plainly: "You are free to copy, distribute, transmit and adapt our data, as long as you credit OpenStreetMap and its contributors" (OpenStreetMap copyright page, fetched 2 August 2026).

That origin has a consequence worth stating. A distance is the length of a line drawn by mappers, not a wheel measurement, and the corner counts are derived from that line by a script, not checked on foot by anyone. Where a park authority publishes a different figure, as the National Park Service does for Crissy Field, neither number is a correction of the other.

The upside is that it is all checkable. The same ground is open in OpenStreetMap, so you can look at the line, see where it starts and stops, and decide whether it matches the run you had in mind before you leave the house.

1. Read the four measured fields firstDistance, surface, loop or not, corner count. Everything else about a route is somebody's opinion, and these four are the parts you can plan against.
2. Settle the return leg before you startTwo of the three San Francisco routes do not come back. Decide whether you are doubling the distance or arranging a way home, because the route will not decide it for you at the far end.
3. Turn the surface value into a picturePaved needs no further thought. Mixed does, so spend one minute on a park page or a photograph instead of finding out at 3 km.
4. Use the corner count as a rhythm forecastA corner every 500 m is a route you steer. A corner every 3 km is a route you settle into. Neither is better, but they ask for different kinds of attention.