When the seasons change, especially as winter approaches, many people habitually pull back the curtains after waking up to see how others are dressed, using that as a gauge for whether the day's temperature is cold or warm.
This is essentially a simple form of "data analysis"—inferring overall trends by observing nearby samples.
The recently launched "Street Ranking" by Amap follows a similar logic, using data on navigation destinations and travel behavior to determine whether a store is worth visiting.
Limitations of Observational Indicators
The problem with this judgment method is obvious.
Winter in Guangdong is a classic example of chaotic samples—some people wear short sleeves, others wear down jackets, and still others pair trench coats with shorts or scarves with short sleeves.
In such an environment, even if you see most people in short sleeves within a minute, you can't conclude that "it's hot today."
From a statistical perspective, this is a combination of "small sample bias" and "high variance noise."
The people outside are just a random sample of the urban population, and their clothing choices are influenced by too many random factors (commuting methods, physical differences, psychological expectations, etc.).
The result is that the signal you get carries strong volatility and randomness.
When this judgment goes wrong, the misattribution also becomes skewed.
If you dress lightly because you saw people in short sleeves and end up freezing, you'll curse those short-sleeve wearers as crazy; if you fall into a trap due to a Street Ranking recommendation, you'll blame Amap for being unreliable.
By simply using navigation data for algorithmic analysis, Amap struggles to provide users with recommendations that meet expectations (or align with their marketing claims). If Amap wants to avoid failure, it must introduce data beyond navigation for cross-validation.
Transaction Data Is the Most Indicative Metric
What truly reflects a store's performance is "transaction data."
When the unit business area and product prices are clear, stores with "high transaction volume, high revenue, high repurchase rate, and wide customer distribution" are more stable and attractive choices.
These metrics are "high signal strength" data, better revealing true commercial quality.
However, the problem is that transaction data is more sensitive than navigation data, involving merchant privacy and platform compliance. Even though Amap is a wholly-owned subsidiary of Alibaba, it cannot use such data openly.
Therefore, products like Street Ranking, if they reflect real popularity to some extent, likely achieve this by analyzing "proxy variables highly correlated with transaction behavior," such as navigation visits to stores, user repeat visit frequency, and the proportion of long-distance navigation trips.
At the same time, from the currently displayed ranking types, it's evident that Amap uses data beyond navigation that is highly correlated with transaction data (or location data at the time of payment) and user location data when not navigating (such as real-time device positioning).
Although these behavioral indicators are not transaction data themselves, they are statistically highly correlated with consumption behavior.
This approach is called "correlation inference" or "proxy modeling": when target data is unavailable, the target value is indirectly inferred through observable related variables.
So, Amap's Street Ranking "is likely constructed through correlation analysis between navigation behavior and consumption trends," rather than directly referencing transaction amounts.
This is also a typical "weak signal amplification" strategy: using enough indirect indicators to construct a judgment close to reality.
The Boundary of Rankings Is the Greatest Common Divisor, Not the Optimal Solution
However, no matter how advanced the algorithm, it cannot eliminate the bias caused by "group averages."
Rankings like Street Ranking essentially only reflect stores that "most people think are good"—statistically close to a "majority consensus solution," like the McDonald's near your home.
It can significantly reduce the probability of falling into a trap, but it cannot guarantee you'll choose the "optimal" option.
Just like the result of a public vote is usually a plan that "most people can accept or won't strongly oppose," rather than the most efficient one.
Moreover, navigation data itself is extremely noisy. Many "check-in" travel behaviors do not represent actual satisfaction—who hasn't fallen for a trap at a trendy internet-famous store?
Nevertheless, this methodology is still worth learning from.
When we cannot obtain first-hand data, we can build a reasonable judgment system through a series of "indirect indicators."
For example:
- To estimate foot traffic in an area, look at surrounding residential density, urban village distribution, and the number of subway entrances;
- To determine if a road section is prone to congestion, check nearby schools, office building distribution, and intersections with main roads.
This is essentially the thinking of "feature engineering": when core variables are missing, approximate the real-world operating rules by establishing a reasonable combination of proxy variables.
Reasoned "Nonsense"
Real-world data is never perfect.
The crowd you see outside, the rankings on Street Ranking, and the heat maps are all noisy signals.
But as long as you understand these data with "logical and probabilistic thinking," you can extract relatively reliable judgments.
In statistics, this is called "Bayesian inference": in an uncertain environment, combine limited information with existing experience, gradually revise beliefs, and approach the truth.
So, when you can't get real data, judging the weather by "looking at people outside" or evaluating a business district by "Street Ranking" isn't wrong.
The key is whether you know its limitations and can identify signals amid the noise.
After all, compared to being completely blind, reasoned nonsense is often closer to the truth.