Read a Chinese Delivery Address
and Phone Number From a WeChat Chat Screenshot

What a WeChat Address Message Actually Looks Like

A typical Chinese delivery address sent in WeChat chat follows a big-endian hierarchy: the largest geographic unit comes first, the smallest last, followed by the recipient's name and phone number. A complete message reads like this:

In that single string, a human reader instantly identifies six distinct data points: the province (省), city (市), district (区), street and building (路/号/楼/室), recipient name (收件人姓名), and phone number (手机号). The address is structured — Chinese address format is one of the most consistently hierarchical postal systems in the world — but that structure lives entirely in the meaning of the words, not in any visual separation. There are no line breaks, no column dividers, no form fields. Just one continuous block of Chinese characters and digits.

The address is perfectly structured. It just does not look structured. And that is exactly the kind of input that traditional OCR handles worst.

The same format repeats across every WeChat order scenario. A customer buying clothes from a WeChat seller (微商) sends their address in one message. A daigou buyer confirms their shipping details the same way. A restaurant supplier receiving delivery instructions over WeChat gets the same kind of text — province through phone number, all in one block. The content is identical regardless of who sent it: geographic hierarchy + name + phone, everything concatenated into a single string that a human parses in half a second and a machine misreads as a single undifferentiated blob of text.

The Hierarchy Problem — Why a Flat String Is Not an Address

The difficulty is not that Chinese characters are hard to recognise. Modern OCR reads printed Chinese with high accuracy — the technology for character recognition has been reliable for years. The difficulty is that a Chinese address is not a sequence of characters; it is a hierarchy expressed as a sequence of characters. Traditional OCR treats the entire string as one line of text. It outputs "浙江省杭州市西湖区文三路478号华星科技大厦12楼1208室 王芳 13812345678" as a single field — one continuous string with no semantic segmentation.

The problem compounds when addresses are not perfectly formed. A customer might omit the province entirely when ordering from a seller in the same province: "杭州市西湖区文三路478号" without "浙江省" at the front. Another might include only the city and district: "深圳南山科技南路18号" — dropping the province and using a well-known shorthand for the district. Some addresses omit the building name and give only the street number and a nearby landmark. Traditional OCR has no way to normalise these variations because it does not know it is looking at an address at all.

The phone number, which looks like the easiest piece to separate, has its own traps within the address block. A Chinese mobile number is 11 digits starting with 1 — a clear pattern in isolation. But when it sits at the end of an address string that also contains building numbers (like "12楼" or "1208室"), a tool that naively extracts all digit sequences might grab "478" (the building number), "12" (the floor), "1208" (the room number), and "13812345678" (the actual phone). It has no way to know which digit group is the phone number and which are address components.

The phone number pattern (11 digits, starts with 1) is easy for a human to spot — but a tool that does not understand it is looking at an address will grab every number on the page and call them all "data."

How Visual AI Reads the Same Screenshot Differently

This is where the extraction approach shifts from character recognition to semantic understanding. Instead of scanning the screenshot for all text and outputting it as a flat block, a visual AI tool reads the image by asking: "what kind of information is this, and where does each piece belong?"

With Custom Column Extraction, you define the columns you want — "Delivery Address (收货地址)", "Recipient Name (收件人)", "Phone Number (联系电话)" — and the AI locates the matching values by understanding what each one means within the context of the Chinese address format. It does not need to be told that 省 marks the province or that 路 marks the street name. It identifies these markers because it understands the structural grammar of Chinese addresses — the same way you instinctively know that "浙江省" is the province even if you have never studied Chinese geography.

Here is how the semantic approach handles each component of the address block:

This approach works because the AI does not need to know exactly which city or which street name appears in the screenshot. It needs to know that something in that position functions as the city, something functions as the street, and something functions as the phone number. The extraction is driven by the structural role each segment plays in the Chinese address format — a format that, despite local variations, is one of the most consistent postal hierarchies in the world.

The same principle that applies to extracting data from payment screenshots that are not formal receipts applies here: you define the output columns, and the AI locates the matching values on the screenshot by understanding what they mean, not by matching a template. Whether the input is a WeChat Pay confirmation screen or a chat message with a delivery address, the method is the same — semantic extraction reads by meaning, not by position.

Batch-Processing Multiple WeChat Orders

A WeChat seller does not receive one order at a time. On a busy day, the chat log fills with address messages from multiple customers — each with a different name, a different city, a different set of delivery instructions. Manually transcribing ten addresses means reading ten messages, copying ten phone numbers, and typing ten names into an order sheet or a package label. The process is the same every time, and the error pattern is the same every time: a digit swapped in the phone number, a street number misread, a district misidentified.

Batch processing changes that workflow. All the screenshots — each showing one customer's address message in chat — are uploaded together. A single set of column definitions ("Delivery Address", "Recipient Name", "Phone Number") is applied once. The AI reads every screenshot in the batch and outputs a unified table: one row per customer, with the address, name, and phone number in three clean columns.

The output table can be exported directly to Excel or written straight into Google Sheets using the ImageToTable.ai Google Sheets add-on — without leaving the spreadsheet. Each row is one order, ready to print as a shipping label, import into your order management system, or share with your fulfilment team. The per-order effort drops from roughly a minute of reading and typing to a few seconds of verification.

The value of extraction is not about saving a few seconds on one address. It is about scaling from ten orders a day to thirty without adding more manual work — because the bottleneck is not reading the address, it is retyping it every single time.

The same batch workflow applies across the address format variations discussed earlier. A batch might contain ten messages with perfectly structured addresses (province→city→district→street→name→phone), two with the province omitted, one where the customer included a landmark instead of a street name, and one in a municipality (Beijing or Shanghai) that has no province level at all. The AI handles all of them in the same pass, because it reads each address by its structural components — not by assuming every address follows the exact same template.

For sellers who also handle orders on WhatsApp — especially those serving customers across different markets — the same approach extends to address messages sent via WhatsApp Business. The extraction columns stay the same; only the chat interface changes.

Frequently Asked Questions