"He's on High and 23rd", "I live on Monroe", "Near Jomo Kenyatta Airport", etc.
The biggest mistake what 3 words made was not making the first two words be a geocoding of the approximate area and the last word could provide additional precision. It looks like google plus codes fixed this which is actually a huge improvement.
Approximate locations are dramatically more useful for navigating in much of the rural developing world. "How do I get to village X" will have an answer for 10 miles around. "How do I get to X lat/lng?" will be met with a shrug, both from google maps which will suggest you drive through a field and from anyone you ask.
If you can make something that works at the village level, you can bootstrap people getting their location. If the first word provided approximately "country" level information, the second provided approximately 1km bands (what you get if you evenly divide the three little words partitions), and the third gave you the 3m resolution grid, people without addresses could know their approximate address from others.
Basically, there's a bootstrapping problem and a usefulness problem. Right now, no one knows their 3 little words number. To become useful, most people need to know their 3 little words. How do you cross that gap? One way would be to have it be useful to know "2 words" or some approximate identifier for your region and then after locating you approximately, your service provider can tell you your last word.
The startup I'm involved with has developed what we call a geohash phrase. The core idea is to map English (or whatever language) words to the characters of a geohash. We usually map a word two-characters at a time, giving us 10-bits of precision for each additional word. So your location might be something like: "The big dog walks near the red house."
What is nice is that the phrase telescopes. So you can be vague when you want and only provide say the first three words of your location.
https://news.ycombinator.com/item?id=8614198
https://news.ycombinator.com/item?id=17423251
Also, I believe that they require large dictionaries, which makes it hard for apps to support it on low-end mobile. Wikipedia says ~10 MB for their database.