How many stores do you need in a parking-lot sample to track comp sales?

There's no single store count that works across every retail chain. Anyone who hands you one without first asking about the store base is skipping a step.
Sample size for a same-store sales proxy is mostly a function of variance inside the chain. A 600-unit chain with tight format discipline, similar box sizes, and a footprint concentrated in a few states behaves very differently from a 600-unit chain split across urban strip centers, suburban power centers, and rural freestanding boxes. The second chain needs a bigger sample, or better stratification, to get the same confidence in the trend line.
What drives the minimum count
Three things push the required sample size up or down, and none of them is the raw store count on its own.
Store-level variance. If weekly traffic at individual stores bounces around a lot relative to the chain average, more stores have to go into the sample before the noise cancels out. Grocery and off-price tend to run tighter variance store to store. Specialty apparel and anything with a strong promotional calendar runs wider.
Stratification. A panel that grabs the first 80 stores it can land a sensor or imagery contract for produces a convenience sample, skewed toward wherever the vendor already had access. If those 80 happen to lean toward malls when the chain is two-thirds strip center, the trend read comes out wrong no matter how many stores get added later. The textbook answer from survey sampling theory says each meaningful stratum, by format, by region, by urban versus rural, needs enough stores on its own, generally landing somewhere in the 30-to-50-per-stratum range before the central limit theorem starts doing real work. Multiply that by however many strata matter for the chain and the "minimum" climbs fast.
What you're trying to catch. Catching a broad directional move, comps accelerating from flat to up mid-single-digits, takes a smaller sample than catching a 100-basis-point inflection a quarter before management flags it on the call. The tighter the signal needed, the more stores it takes to get there, and the app-based panels advertising single-digit-percent coverage of a chain's locations are usually built for the first kind of read.
Why the sample-size question disappears with a census
Most foot traffic panels are built from a subset of locations, usually wherever a mobile app happens to have device penetration or wherever a sensor vendor has an existing install base. That subset gets picked by whatever access the panel already had, independent of the chain's actual footprint. Figuring out whether a panel's sample resembles a chain's real store base means working through the stratification math above before trusting a single number out of it.
A daily vehicle count pulled from satellite imagery sidesteps that work. It counts cars in the lot at every location, store by store, indexed and seasonally adjusted, instead of sampling a slice of the chain. There's no weighting scheme standing between the raw count and the chain-level number, since the chain-level number is just every store's count rolled up. That's the idea behind daily vehicle counts across every lot in a chain's footprint, and it's why the sample-size question that opens this post mostly stops applying once the read comes from a full-footprint count instead of a chain-wide guess.
The short version for a model
Building or vetting a comp sales proxy from a store sample means answering three questions before trusting the trend line: how many stores per format and per region, how was the sample selected, and how big a move needs catching. If the answer to the first question is thin and the second is "whatever the vendor already had," treat the output as a rough direction, not a number to put in a model next to the print.
Worth checking what full-footprint coverage looks like for the names you cover ahead of the next print.