Introducing Predict – see who will convert, before they do | Hacker News Reader

Introducing Predict – see who will convert, before they do

mixpanel.com

32 pointsamitmathew15 comments

I love mixpanel, and this looks really promising. There's still a bit too much magic to my liking here. Not entirely sure how accurate this thing is... I asked their support how their algorithm calculates the grades, and this is what they said:

> "In essence, it analyzes users who have converted and not converted on the specified event in the past, and attempts to construct a few default profiles based on the characteristics of those users, both in their events behavior and their people properties. Then, it compares these default profiles to your actual people profiles, and places each profile in the bucket that that profile most resembles. So that's how we end up with the grades A, B, C, and D, and place users in those buckets."

It sounds good in theory, but I'd be quite interested in false positive / false negative rates or something a bit more transparent.

Still, this can potentially be a big differentiator for Mixpanel. I guess we'll try to target those 'high confidence' users and see how well they convert...

> It sounds good in theory, but I'd be quite interested in false positive / false negative rates or something a bit more transparent.

I believe it's machine learning, difficult to make transparent. That said it should be easy to test if changing your marketing to different levels of likely customers changes your total conversion rate.

Are there any details on data security/ownership here? If they are doing all of this lookalike modeling, are they using your audience data for other customers? Are they potentially selling it to Adobe's new audience exchange?

When you give an analytics company the level of conversion/revenue detail about your customers that one gives Mixpanel, there's a lot of cool things they can do, but there's always the nagging question of how much of that data do you really own. When it gets into modeling audience profiles, that gets really murky and I'd question whether our 1st party data is being being used by and/or sold to other parties.

Would love some official clarification on this.

It might be a better test to target the B and C users, and leave the A users alone. According to the algorithm they should convert anyways without your intervention.

As a conversion optimization consultant this is interesting, but I'm not jumping to pay for the extra $150/month (at least) for this until I see a convincing case study or experiment result.

Here's a low-tech way of increasing conversions: Talk to your customers, your sales prospects, and people in your target audience. Ask them what information they look for when looking for software (or whatever you're selling), whether they're able to find it on your site, whether it's convincing, do they still have any doubts and what are they, what next step would they take if they're interested (request a demo? call someone? email someone?), etc etc.

You may find, for example, that a lot of people misunderstand something critical about your product. Or that they prefer to talk someone before signing up for a trial. Or the pricing page wasn't clear to them. Or they didn't see a certain detail about your product and assumed it doesn't have it.

If you can't talk to those people, then talk to your sales team and get this info from them... They talk to prospective users all day so they should know about their concerns and questions. If they're getting the same question frequently, put the damn answer on your homepage, pricing page, contact page, ...

An algorithm isn't going to figure these things out for you, at least not yet.

(To turn the tables, I suspect Mixpanel would get a lot more upgrades for this feature if they included a case study or test result along with this announcement.)

There are too many products out there (probably less sophisticated than Mixpanel, I admit) that promise higher conversions with super-duper intelligent timing and segmentation... And most of them fall short.

If you're already paying for an event plan, you get 10 thousand user profiles for free. Gives you a little bit of room to test it out without spending the extra 150/month.

Huh... Interesting. Their public pricing page doesn't mention that (or I missed it), and the account upgrade screen (when logged in) doesn't mention any pricing at all, so this wasn't clear to me. Thanks for clearing that up.

EDIT: Seems the terms are different if you're on their "badge" program. I suggest just chatting with them and asking what it would cost, if anything.

I'm a Mixpanel customer and this is very interesting. Mixpanel suggests the main thing to do with this data is messaging the users that are likely to convert, but I'm more interested in finding out what are the indicators that makes someone likely to convert and optimizing for that.

This feature also supports my theory that analytics companies are really going to become broader marketing companies, so instead of focusing on just gathering and visualizing data, they will give you more ways of acting on that data.

It has been a while since I last spoke with them, but don't they charge based on audience size for their messaging capabilities? Seems like they and other marketing platforms are all trying to get customers to pay for messaging on their own properties (vs. email). I'm ok with paying for storing audience data (although portability/vendor lock-in are HUGE concerns), but paying for reaching my own visitors outside of email seems a bit much.

> analytics companies are really going to become broader marketing companies, so instead of focusing on just gathering and visualizing data, they will give you more ways of acting on that data.

This tends to happen when they run out of money or start looking to make more money. See UrbanAirship, Localytics and many more. Good thing: It works for them. Bad thing: Products tend to get more expensive and to expensive for the small/hobby guys.

I agree that the products get more expensive, although I like the idea of centralizing some of these data-driven marketing efforts through our analytics. Otherwise you have to buy 10 different tools and get them all to talk to each other.

> This feature also supports my theory that analytics companies are really going to become broader marketing companies, so instead of focusing on just gathering and visualizing data, they will give you more ways of acting on that data.

Soomla is already doing this in gaming with their Whales Report https://youtu.be/qj4x8GInMfE

If this really works (well), this could really be an awesome tool. Especially in connection with a notification, an in-app message, or email.

(By the way: The "people" features are an add-on that you have to pay by user, additionally to the engagement data points)

I've done this in the past with a pretty basic time series model. The accuracy was incredibly high and allowed us to forecast financial results months into the future.

am I the only one who expected some kind of intelligence surveillance project after reading the title?

Reply on news.ycombinator.com