So, here are some problems with gathering that data:

1. There is no "X unique readers" data. We do not track uniques, and even a very entropic and brittle way of assessing uniques (using unique IP/UA tuples, say) requires data that isn't in the search logs. Very deliberately, our search logs contain no data that can be traced to individuals, and our logs that contain *that* data does not contain search information. Absent any way of sanitising, we would prefer not to surface this data because actually sanitising the search strings themselves is almost impossible.
2. Honestly, having hand-reviewed the high-frequency requests, I don't see the community getting a tremendous amount of value out of this. When we discuss "failed" searches we mean searches that produced 0 results - and most of those that do, do for a reason. We're talking automata or queries that are nonsense text. There is work we can do on the search side (like detecting languages and translating them, or localising the results) but knowing that people search for 'Barack Obama' in Hindi on en.wikipedia does not necessarily mean a redirect can be justified (unfortunately!)
3. The really high-frequency zero result requests are usually automata, not actual people.

1. Even in this case there are other ways to sanitize the data. Other possible option: whitelist keywords that already appear as red links (or as links), e.g

select pl_title from pagelinks left join page on page_namespace =pl_namespace  and page_title =pl_title where page_title is null;

2. It may or may not be justified to do manual redirect. If someone will go manually over such list he/she will probably find many other examples where redirect could be reasonable.

If keywords already exist as links, they most likely either (a) exist as pagetitles or (b) exist as terms within the article. In either case a result would be returned, and so it's not a failed query (in the sense of returning zero results)

It is not a technical fail of the search, but a fail for the reader - someone is looking for article that doesn't exist yet. Wikipedia community can use this data to prioritize articles for creation and anyway, it is interesting non sensitive data.

You're missing my point; suppose we include the words in the titles of redlinked pages as keywords, and we do not require that a search return 0 results to have failed. We then have several scenarios in the resulting data:

1. A user searched for a redlinked page and didn't find anything. Success, from our point of view.
2. A user searched for the topic covered by a redlinked page and actually did find what they were looking for in another article covered by the same keywords.
3. A user searched for something largely unrelated to the topic, but the redlinked page has some of the same words in it, so their query gets in the list anyway and gives a false impression of which redlinks are most needed.
4. A user searched for something largely unrelated to the topic, but the redlinked page has some of the same words in it, and the query also contained sensitive information we can't sanitise, so not only are we giving a false impression of redlink-filling necessity, we've also violated the privacy policy.

The only way of doing this to avoid that would be to require a query to *precisely match*, down to the phrasing and wording, the phrasing and wording of a redlink. Not only is this likely to be uncommon (Wikipedia's titling standards do not match standard human idioms for search, as our SEO results unfortunately reflect) it's also likely to be largely duplicative of looking at the pageviews for redlinked titles.

I don't know whether it is common or uncommon and whether the valid searches are representative sample or non representative sample of what the reader are interested in, but it would be nice to have even small chunk of the search data available to the public and this (red links white-listing) is very simple and clear method to sanitize. I'm not sure there is data of pageviews for redlinked titles, and if there is, I doubt it would match the same data for searches.

BTW it doesn't have to be precisely match raw data if the data isn't given as raw data but as processed/analyzed data:

1. It is possible to score links based on number of their occurrences in searches -e.g "My password is Barack Obama" results in {"Barack Obama": 1, "Password": 1, "My":1, "is": 1 } etc . [Other way to think about it would be to build a search index not on titles but on pagelinks, and see what are the most common terms the search engine suggest]
2. Co occurrence/ mutual information of substrings/words in terms appearing in other search terms would likely be a valid way to extract "native" terms from human idioms