Microsoft, Facebook, and Slack have introduced bot APIs. Ray Kurzweil has announced that Google will also introduce a bot later this year. This made me wonder what features a fully-functional Google bot would possess, and how it might support Google’s business model.
Summary of Predictions
Google search is presently characterized by two types of responses to keyword queries: ranked web pages and advertisements. Search will expand to include natural language dialogue responses to natural language queries. (There is not a black-and-white distinction between natural language queries and keyword queries. The responses could be a similar blend.) (See Google Assistant.)
Google search presently exploits context that is implicit and not volunteered by the user, at least not at the time of query. This has the benefit of “magically” relevant responses, but also has a “creep factor.” Natural language dialogue will consume more of a user’s time, but have the benefit of explicit, immediate, and voluntary context. This in turn will be rewarded with precise, relevant results.
Google search dialogue could differ dramatically from human-to-human dialogue if it presents multiple responses, allows the user to choose the course of the dialogue, and improves over time based on these choices.
Google will continue to promote schemas and standards for authors to explicitly structure content for easy, relevant access. Google will continue to pursue natural language processing algorithms to automatically structure text that remains unstructured.
In the same way that the look and feel of site search can be customized, it will be possible to customize the personality of the dialogue aspects of site search.
Presently, ranked web page results sometimes overlap with advertisements (sponsored links). Similarly, natural language dialogue responses may direct users to web site destinations that are also advertised. Such directions will have a higher closure rate (revenue per impression) because of explicit, immediate, and voluntary context.
Google presently allows advertisers to associate keywords with advertisements, and encourages expansions to ensure that all relevant keywords are included. Google will allow advertisers to associate some form of dialogue structure (for example, intents) and encourage expansions to ensure that all relevant ones are included. Advertisers will tolerate the additional labor because of the increased revenue per impression.
The Importance of Context
Google’s mission is to organize the world’s information and make it universally accessible and useful.
Increasing amounts of “the world’s information” are certainly becoming available online. However, Google recognized early on that accessibility, usefulness, and even organization were highly dependent on the time, place, and other aspects of the context in which it was being accessed. By accounting for the time and place of your search, the recent history of your searches and even searches by all other users, Google is able to better filter and sort results, frequently ranking the most relevant result at the top. This is by no means an easy task, and in the beginning it seemed magical. It is now very much expected by users, who can become frustrated when the results fail to satisfy their immediate demands. However, use of context can also trigger the “creep factor”—that unpleasant but common suspicion that “Google knows everything about you.”
Accessibility of information was also frequently degraded when the answer to a specific question was buried in the text of a web page. Google has made strides on this front by returning “snippets” at the top of the results page—snippets that could have been extracted automatically by Google software or defined by web page authors themselves. Of course, snippets are only helpful if they correctly address the specifics of a search query and the context in which it was made.
Custom search could be thought of as a specific kind of context that filters search results to a single web site.
Dialogue for Context
One way to acquire context is through question-and-answer dialogue. This is a very natural mode by which humans answer questions, and avoids the creep factor—you, the search user, do not feel the other party “knows everything about you” because you have supplied information voluntarily and recently during the course of the dialogue.
A very rudimentary form of dialogue was introduced by Google with the “Did you mean?” feature. If you use an unusual spelling in your query, you are presented results as usual, but at the top of the page is a link titled “Did you mean [expected spelling]?” That link lets you re-do the search with the expected spelling, in a single click.
More recently, Google introduced anaphora resolution in its voice search. You can (verbally) ask the search engine, “Who is the President of the United States?” and get the correct result, “Barack Obama.” You can then ask, “How old is he?” Google will correctly interpret “he” to mean “Barack Obama” and return Obama’s age.
Natural language dialogue does consume time compared to fast keyword search. It will never replace keyword search, but it does provide a useful and natural extension. This is especially true as Google asserts Android’s dominance on smartphones. Forbes recently published the results of a Gallup Poll: “Text messages now outrank phone calls as the dominant method of communication among Millennials.” It is reasonable to conclude, therefore, that chatbot dialogue is a natural next step in Google’s pursuit of its stated mission. (See also: Apple Lays the Groundwork to Kill Online Advertising)
Show Me the Money
Not everything Google (or Alphabet) does has a direct impact on the top line, but search does:
Advertisers pay Google to present advertisements along with “organic” search results when certain keywords are queried
The more an advertiser bids on a keyword, the more likely the advertisement will be presented when that keyword is used in a query.
An advertisement targeted for one or more keywords may or may not be presented (an event called an impression) when those keywords are queried. The advertiser is only charged when the user clicks on an advertisement (pay per click or PPC).
The advertiser can easily track when the user clicks on an advertisement, and even when that clickthrough results in a sale, download, or other action
Google provides tools for advertisers to construct advertisements, target keywords, count impressions, and track clickthroughs.
Web site owners can earn money from Google by become advertising affiliates. They reserve space on web pages for advertisements to appear. Google runs an automated competition for that space, choosing a relevant advertisement based on the content of the page, the advertiser’s bids, and the visitor’s context. The user’s visit to the page becomes a kind of implicit query whose result is the advertisement that appears in the reserved space.
Advertisements could be thought of as a class of search results whose “relevance” is determined (partially) by money. Sometimes the results overlap, and the top advertisement links to the very same page as the top “organic” search result. However, the ideal for all parties is when the advertised product or service truly meets the need implicit (or explicit) in the user’s query.
What does this mean if keyword search is extended to chatbot-style dialogue, or if chatbot-style dialogue becomes the dominant mode of search?
Search as Chat
We can certainly imagine queries being processed in two ways simultaneously:
as keyword searches, just as they currently are, with results being web pages. Of course, multiple web pages may be relevant to a keyword query, and they are ranked according to Google’s proprietary algorithm. When a user clicks on one of the returned links, this choice is used to improve the algorithm’s ranking in the future.
as natural language chat statements, with the result being the answer to a question, the result of executing a command, or a clarifying question.
The fact is, in response to a natural language chat statement, a multiplicity of responses could be considered. Differing assertions might be present in the Google knowledge graph and in various web sites. These responses might be ranked the way web pages are, and the top-ranked response might be chosen for presentation. This is analogous to what happens in a human dialogue. However, the search page affords an opportunity to do something different: unlike human chat, Google chat could return multiple ranked responses. The user could choose from one, thereby taking the dialogue in a particular direction. This choice could be used to alter the search system’s ranking of responses in the future. (See Choose Your Own Adventure.)
The number of natural language responses (vs. the number of organic search results) presented could be higher if the query consisted of a parsable natural language question or statement.
To the extent a knowledge graph can be extracted from text on a web page, Google seems to do so automatically, using techniques like information extraction but a special importance should be given to knowledge explicitly structured by web page authors.
When a multiplicity of answers is given to a natural language query, it will be important to communicate their sources. Popular answer sources could be considered more trustworthy than unpopular sources.
The chatbot equivalent of site search would converse based on the knowledge graph in a single web site. The results would provide the “perspective” of that source on a topic of conversation.
When a query is processed, a (potentially large) set of advertisements “compete” to be among the few presented on the results page. The algorithm that selects an advertisement is proprietary, but considering that advertisers associate keywords with their advertisements, we can safely assume that if an advertisement is associated with a query keyword, it is included in the initial set. (After that, it may get filtered out for various other reasons.)
The risk here is that the search user may, for various reasons, use a keyword that the advertiser did not anticipate. The user misses seeing a relevant advertisement, and the advertiser misses a potentially lucrative impression.
Google mitigates this risk on the advertisers’ side by suggesting expansions of the keyword list associated with an advertisement. There appears to be some expansion happening on the user side as well, because the web pages in search results sometimes contain only synonyms, and not the keywords themselves.
The point is that the match takes place between a derivative of the query and a derivative of the advertisement. These derivatives may be specified by the author (as in advertising keyword expansions or structured snippets) or they may be extracted automatically by Google algorithms (as in query keyword expansions).
We can certainly imagine treating statements in a dialogue as queries for the purpose of selecting advertisements, using the same algorithms as for keyword queries. However, irrelevant advertisements inevitably appear, at least partly because the context used to select advertisements is not obvious, immediate, nor in the user’s control.
It is more interesting to consider that a properly-structured dialogue—which is explit, immediate, and within the user’s control—could lead to a product or service that perfectly meets the user’s needs. The increased likelihood that an “impression” leads to satisfactory resolution would compensate for significant incremental effort on the part of the user (in engaging in dialogue), the advertiser (in structuring dialogues, analogous to associating keywords and advertisements to products and services) and Google (in automatically generating derivatives to close the gap).