Home > Paid Traffic Sources > POP / PPV / Redirect

---

Voluum - Anyone Used the Traffic Distribution AI? (27)

---

Hi Everyone,
I'm thinking of testing out the 'Traffic Distribution AI' on Voluum.

They have it at 3 levels
* Offer level
* Lander level (for advanced plans)
* Path level (for advanced plans)

The Pros of using it are:
* Spend less money and time to find out best offer/lander/path

The Cons of using it are:
* Probably less accurate then spending money evenly and waiting for statistical significance

I've used their 'optimization calculator' to manually help figure out which is the best lander or offer BUT I'm a little cautious since I'm not sure whether they start making judgement call too fast (which I've seen in other auto-optimization programs).

So, I was wondering if anyone had used it before and could give any insights on how well it did?

Especially or how much faster/cheaper it was then doing it manually AND/OR the loss of accuracy using this method?

thanks - Suneel

---

I've only recently started using it for Landing Page vs Direct Linking and at offer level and I think it's better than doing nothing / waiting for weeks to have results.

I haven't found a documentation on which optimization system Voluum uses, but I guess they use the one-armed bandit approach (like Google uses for Adsense)
- In manual optimization, I'd normally wait until I have a p-value of 0.95, and then send 100% of the traffic to the winning variation.
- With the system Voluum uses, they start sending more traffic (but not all) to the currently better performing variation.

This makes sense for me. It increases profitability because it starts sending traffic earlier to the variations that are probably going to win the experiment, even though statistical significance is not yet reached.

I'm not sure if there will be 100% of the traffic sent to the winning variation you have reached a statistical significance of 0.90, 0.95 or even 0.99.

---

Thanks Osmiuman - I usually try to check lander/offer/paths twice a day (i..e looking for 0.95 p-value) when testing but it does take up time when doing multiple projects.

I'll test out the auto-optimisation and see if I can cut back on the manual checking to maybe once every day or once every 2 days.
I'm always looking for ways to do more with less of my time :-)

---

I'm using this function.

I would say it works quite well with campaigns that are getting good amount of conversions.

Then it can do its job and autooptimize more effectively.

I still have to check it sometimes manually and do "hard-cut" of worst offers (because it will still be sending small % of traffic to them), but otherwise its doing its job just fine.

It takes into account just the last 24 hours though, so you really need those conversions coming in.

I did not find it as effective with COD nutra offers, where there is less of data in the range of last 24 hours.

---

Hey,
just to let you guys know that Traffic Distribution AI for Paths, Landers and Offers is now available on all plans.

Feel free to check out our blog post on Paths Automation.

Hoping to see more feedback here

---

I tried using it, but got back to manual from the most part. Until they come up with some solid tool to block bots, proper automation isn't really possible imo, as the system takes bot visits/clicks into equation and those are not equally distributed between campaigns. So some of the auto decisions simply arent done right.

---

Using it on some campaigns, only with CV calculation method... When you are buying from big, reputable sources that don't send bot traffic, then it makes sense to use AI feature. However, it's always good to check in manually and simply cut off completely bad LP/offer, even though the AI will give it certain small percentage.... No point in running any losers at all IMO.

But anyways this feature is just a temporary thing for me... just to find which lander/offer combo converts the best... Then I keep just the best offer..... Then if I want to run second offer on the same targeting, I make a different campaign (preferably with different banners and LPs)... Because these offers will most probably have different target CPAs, they will need different bids on the traffic source (speaking from the PPS methodology here).

---

Oh... Are you sure it's just 24 hours? Then this feature can only be used with massive campaigns (I'd say generating 20 conversions/24 hours is the minimum).

---

Yes, in Voluum manual it says: "In order to take advantage of fresh data, they are based on the last 24h data set.".

---

Exactly, the AI keeps on testing all of the offers by default, which means certain % of traffic is always lost to the poor performers. With a big amount of offers, it can account for a significant loss for sure. I mean, I understand it has do be this way, otherwise the system couldn't actually work ... but I prefer to handle it all manually and pause what I think needs to be paused.

---

Ok Erik and Forgamon, thanks for pointing me to this important detail.

I analyzed some data and came to the conclusion that the AI auto-optimization traffic distribution feature of Voluum is useless for me.

I think it can be helpful if someone drives a steady flow with really lots of conversions to different offers or LPs and is currently clueless about what LP / offer will perform best.

But my campaigns don't generate many conversions during a 24 hour period, and they sometimes stop. And if they stop for more than 24 hours, then all the past data for the "AI" is lost and the traffic distribution restarts evenly.

Take a look at these screenshots (modified to just show the relevant data):




Offer 2 was clearly better as offer 1 had a CPA more than 3x as high (payout is the same).
But V AI wants to send 83% of the traffic to offer 1.
So back to manual optimization

---

No problem.

I think @Voluum should add option to select time range which we want to include in the calculation, same as they do currently with their Anti-Fraud/MTTC feature.

And yes, you are right, this happened to me a lot as well.

Especially with COD offers, where there was sometimes longer delay as conversion depends from call-center operation.

It's working well though with offers that have more "instant conversion flow" ie. app installs, SOI/DOI, CPL etc. and many conversions coming in

To me personally it helped when offer started underperforming (for example because it reached global cap and was redirected to another offer and so on) or testing multiple paths and landing pages.

Later on I also still kill the worst performers and also switch to manual optimisation when I see a clear winner.

But in initial stages of the campaign this function is helpful.

---

Im using the AI now as well in order to optimize my smartlinks better. Currently I'm with Erik for simple flows seems it banking in. For long flows / call-centers conversions it not that useful.

Sent from my SM-G930F using Tapatalk

---

Yup, for simple flows and in cases where you get a ton of conversions, the AI would probably work better. But even with dating, where I get quite a lot of conversions, the AI tend to mess up a bit, especially when I add some new element to test it.

---

Thanks Matuloo for pointing this out.
I am a little confused though?

If overall you get 10% bot traffic, won't east of the paths and offers get 10% of their traffic as bots?
So, if you set the auto-optimisation to be based on on EPV or CV (not ROI), won't bots become less of an issue for optimisation purposes?

Or am I missing something?

thanks - Suneel

---

I'm trying to figure out best times to use the AI at the moment, so if you don't mind sharing ...
... how does it mess up if you add new tests?

Does it work ok if bots are not an issue and you turn it on right at the start of a completely new test (i.e. new lander/offer combos being tested).

---

Guys,
many thanks for such a useful feedback. Glad to see so many of you testing it and sharing your experience with others. We will definitely look into implementing the option to select time range.

Just want to point out that suspicious events do not have huge impact on your data - anyway - we do filter ‘good’ bots (google crawlers etc.) so it is not taken in calculations.

As for adding new elements to the campaign flow - we provide quite a solid mechanism for new offers added to the existing configuration - so we would appreciate more details here.

---

Does Voluum AI analyze each click and make a decision in real-time? Or does it just move the weighting of your paths/landers/offers around to the most successful?

---

In the ideal world, yes the % would be equally distributed ... and with a FUCKLOAD of volume, it's also probably going to happen ... more or less.

But with limited tests, it's not possible ... the % of bots on all placements is not equal, but more importantly, they are not coming-in in set intervals ... so it's not like 1 human click, 1 bot click, 1 human click, 1 bot click ... you can get a sequence of 20 bot hits, then 3 normal ones, then 2 bots, then 10 normal clicks ... try to distribute that evenly over 6 or 7 paths... it would take a lot of volume.

New paths will automatically get some traffic send to them, in order for the AI to determine their performance ... so the good paths will lose some as a result. Since the distribution of the traffic changes too, so can the performance of the originally best paths and the AI might start pushing the new ones... at least that's what happened to me a couple times.

---

Hey,

We're adjusting weights within the campaign/path so that the best performing offer/lander/path receives most of the traffic, and the traffic distribution is based on the last 24h of data.
It's worth noting that newly added offers initially might get more traffic, as we can estimate their performance, but since there's small amount of data, the range of possible CR/ROI/EPV will be very wide. For example - when you add a new offer and it has very little traffic initially, we could estimate that there's a 90% chance that CR will be somewhere between 3% and 90%, so it will be favored and start getting more traffic. However, once we accumulate more data, we can narrow it down - ie. estimate that there's 95% chance that CR will be between 3.1% and 3.8%, and thus the offer will start getting less traffic (assuming that there's another one with better CR, and that both have same payout).

Hope this make things clear now.

Feel free to ask more questions!

---

Thanks @Voluum for making things clear here!

I really enjoy this discussion!

As for questions -> I wonder what happens when we manually remove offer from the AI based path (as so far it's not possible to "pause it").

Will it restart the optimisation process again or it's smart enough to continue with the previously collected data in weight %?

---

Would like to know this too

---

Would also love to know if there is a way to set minimum conversions before it starts optimizing paths or offers.

Just had a case where one lander/offer combo got 5 conversions pretty fast and ended up getting about 3x the traffic of others.
After 3 days it is actually only the 3rd best CV% - and others might actually still be better but not getting enough traffic to them to figure it out.

This is probably great if you want to set and forget, but I really wanted to get enough traffic to all combinations to find the best lander/offer combo - so being able to say not to start optimise until the best CV% combo got to 10 converions (or some other min) would help.

---

For now we don't support it, but it's a good piece of feedback. Thanks a bunch! If you are up for discussing it with one of our Product Managers, just PM me and I will put you guys in contact.

It will not restart the whole process, it will continue distributing weights (the offer gets excluded from calculations). No harm to the algorithm to find the best performing elements and adjust their weights appropriately.

---

OK there is first time that Traffic Distribution AI screwed me.

One of my offers capped out, and it was redirected to some smartlink that payed out $0.01.

There were hundreds of conversions coming in, so AI decided to send there as much traffic as possible and feed that nonsense.

I used the "AUTO" option.

Thought it will see that those conversions are too small and the ROI and PROFIT doesn't add up.

Lost around $300 during night.

I can live with it and I do not blame Voluum nor affiliate network.

It's my mistake because I should be more foresight, but I still want to point out this issue.

In ideal case the AI should notice that even this offer converts well, it doesn't make profits so it shouldn't send so much traffic towards it.

Would like to hear your inputs, also from others how do you guys usually solve these kind of issues? (the cap is global and can cap out at random times)

---

Yeah, it's prone to errors and people can lose money that way.

The AI should have "Rules" feature, so that even though it is automatic path distribution, it should "cooperate with me" a little bit.

For example, if I have 3 LPs and I am disitibuting traffic according to the calculated AI Landers CV, then if the weights are:

LP1: 55%
LP2: 30%
LP3: 15%

then STOP LP3 (LP with the smallest CV).

In other words, I wanna make a rule, something like this:

"IF after 200 visits the CV of LP-n is the lowest out of 3 LPs in rotation and that CV% is at least 2x less than the next LP CV, then STOP that LP-n".

or maybe STOP LP-n, and instead of it, turn on another LP that I already prepared and put it into mix with other 2 LPs...

There are many scenarios and cases where Rules of this kind could be implemented with many dimensions and metrices (landers, offers, CTR, CV, ROI, Profit, eCPA etc etc...)

I am just giving 1 example from the top of my head... But yeah, implementing some type of Rules with the AI feature _might_ improve things..... Rules, something like those on Exoclick Bidder.

---

Hey Erik!
Thanks for reporting it, and yes, we can confirm we've identified a bug that in a very specific scenario would cause the algorithm to fall back from ROI/EPV to calculating via CVR - a fix to this issue was deployed yesterday.

Again, erikgyepes & forgamon, thanks a million for the continuous feedback, we keep a close watch on this thread and appreciate the discussion - all your suggestions are truly valuable to the development of AI in Voluum!

Karolina

---

Home > Paid Traffic Sources > POP / PPV / Redirect