The most complete analysis of bidding for search ads!

Currently, the common search advertisements in the market are basically bidding transaction models, whether it is Baidu Fengchao, Google's adwords, or Taobao Express.

Last time, we introduced a search advertising product in the recruitment industry: Refresh. Refresh is charged by the number of times, not a bidding transaction model. Today we will continue to sort out another type of search advertising: bidding top.

I will explain the internal principles of this search ad so that everyone can have a general understanding of search ads and also understand how actual vertical industries, such as the recruitment industry, use search ads to monetize.

1. What is search advertising?

Simply put, it is the advertisement you see when you enter some words into the search engine to search.

For example, in the picture below, I have outlined the part in red, where a job seeker enters the search term “sales manager” in the search box.

After clicking search, you will see some advertisements for sales positions, as shown in the red box in the figure below. These positions are paid bidding top ads.

To put it bluntly, bidding for top spots is a bit like the process of selecting a girl group. A series of advertisements are like girls chasing their dreams. They go through auditions and assessments by celebrity coaches, and finally X trainees are selected to form a group and debut.

Advertisements also need to go through these levels of assessment, and finally the appropriate advertisements will be selected and displayed to users in order.

2. The decision-making process of bidding top

There are about 5 main steps in the decision-making process of bidding on top ads. I will try to express these decision-making processes in an easy-to-understand way so that everyone can explore the internal principles of search advertising together.

Decision 1. Girl group audition recruitment: query understanding

We know that before every talent show meets the star mentors, there will be open auditions and recruitment registrations. Advertising also has it, and this advertising process is called query understanding.

What are the specific steps?

Query cleaning, ignoring stop words. For example, we will automatically ignore or filter out some meaningless empty words in the query to save storage space and improve search efficiency. For example, "的", "也", etc.

Word segmentation: When the user enters [product manager], we need to segment the query. [Product manager] will be segmented into [product][manager] for query.

Query expansion: There are three main matching modes for query expansion:

1. Exact match: refers to the exact match between the user's query term and the advertiser's keyword. For example, if the user query term is [product], the advertiser's keyword must be [product] for this ad to be triggered.
2. Phrase match: refers to the user's query term completely containing the advertiser's keyword. For example: when the [advertiser's keyword] is [product manager] and the [user query term] is [senior product manager] or [senior product manager], this ad will also be triggered.
3. Broad match: When using broad match to place a keyword, the ad can also be triggered when the user's query is highly relevant to the keyword placed by the advertiser. For example, if the keyword placed by the advertiser is sales manager, when the user searches for sales representative, the ad will also be triggered.

Because the keywords placed by advertisers and the words searched by users may not necessarily match completely, in order to reach as many users as possible and not miss any possibilities, we need to do query expansion.

The purpose of query expansion is to automatically expand relevant query terms for advertisers and increase the advertisers' purchasing traffic.

Decision 2. Girl group audition: search for ads

After the auditions, we need to find girls who meet the recruitment requirements. The same is true for advertising. We need to find all the ads that match the user's query terms in the advertising library.

For example: If a user searches for "sales manager", then we need to find all advertised positions that meet/partially meet the requirements of "sales manager".

For example, in the table below: Both the advertisements [Sales Manager] and [Sales Representative] will be selected, but [JAVA Engineer] is completely irrelevant and will not be selected.

Which one, ad 01 or ad 02, is more relevant to the user's query? We intuitively know that ad 01 is more relevant than ad 02, right? So how do you measure relevance? This is when we introduce the [inverted index].

The inverted index is used by users to calculate relevance for us. We first create an inverted index.

Sales->[01,02]

Manager->[01]

Representative->[02]

For example, if [Sales] appears in two ads, it will correspond to two ad IDs: Ad 01 and Ad 02.

For example, if [Manager] only appears in one ad, it will have one ad ID: Ad 01

…

Relevance:

Relevance is a measure of how relevant [user query] is to [ad keywords].

To simplify the explanation, we give a simple relevance score formula: Relevance score = [number of keywords matching the ad query] / [total number of ad keywords]

Let’s take the example of user searching for [sales manager]. According to the above formula, the relevance score of Ad 01 is 1, the relevance score of Ad 02 is 0.5, and the relevance score of Ad 03 is 0.

The industry does not do this. It may extract some information from the historical click logs of advertisements. For example, for these queries, which ads were clicked? All this information is extracted and then manually labeled, as shown in the following figure. Each label will correspond to a relevant score. After the labels are attached, there will be a part of training data. At this time, supervised learning can be done, or more features can be used for supervised learning.

Decision 3. Girls Group Preliminary Selection: Advertisement Filtering

We have come to the preliminary round of the girl group. At this time, we need to eliminate a group of candidates. What kind of candidates will be eliminated? No topics or stories, no talents or ideals, not pretty, no body, no background...

The same is true for advertising. There are two main considerations for elimination:

• First: Is the advertisement relevant enough? Advertisements with low relevance need to be filtered out. For example, the JAVA engineer above is completely irrelevant to the [sales manager] that the user is searching for. No one clicks on it even though it is displayed, so the advertiser cannot get the desired effect, and C-end users see irrelevant ads, affecting their experience.
• Second: Is the bid of the advertisement high enough? Ads with low bids need to be filtered out. You didn’t even give me 100 yuan, you only gave me 1 yuan (Little Red Riding Hood face), and the platform still needs money to eat...

Decision 4. Mentor Assessment: Ad Ranking, Select TOPK Ad

After the preliminary selection, it is time to meet with the mentors. The mentors need to score and rate the contestants. The same goes for advertisements. The ads need to be scored and ranked, and ultimately the TOPK ads will be selected for display, where K is the number of ads placed.

How are the selected ads ranked? We generally rely on two dimensions to calculate the ranking score. One is the quality score of the advertisement itself, and the other is the price offered by the advertiser.

Ranking score = Ad quality score * Ad bid

How is the ad quality score calculated? This is a company secret and is generally not disclosed. We can give a simple formula to express this principle: Ad quality score = 0.8 * click-through rate + 0.25 * relevance score.

In general pay-per-click bidding ads, the click-through rate is calculated through a machine learning model, where 0.8 and 0.25 are the weights of the click-through rate and relevance score.

Why do we design ad filtering in the first place? In fact, this is to avoid the situation where some contestants are not strong enough, but have a strong sponsor behind them who offers a high bid and raises their ranking points.

When all ranking scores are calculated, we sort all ads according to the ranking scores. For example, if we display 5 ads, we will select the TOP5 ranking score.

Decision 5. Advertising Fees

The advertising ranking has been determined, but how do we charge advertisers? Should we charge advertisers based on their bids? No, we charge the second highest price.

Simply put, the highest bidder wins the item, but the second highest bid amount is used as the transaction price.

Why use the second highest price? Because the second highest price can ensure that the auctioneer will get the desired price. Because if the fee is based on the bid, the buyer will often bid slightly lower than the psychological price in order to protect his own interests. When using the second highest price, the buyer may bid at or slightly above the psychological price.

Here, I give a formula for the final fee, as follows: Final fee = (next quality score / current quality score) * next bid + 1

Let's take the example above. If you are currently ranked first and bid 100, and the advertiser below you bids 80, then the actual transaction price is 59, so we will charge you 59.

In our actual application, we did not give the buyer a bid, but provided the buyer with a price range, a minimum bid and a maximum bid. This maximizes the protection of the buyer's interests and saves the buyer the time cost of frequently logging into the system to change prices.

For example, if the buyer's bid is 71-100, 100 is the buyer's bid upper limit, we will first bid 71, if it fails, we will bid 1 yuan more until it reaches 100.

Conclusion

The above is the principle of search advertising: bidding at the top. Due to time constraints, I did not summarize the advertiser backend functions, including the rules for recommending bids, the rules for generating word packages purchased by advertisers (we do not sell keywords individually, but sell them as word packages), and the rules for recommending word packages.

If you are interested, I will write another article another day to talk about this specifically.

As product managers, we should focus on product strategy and mechanism design. How should we design at the product level in the future to optimize the bidding top product?

Like design price squeeze? Price squeezing is to more actively influence the bidding system to develop in the desired direction based on market conditions. Simply put, when bidding is very fierce and there are a large number of ads with high bids but low quality, the ranking can be influenced by adjusting the quality weight.

When the bidding is not fierce enough, you can increase the weight of the bid to encourage competition.

For example, in the recruitment industry, spring is the peak season, so the weight of advertising quality can be increased; during the normal off-season, in order to encourage competition, the weight of bids can be increased.