Method · How this site is made
How Softrecs is made, and where AI is used
Every number on this site comes from a prompt run in the public interface of ChatGPT, Perplexity or Google AI Mode, or from a study named on the page. An AI assistant, Claude, operates the browser for those runs and drafts the pages. Alauddin Chowdhury decides what gets measured, checks the figures and the text, and publishes. This page says which part is which.
Who is behind it?
Alauddin Chowdhury, a digital marketer in Dhaka with eight years on the revenue side of WordPress products. Softrecs is one person, not an agency. Every page is published under his name and he answers for what it says.
How is the data collected?
The questions are real. Prompts are taken from Google autocomplete, so they are phrasings people actually type, not ones written to get a good result. A re-run uses different real phrasings from the same autocomplete set, because Run 04 showed that repeating an identical prompt returns a cached answer.
The answers are read where a buyer would read them. Each prompt is put to the public interfaces: ChatGPT in a temporary, unpersonalized chat with search on, Perplexity, and Google AI Mode. No API is used, because an API answer is not the answer your customer sees.
Sources are read in full. The sources panel is opened on Perplexity and Google. On ChatGPT the conversation’s stored data is read, because the interface shows only a few of the sources it retrieved (Run 05 found 17%).
Counting is mechanical. Plugin names are matched against a fixed candidate list written before the run, and domains are counted from the extracted source lists. The per-answer results are saved as a CSV, and the figures on the page are computed from that file.
Where exactly is AI used?
| Task | Who does it | How it is checked |
|---|---|---|
| Deciding what to measure and what to publish | Alauddin | — |
| Running the prompts and extracting source lists | Claude, operating a browser in the public interfaces | Raw results saved per answer; a branded control query confirms the extraction works |
| Counting and analysis | Scripts written with Claude | Figures recomputed from the saved CSV |
| Drafting the guides and index entries | Claude | Reviewed by Alauddin |
| Code for the free tools | Written with Claude | Tested on real plugins before each release |
| Positions, prices, audits and client advice | Alauddin | — |
What is AI not used for?
Inventing numbers. Every figure traces to one of our runs or to a named external study. If a number cannot be traced, it is not published. Drafts have had claims removed for exactly that reason.
Quotes, reviews or testimonials. This site publishes no testimonials, and no quote on it is written in anyone else’s voice.
Volume. Each guide answers one question that people ask, with our own data or a named study behind it. Google’s spam policy on scaled content abuse targets pages generated in bulk without adding value, and that is not what the tool is used for here.
Why use AI at all?
Because the research would not exist otherwise. One person cannot read 45 AI answers and every source behind them in a day, and cannot see ChatGPT’s full retrieval set by hand at all. The assistant makes runs of that size possible. What it cannot do is decide which question is worth asking, or whether a finding is true. Those stay with a person.
Google’s guidance on helpful content suggests telling readers who made a page, how, and why, including whether automation or AI was used. This page is that answer for the whole site. Each guide and index entry links back to it.
What happens when something is wrong?
It is corrected in the open, on the page, with the date. The clearest example is Run 03, which said a pattern looked specific to Perplexity. Run 05 showed ChatGPT does it more often, and Run 03 now carries a dated correction linking to that evidence rather than a silent edit.
If you find an error, tell me. If it holds up, it gets corrected with credit.
Want your category measured the same way?
Send me your wordpress.org slug. I will run your category on all three engines, in the public interfaces, and show you what they read. Free, no call, usually within a day.