Someone who does product sourcing copies three things off his phone every morning: competitor prices, ranking positions, and his own shop’s sales.
Three to five rows on a page, five or six pages. It does not sound like much. Once he put two days of figures side by side, one number was off by a zero. That day he slipped on page four. He said, “I know it goes wrong later on. I just don’t know which page.”
Manual copying starts shifting rows by page six. His case was earlier. It went wrong on page four, and afterwards he still could not match which page it was.
Why this work suits a machine
Data collection sits on three things at once. It happens every day, sometimes several times a day. Open the page, read the figures, write them down: the same actions on every page. And a wrong number is obvious at a glance.
The third one lets you check. Spot-check the first few runs, confirm the reading is right, and you can keep using it after that.
“Should I follow up with this customer” does not fit here. It happens often, the steps are not fixed, and the result cannot be checked.
Three stages from the screen to the sheet
Read it first. Where the page itself has text structure, read the text. That route is the most accurate. Where there is only a picture and no text, for example when the figures have been turned into an image, use text recognition. The same page can mix both. Read the text where you can find structure, and recognise the fields where you cannot.
Then turn it into rows and columns. Name the fields first: this column is the date, this one the product name, this one the price. Leave that unclear and what you read is a lump of text. Three months of it is still unusable.
Once the fields are set, decide which place on the page corresponds to which column. The failure that shows up most is a change in order. One day a field moves, the machine keeps reading the old place, and the columns shift. When you read, look the field up by name. Do not look it up by position.
Last, write it into the sheet. Append to one fixed sheet. Each run adds a row. Do not overwrite. Three months later you have data you can chart. Overwrite, and what remains is that day’s snapshot.
The write step is the one most often skipped. Skip it, and what you read that day is gone once you have looked at it. Thirty days of figures side by side is what shows the change.
Two safeguards, then one page
Leave a record at each stage: which page you reached, how many fields that page yielded, and whether anything failed to read. When something goes wrong you can point to the stage.
Stop when a value is far from the last reading. If a price suddenly doubles, do not write it into the sheet. Save a screenshot and wait for someone to confirm. A misread can still be corrected. A wrong number that has entered the sheet, and then been used for a decision, is hard to pull back.
Get one page working first. Read it, turn it into fields, write it in, and confirm three things: whether the reading is accurate, whether the fields are right, and whether it lands in the right place. Spot-check the first few runs yourself. Move on only after they match.
Add a second page once the first runs steadily. Start five pages at once and, when something fails, you cannot tell whether the reading, the fields, or the placement is wrong. Sorting that out takes longer than running them one by one.
Reading and writing for one page usually takes an hour or two for a first version. The rest of the time goes into checking.
Most tools only cover how to operate: where to tap, where to type. After that, how the data becomes something you can look at is rarely followed through. Having read it, and being able to see a trend, still means assembling fixed rows and columns. The same screen reading, with that output attached, turns a record into something you can use for a later judgement. What you save is the copying, and the extra pass of assembling that used to follow it.
For a related idea, read let the AI look first, then decide. For watching competitor figures every day, see how to watch competitor prices and content.
Frequently asked questions
- Why does data collection suit automation?
- It meets three conditions: high frequency, fixed steps, and a verifiable result. Jobs that meet all three are rare, and this is one of them.
- What are the ways to read data off a screen?
- Two. Where the page carries structured text, read the text. Where only an image exists, run text recognition. The two can be mixed, whichever works better for a given field.
- Which is more accurate, manual or machine?
- Manual is better for the first five pages; the machine is better from the sixth onward. The issue is not ability but attention, which necessarily declines with repetition.
- What if it reads something wrong halfway through?
- Two safeguards. Leave a record at each stage so a failure can be traced to a page; and add a boundary where a value far from the last reading stops the run for confirmation rather than being written straight in.
- What format should the output be?
- Append to one fixed sheet, one new row per run. Three months later you have data you can chart, rather than a pile of daily snapshots.
- How do I sequence several pages?
- Get one page reading and writing correctly before adding a second. Five pages at once leaves you unable to tell which stage failed.
- How is this different from a patrol?
- A patrol looks and records. Data collection adds one step: turning it into something analysable. The first is a record, the second is an output. Adding the output is what turns the same screen reading into something that supports decisions.
- How long does it take to get running?
- An hour or two for the first version, covering one page. The remaining time goes into verification — spot-check the first few runs to confirm it reads correctly.
How many pages a day?
Let it read the screen and write the sheet
Software is completely free and runs on your own computer. Start with a single page.