It's the ultimate location to enjoy Hida beef, Japanese sake, and hot springs. On holidays, it's swamped with tourists, and the mountain city with a population of 90,000 welcomes 3 million visitors annually, seemingly recovering smoothly even after COVID. Indeed, when I visited, it was extremely crowded, and most on the trains were foreigners.
This project was initiated in response to the tendency towards overtourism in Takayama, where the charm of the adjacent Hida as a city has not been fully appreciated, posing challenges in creating buzz for neighboring Takayama city. Therefore, we decided to undertake this project using quantum technology to promote the attractions of Hida to visitors from outside.
Ad optimization has been a popular offering of our company for some time, but it has been underutilized for a reason.
The reason is that gathering materials has been very challenging.
Collecting advertising materials requires significant effort, so it was a technology used when budget and opportunities allowed. However, with the advent of generative AI, it has become easier to acquire these materials! Thus, we are reutilizing it.
It's possible with D-Wave, but since it can also be utilized with quantum gate computers, we plan to modify the model for an advertising optimization model suited for quantum circuits. The process will be as follows:
1. Automatically generate ads and learn from the effectiveness measurement to improve generation accuracy.
2. Prepare content that conveys the attractions of Hida as the destination of the ads to facilitate conversions (in this case, visits to Hida).
While the technology can also automatically generate websites, we will start by generating ads to proceed in an orderly fashion.
Quantum annealing utilized models based on a probability distribution known as the Boltzmann machine or Boltzmann distribution. However, with quantum gates, a model called the Born machine, which has enhanced expressiveness, can be used.
The mechanism is simple. The base materials for the ads are generated using LLM (Large Language Models). This time, the photos were provided by a local cooperating company. We aim to stimulate the demand for tourism in Hida and identify the target audience.
For this campaign, we partnered with a local tour company, Happy Plus.
[https://happy-plus.co.jp/](https://happy-plus.co.jp/)
There were several conditions and notes for this project:
- The quality of the advertisement was largely dependent on the photo and layout, with the layout being fixed.
- Photos were provided by a local cooperating company, knowledgeable about the area's attractions.
- In terms of attracting customers to the ads, combinations were considered, so the wording was not made overly complex. While it's possible to tailor the wording to match the photos, this was excluded to enable comparisons.
- The banner colors, among other elements, were all determined by LLM.
- The wording was also decided by LLM.
- The selection of photos was done by human judgment.
As a result, the system automatically generated 20 ads.
The prompts for generating the text were kept general:
"A short phrase to encourage clicks for an advertising banner intended for tourist attraction."
"Specify five nice, somewhat bright colors with their color codes."
This way, we could specify the length of phrases and colors using color codes.
The logo is fixed, located in the top left, which I made myself. This fixed approach is applied to all corporate banners and such.
Surprisingly, the ads seemed to reach the minimum quality required for advertising. Some ads appeared almost identical. These could have been excluded by calculating similarity, but we decided to submit them as they were this time.
Additionally, some ads had overlapping text layouts. Although we could exclude these based on text length, we kept them to assess their impact on the ads. Of course, undesirable ads could be pre-excluded manually.
The choices were relatively limited this time:
5 types of main text
5 types of background images
5 banner colors
5 banner texts
3 button texts
This results in only 1,875 combinations, so the ad's search space isn't very broad. However, the aim was to quickly and cost-effectively launch the ads, so this was deemed acceptable. Looking forward to the results.
The main texts were:
- Experience something unique here!
- Your next adventure starts here
- A journey through history
- In search of breathtaking views
- An unforgettable vacation with the family
Banner colors were:
- #87CEEB
- #F88379
- #40E0D0
- #E6E6FA
- #FFFACD
Banner texts were:
- Fulfilling your dream journey
- Let's find beauty
- An adventure to remember
- A journey to discover hidden gems
- Exploring tradition and charm
Button texts were:
- Check it out now!
- Go to site!
- A limited-time opportunity!
I'm excited about the results. Now to launch the campaign.
The conditions were as follows:
- Google Display Ads network
- Submitted 20 banners
- Max impression bid 100 JPY (I confused it with the click price)
- No target setting
- Limited to Japan (an oversight)
Overall campaign results:
- Executed on 2024/01/22
- Budget 14,500 JPY (could have been less)
- 161 clicks
- 546,227 impressions
- Click-through rate 0.03% (low!)
- Average cost per click 90 JPY (high!?)
- Average cost per impression 45 JPY
- Completed 546,000 impressions in just minutes? It was quick.
Let's look at the results. I'd like to delve into the details on another occasion.
The best-performing ad was...
The number of clicks was 13, the number of impressions was 27,118, and the average cost per click was 56 JPY.
The number of clicks was 4, the number of impressions was 27,404, and the average cost per click was 179 JPY.
It's fascinating how different combinations of the main text and banner colors, even with the same photo, can yield different results! The fact that nearly identical advertisements can have a threefold difference in cost per click due to different combinations is intriguing. With a similar number of impressions, fewer clicks mean the revenue from the destination could potentially be three times worse.
Moreover, since the overall average cost per click was around 90 JPY, averaging out is not efficient. There were several high-performing ads at the top, so by focusing capital on those, advertising costs can be significantly reduced and used more efficiently.
From this trial, it was found that focusing the budget on the top-performing ads could save up to 39% in advertising costs.
This is very pleasing. The beauty of this system lies in its ability to analyze outcomes with a significant degree of explainability, despite the learning process being somewhat of a black box and lacking explainability.
The user demographics for this campaign were roughly as follows:
- The age group 25-34 years old had about 50 clicks, 35-44 years old had 30 clicks, and 18-24 years old had 20 clicks, indicating a younger audience predominated.
- By gender, males had 100 clicks and females had 50 clicks.
- Regarding families, 40 clicks were from those with children, 20 clicks from those without children, and others were unspecified.
This suggests that younger individuals were more engaged, and perhaps phrases targeting families did not resonate as much as expected. The output results are not a black box but come from combinations, making them very analyzable.
In this presentation, I introduced a system that uses our company's ad optimization technology to analyze tourism promotion and attractive advertisements for users all at once. Recently, generative AI has become available for material creation, allowing for the quick and inexpensive production of advertisements. However, without knowing if the created ads were effective, we deployed them on Google Ads and used the number of user clicks for feedback. As a result, we achieved a 39% reduction in the advertising budget and did not need to spend time on creative efforts. Quantum technology also entails no cost for ad generation since calculations are completed instantly, making this a great example of cost reduction using quantum technology.