www.woshipm.com/marketing/1…

Ok, a clickbait headline, but there’s nothing inherently exaggerated about it. The algorithm does know how much you’re worth and decides what ads to show you.

Since The introduction of oCPM bid by Facebook, domestic headlines, wechat, Baidu information stream have launched their own oCPX bid, and oCPX bid has become the standard of mainstream advertising platforms.

In simple terms, oCPX is a bidding mechanism that allows advertisers to bid according to the cost of conversion actions. Assume that for game advertisers, they want to optimize the activation of the App. After setting the activation cost in the advertising system, the algorithm will automatically screen valuable groups according to the previous converted data and the bid of advertisers, and raise the bid for those with high activation probability to win the advertisement exposure. Reduce advertising waste by lowering the bid for people with low activation probability and reducing advertising exposure.

Other common advertising bid methods are CPM and CPC.

  • CPM: Cost Per Mille, Per thousand exposure Cost, that is, according to the number of exposure bid, such as wechat moments video form advertising in Shanghai according to ￥180 one thousand exposure deduction fee, regardless of whether you click on the AD. This is obviously best for the media in the short term because it doesn’t care if the exposure is valid or not; But why short term? Very simple, advertisers are not fools, if the subsequent conversion of advertising has been found to be poor, is bound to reduce the media.
  • CPC: Cost Per Click. If the AD is only exposed and not clicked, it will not be deducted. It seems fair, but advertisers need a lot of tests to find the best delivery method. In fact, some clicks are wasted, which is also a loss to the media.

So oCPX model emerged, it can help advertisers optimize deeper conversion behavior, at the same time improve the effective click, improve media benefits.

Recently, I found a taobao paper on oCPC on the Internet. The logic is very clear. Here is a simple paraphrase.

The article is full of details, but for those of you who are not interested, just pay attention to the following points:

  1. The ratio of this estimated CTR * conversion rate to the historical average CTR * conversion rate determines the coefficient of the system bid adjustment;
  2. Taobao paper, advertising to ensure the maximum eCPM and the sum of the interests of all parties.

Definition: first

  1. Bid: the original bid of the advertiser is B0, and the bid adjusted by the algorithm is B1 (as we often say: “the system will automatically adjust the bid”);
  2. Estimated conversion value p=pCTR*pCVR*v; V represents the average value of each purchase, which can be assumed to be constant; PCTR represents the estimated click-through rate of an AD, pCVR represents the estimated conversion rate (for example, the conversion behavior can be defined as “purchase”);
  3. Historical conversion value h=hCTR*hCVR* V; HCTR represents the historical click-through rate of advertisements, WHILE hCVR represents the historical post-click conversion rate. In the practice of Taobao, hCVR uses the advertising data of competing products in the past period of time (it is estimated to solve the problem of cold start; if the account has accumulated conversion data, it will definitely use the data of advertisers), calculates by pCVR estimation model, and takes the mean value of the highest and lowest 10%.

To simplify the illustration, make two assumptions:

  1. Assume that the advertiser’s goal is to ensure (or increase) ROI;
  2. Single click ROI = (pCTR * pCVR * V)/b0; B0 is the advertisement bid, such as the purchase GMV brought by the advertisement divided by the advertising cost, namely the ROI of the advertisement (to simplify the calculation method, it is assumed that the advertiser bid is equal to the advertising click cost, that is, the logic of generalized sub-high price is not considered, the generalized sub-high price GSP mechanism can be referred to the previous article).

To maintain or increase ROI, just ensure that B1 / B0 ≤ P /h.

B1 represents the bid after algorithm optimization; For example, assuming that this exposure is expected to bring 1.5x conversion value, the overall ROI will not decrease as long as the optimized bid/advertiser bid is less than or equal to 1.5x; Conversely, if the estimated conversion value is lower than the historical average, then reduce the bid; Since b0, P, and h are all known, then only B1 needs to be computed to satisfy this constraint.

In the real advertising environment, in order to give consideration to brand safety and account stability, the range of B1 / B0 is controlled within the range of [1-α, 1+α]. In taobao paper, the value of α is set as 0.4, that is, the highest advertising bid will not exceed 140% set by advertisers, and the lower limit will not be lower than 60% set by advertisers.

  • If P/H ≥1, then the lower limit l(b) of b1 after system optimization = B0, and the upper limit U (b)= B0 *min (1+α, P /h)) (because of ROI constraints, the p/ H constraint cannot be broken);
  • If P/H <1, then the lower limit of bid B1 after system optimization is L (b)= B0 * (1-α), and the upper limit is U (b)= B0.

Through the above constraints, we can also achieve what we often say: for the traffic with high conversion probability (P/H > 1, that is, the advertising conversion rate is higher than the historical average conversion rate) increase the bid, get a higher probability of presentation; On the contrary, for the traffic with low conversion probability (P /h<1), the bid is reduced to reduce the opportunity of presentation.

So the question is, if multiple ads compete, who wins the bid?

Advertising sequencing we all know that the use of eCPM sequencing, eCPM= B0 *pCTR, eCPM higher to win the opportunity to display advertising. Therefore, under the ROI constraint, the maximum value that eCPM can take is the upper limit u(eCPM)=pCTR* U (b), and the minimum value that eCPM can take is the lower limit L (eCPM)=pCTR* L (b);

The sorting mechanism given by Taobao is: sorting according to eCPM, while ensuring the maximum sum of interests of all parties.

To calculate the sum of the interests of all parties, Taobao provides two formulas:

f(1) = pCTR1 * pCVR1*V

F (2) = pCTR2 * pCVR2*V+β*CTR2* B0

F (1) calculates all GMV generated by advertising; F (2) calculates GMV brought by advertising and advertising revenue of the platform. It should be pointed out that f can be extended and therefore optimized for any goal; Here, f is considered to be a monotonically increasing function. As the bid b1 after system adjustment increases, the yield also increases. Although F (1) does not take B1 as a parameter, it is assumed that the increase of B1 can obtain better resource bits, thus bringing higher CTR and better traffic.

The AD sorting process is as follows (a little tedious, remember what each letter stands for) :

① Calculate F (u(eCPM)) for each advertisement in candidate advertising pool A, and sort in reverse order; In other words, f is calculated according to the highest bid. Since f is assumed to be an increasing function of B1, the calculation based on the upper limit of B1 can theoretically guarantee the maximum of F, that is, guarantee the maximum sum of benefits of all parties.

(2) Calculate the maximum of all eCPM limits t= L (eCPM) in all alternative advertisements;

③ According to the sequence, the first AD K satisfying U (eCPM)≥t is found, and the AD K wins the bidding and is removed from pool A; That is, as long as the ECPMs under the highest bid of an AD are greater than the ECPMs under the lowest bid of the other ads, the AD will win the exposure. This is to ensure media revenue, if only according to f without considering eCPM, then it may be a loss for the media business;

(4) Compare the eCPM u(eCPM) under the maximum bid of all remaining ads with the eCPM u(eCPM) of the last AD k, and take the minimum value between the two, so as to ensure that the eCPM of the AD K that has won the opportunity to display is the largest among all the candidate ads; If the eCPM upper limit u(eCPM) of advertisement I is adjusted, it means that the eCPM upper limit U (eCPM) of advertisement I is greater than the eCPM of advertisement K, and it is adjusted to the eCPM of advertisement K.

At this point, then the bid of advertisement I also needs to be adjusted, and the code of Taobao is more intuitive:

  • U (i-ecpm)=min(u(i-ECpm),u(k-ecpm));
  • Step 2: u(b)=min(u(b),u(i-ecpm)/pCTR);

The first step is to compare the upper limit eCPM of AD I with that of AD K that has won the opportunity of presentation, and take the minimum value between the two, which can ensure that the eCPM of AD K that has won the exposure is the maximum.

The second step is to update the bidding ceiling of AD I; If u(i-ECPM) is updated in the first step, then AD I’s bid is updated in the second step; Note that f(u(eCPM)) of AD I will also be updated (since f() is a monotonically increasing function, f() will decrease) as the bid cap changes.

Why adjust the eCPM for remaining candidate ads to be no higher than the cap for winning ads? Firstly, it is compatible with eCPM sorting mechanism, that is, it ensures the maximization of media revenue. Secondly, the consideration of f(u()) ranking is to ensure the interests of the platform, advertisers and consumers.

Repeat the above 4 steps until all ads have filled up or no ads have won the display, then end the loop and set bid B1 for all ads to U (eCPM)/pCTR, i.e., all ads bid according to the upper limit.

Release revelation:

(1) Purely logically, the real offer is the best strategy for advertisers. If the initial bid b0 is less than its true cost, then the expected traffic can not be obtained, which is unfavorable to the advertisers; If the initial offer b0 is higher than its real costs, it is good to take the amount, but can lead to a model will continue to the offer level near (p (c | u, a)/h (c | u, a) constantly tend to be 1), the late to control costs will inevitably need to accumulate data to study;

(2) Before starting oCPC, steady drop is required. In other words, the user conversion path works and costs as expected. For example, although the cost of activation is up to the standard, the subsequent transformation (registration, purchase, loan application, etc.) is not up to the standard. Even if oCPC is opened to optimize the activation, it is not meaningful, because the subsequent conversion rate of activated users will not change much.

This is taobao oCPC thinking details, we jump out of these details, take a look at an advertisement from the request to show involved in the whole process:

To make sure I don’t lose details, I just posted the description of the paper: After receiving the page exposure request, the Front Server passes the user information to the Merger Server, and the Matching Server analyzes the user’s characteristics (a series of tags), which are passed to the Search Node Server to retrieve the qualified candidate ads. The number of candidates dropped to about 400.

Real-time Prediction Server predicts pCTR and pCVR. Stragedy Layer contains oCPC logic and GSP mechanism. After this logical layer, the winning ads are optimized by Data Node Server and Smart Creative Service, and finally the Front Server returns the advertising elements and presents them.

The link to the original paper is here.

Everyone is a product manager. Reprint without permission is prohibited

The picture is from Unsplash, based on CC0 protocol