A benchmark and a poll are apples and oranges. A benchmark is measuring a reproducible result based on the performance of the part. A poll is taking a sample of what is out there and is subject to an error variance.
Take for instance presidential polls. Those have error ranges upto 5-7%. We are talking a possible .5-1.0% error rage here. That is miniscule. People's GPU's that dont show up in the survey or represents too small of a sample to show as an independent data point fall under the title of "other". "Other" represents .74% of DX11 GPU's as of April.
You're assuming the complete dropoff of one card is representative of the error margin of the survey.
It's not. I cited it as an example of the unreliability of what we see. The error margin is much larger when looking at Steam users vs the overall population.
So basically the Steam survey has a margin of error of at least 0.54% we can assume, meaning that anything with low percentages is unreliable.
We also know that Steam doesn't represent overall sales, because it has a ridiculous margin of error when compared to relative sales numbers in cards which all cost >$100.
We also know it simply fails to be even remotely accurate in some situations, e.g. Crossfire detection vs SLI.
Some cards simply don't show up at times even when they were released concurrently with other cards for some reason.
So we can't use it to compare cards with small % representation.
We can't use it to give overall market stats at a $100 price floor.
We can't use it to demonstrate differences between multi GPU adoption.
We can't use it to reflect sales of brand new cards.
So what it it useful for in any meaningful statistical way?
Maybe looking at broad patterns of adoption of new technology (e.g. DX11 systems) by gamers who use Steam, and a broad idea of the general level of hardware which the majority have, but it doesn't even give an accurate month to month snapshot because of the random nature of the selection and the selection bias of Steam users, and it doesn't accurately reflect sales between NV and AMD or within NV and AMD product lines.
It is not ACCURATE at all, it can be used to gain a broad picture of the overall direction the market is taking and the overall potential level of hardware that might be broadly found in end user computers, but it isn't ACCURATE to any degree, especially not in the way discussed by the OP, and nor in the way suggested by most of the Steam hardware survey threads that pop up.