My use case is a bit different -- I was doing a lot of database cleanups, particularly CRMs. I rewrote/reused code to build a duplicate detector a number of times; always wish there were a service that I could send data to, and it would flag my dupes. Even was using human labelers to train domain specific models.
Specific models might be an interesting addon. Address parsing, normalization, and deduplication (with potential covariates like phone number, email address, etc.) is a massive pain in the ass for any data engineer who works with sales or marketing folks. Their databases (CRMs) are awful -- it was always a chore to clean these up, but measurably saved money (imagine you mail physical cards, and only want 1 per customer... but you have 5 different contacts at that company for 3 unique individuals).
I would have paid for a deduplication service -- say, quarterly batches at somewhere >$500/quarter for e.g. 20-50k contacts.
The 1-size-fits-all isn't really a value add for me, that wasn't so much my issue. For other target users, I can see that use -- for them, the interface is the value add. Especially if you can read/write Excel files directly.
Stop words aren't something I used in my deduplication efforts. How many of your users request or use this? What kind of stop words do you want to exclude from comparing two entries? I would be worried that stopwords still carry information: "The Store" versus "Store" might be significant.
You have interesting insights, especially about the value deduplication offers. In fact, apps similar to mine charge in the thousands for an annual subscription. I didn't know how much of a pain it was until I released my first paid version last year.
As for the stop words: A significant number of my users either asked for it or presented me with issues that could be solved by removing certain irrelevant text. I don't track usage patterns and such so I added the feature based on this feedback, fully trusting that the users could identify the stop words correctly.
Some of these words might be anything from TLDs to definite articles like "the"; it all depends on the user. Sometimes it's a simple word like "very" or "best", etc., and, based on the data they've shared with me, it looks like the feature was a welcome addition to Flookup.
I used the Levenshtein edit distance to generate a list of potentials for human review - I extended MySQL to have a Levenshtein distance function for speed.