Hacker News new | comments | show | ask | jobs | submit login
A/B test of banner ads vs. traffic (gwern.net)
320 points by luu 8 months ago | hide | past | web | favorite | 141 comments

What I learned from running hundreds of forums is that:

1. Banners hurt unless they are relevant

2. Filtering for banners that are relevant tends to reduce revenue rather than increase as there is a smaller pool of highly relevant advertisers

3. Removing all banner adverts significantly increased engagement by every metric

4. Some revenue is possible using affiliate links within the content (re-writing existing links rather than inserting awful links on a word match basis)

That last one is what I now do. Is it as profitable as running banners everywhere? No. But it is the highest possible engagement with some revenue, and if one focuses energy on reducing costs instead it can work well.

Up until a week ago, I had an Amazon banner on my blog. When I set it up (about a decade ago), you had the ability to specify a search term in the URL of the banner, and I used this. For each post, I would have a targeted search term to use to display the Amazon banner.

Sometime in the past month or two Amazon got rid of that feature and so I was only getting a generic Amazon sales banner. I'm not sure why they changed, but they did, and I removed the banner. Thanks Amazon.

They probably determined your page-relevant term was just getting in the way of their killer machine-learning tuned strategy of showing ads for the last thing the user happened to look at.

Or better, slight variations on the most recent purchase

If it was a banner of products related to previous purchases, okay, I can see that. But the banner was just a generic "Shop at Amazon!" Lame.

Do you use Ghostery or other “do not track” tools? It’s possible your users were seeing something different

Nope. I have no ad blockers installed. I also checked the Amazon affiliate website and the new banner ads (at the size I wanted) no longer have the option I was used to.

Thanks for sharing. I really wish there was a general web subscription with revenue-sharing between all visited websites. I hate seeing the ads on my own sites but have no viable way to replace them. And as a user I would like to support several sites, apart from disabling the adblocker.

It wouldn't work anyway. To make up for the advertising revenue, that subscription would have to be in excess of $100/month per household, and you'd need to get virtually every household in wealthy nations to buy into it. The amount of marketing money spent each year is on the order of 10-11% of all revenue made by companies selling consumer goods... you can't replace that with some $10/month subscription even if everyone paid it.

I think your estimate is off by more than an order of magnitude. I follow a number of large youtube channels. Many of them seem to be making more money from patreon than from ads. And patreon isn't even a true subscription model, since you can usually see the content even if you don't subscribe. And it still beats ads.

One of my favorite channels is this super niche thing that only gets about 1000-2000 views per video and makes 2 vids a month. That adds up to maybe $4-8 in ad revenue using some rough estimates of youtube's payout rate. On patreon he's making over $500 a month. Which is really respectable for such a small audience.

Online ads have never generated that much money per user and were only profitable in large quantities. Given how scummy they are and this evidence that they repel users, they should be eliminated.

If you divide US digital ad spend in 2016 by number of US households, that's about $50/month. Same order of magnitude as my earlier guesstimate.

Sure but that's a very indirect calculation. Only a fraction of that is going to end up funneled to banner ads and the websites that run them. And that's the relevant metric; how much revenue you are losing by moving away from ads.

On a first order that's right: We don't need all the ad people now that there's no ads.

But then we have to consider that 95% of the costs/fees for running ad networks would still be needed for cost-sharing models too. You still need third-party JS.

Patreon doesn't need js. A subscription model doesn't need js. Certainly not all the complicated tracking code and huge images and videos required for ads.

I was talking more about a model that works similarly to advertising (pennies/fraction of penny per page view)

I'm more of a fan for direct payments, though I think a lot of sites would struggle with just that.

How many people will subscribe to Android Authority directly? Lots of sites basically get views from searches , but little long term engagement. Those sites are still useful, but only have ads as a revenue model that's "viable"

Something's not right with that equation. You're implying that on average, each household is spending over $1000 per month on things they get via the online advertising that they would otherwise not get at all.

And that's just not realistic. One-off purchases, maybe. Sustained 1k/hh/mth avg - I'd love to see a proof of that.

Total advertising spend worldwide is $600 billion. Online is about 1/6 of that, call it $100 billion.

To a rough approximation, that's footed by the 1 billion wealthy inhabitants of the EU, US, Japan, Canada, Australian, and New Zealand. (I said rough, roll with me.)

That means that the per person costs of advertising are pretty much $100 annually online and $600 annually for all content (mostly TV). Keep in mind, those are real expenses expressed through household purchases.

(This means, by the way, that online content isn't free, you are paying for it, only, you're doing so indirectly, through advertisers.)

By contrast, total direct expenditures on print media run about $125/person in the US. (Audio/video may add to that.)

If you look at what is advertised, an awful lot is very high-ticket items. The FIRE industries (finance, insurance, real estate) were the largest chunk of spend as of a couple-three years ago. Electronics is another large segment.

The notion of some sort of universal content payments scheme, preferably indexed to wealth and income, has a great deal of economic justification. I've been exploring that for the past few years.


So, if I block advertising, I'm indirectly wasting the money these companies spend on me, and are paying more for products without any benefit? That's a funny thought.

You're also avoiding the inherent manipulation of advertising itself.

But yes, the dynamics are a tad bent.

I find it very disappointing and would like to see the ad industry die in a fire :-(

Households do spend >$1000 per month on consumer goods/services (food, clothes, personal care items, etc). Those companies do spend >10% on marketing. I'm not sure what portion of it is spent online, which is what you'd need the subscription to replace.

There's no implication that every dollar they spend on marketing directly translates into $10 of new spending "that they would otherwise not get at all".

Much of it is brand spending. Procter & Gamble is the world's largest advertiser, and the goal of their ads is to make sure you keep picking up Tide every time you're in the detergent aisle. Coca-Cola's ads aren't trying to get you to buy $10 more of soda than last week, just to keep soda on your shopping list every week. Those ad dollars would need to be replaced by the subscription if you're going to remove the ads but keep compensation for the websites the same.

That $100/month is what you'd have to get every household to pay to get rid of all advertisement. Just getting rid of banner ads on websites should be a lot cheaper.

It's possible many ad-supported websites deserve to die... as I see it, they are economic parasites. Advertisers accept bad leads as a cost of doing business, but the cost is simply passed on to consumers.

Many low-quality news papers, TV channels, and magazines died when ad revenue shifted online - consumers did not see enough value to justify the unsubsidized rate. Others survived because consumers saw enough value to pay the higher rates. The internet isn't any different... a large chunk of the internet is spam surviving off of wasteful advertising revenue. Soon, the cruft will die off.

None of us need unlimited access to everything out there, and I don't think it's a sustainable expectation.

I think Mozilla ran the numbers a year ago and found that all the advertising on the internet only nets about ~$11/person/month. Though, hell, I'd pay 10x that to be rid of that malware vector.

That sounds like a content tax.

And is probably precisely how this should work, indexed to wealth / income.

How do you then decide who gets paid to make content?

A fair question, and I don't know that I've got the answer, though several answer-shaped possibilities suggest themselves.

One line of thinking has a Neilsen or Arbitron-like rating system that pays based on readership. I think that is ultimately a false lead as it simply rewards popularity.

Another is to employ a basic income approach. Given the number of writers I know who write largely because they can't not, this might work, though it might not be fully sufficient. The prospect of a UBI and some achievements-based revenue might make this more viable: electronic distruibution is gratis, but live performances and physical media might generate additional revenue, as might various services or other arrangements, all on top of a UBI base.

A third way might be to treat creation something like research: it's an activity, and it's funded as a certain overall percent of national income. Divide that by what's seen as a livable wage, and designate that there are so many positions open, and so many to be filled in a given year, that will be paid. Sorting out how to filter out those who really aren't suited to it is the next challenge, though there are various models: guilds, professional recognition (see the Emmy's, Grammy's, Oscars, Pulizter, Peabody, Nobel, Royal Society, AAS, and similar recognitions).

One thing I don't think should happen is for there to be some sort of quarterly or annual "rank and yank", or the general fight over scientific grants funding we now have. There's an article I'd recently seen about productivity of scientists over their careers, and the distribution is ... almost perfectly random. That is, a given scientist might produce some great work early, late, or in the middle of their career, once only or possibly several times. But there's not much rhyme or reason to it.

Creative work generally is something that evolves over time, and takes time. Days, weeks, or months even for a good piece of journalism. Years for a story, play, or film. A lifetime for works of science or philosophy. Take Wittgenstein, who wrote one book and then spent the rest of his life writing another refuting the first.

(And yes, there are prolific examples as well, though often they are best known only for a handful of works. Attention is a zero-sum game. Alexandre Dumas. Isaac Asimov. There's a large literature on Gresham's Law effects in literature, painting, music, drama, and television as well.)

The inputs to writing are information, sources, and investigations. Otherwise, generally, a quite space and not having to worry too much about room, board, heat, water, and light. Excessive remuneration doesn't much seem to help in its production, and a certain frisson may actually help, but only to a point.

Have you tried Flattr, I've not used it but I see it on a lot of sites and gather it fits that niche.

https://thenextweb.com/tech/2017/10/24/flattr-now-lets-pay-c... may be of interest too.

pbhjpbhj is right to mention Flattr.

Patreon is also a good idea.

Sponsoring is also possible if you can find something relevant and willing: a lot of the podcast and YouTube channels that I follow have audiobook platforms sponsor them.

I have Patreon and a Pay Pal tip jar. I still have ads on some sites. But I have fantasies that I will someday have enough revenue from other things to justify removing ads from most sites.

The Brave browser people are working on a weird cryptocurrency-based solution to this problem called the Basic Attention Token. Hope you don't mind being paid in some Ethereum tokens, though. https://basicattentiontoken.org/about/

Brave is crazy; go with Flattr instead. They both want to distribute monthly subscriptions from users to content creators. Flattr pays in real money and is easier for users because they can use credit cards instead of virtual tokens and hard-to-fill-up crypto wallets. Flattr works with creators on YouTube, GitHub, Vimeo, etc, etc. but Brave only works on the domain-level; meaning only Google can benefit from YouTube with Brave whereas individual creators get the money with Flattr.

There's Google Contributor, which is trying to make another comeback. One of the newspapers in Australia is using it, and supposedly Business Insider as well.


I've looked at Contributor and found it weirdly limited. Google doesn't need publishers to sign up for Contributor, because they are already all signed up for AdSense! The most logical way to implement Contributor would be essentially as a personal buyout of AdSense ads: if you are a Contributor, the AdSense JS checks that and then simply doesn't run, and assigns a fraction of a cent to the publisher in lieu of a displayed ad. Everyone gets what they want and this immediately gets you coverage of millions of websites instead of, like, 10.

That's how the original Google Contributor actually worked, it would remove Adsense ads from all websites running AdSense. You could even choose what to replace them with (eg pictures of kittens). Matt Cutts posted about it here in 2015:


Apparently that model was a huge flop and people just didn't want it. The new version requires publishers to remove all ads (not just AdSense), which helps since many newspapers seem to have dropped AdSense & switched to Taboola & other ad networks. Which might be one reason Google is trying it again.

(I expect the new Contributor will fail too, though. I think the majority of people just don't want to pay for online content.)

That's a pity. I expect the new optin model to work much worse; not sure why they're bothering... Patreon, IMO, shows people are willing to pay for online content but the exact method and experience matters a lot.

How do you have experience running hundreds of forums? :O

I've been running forums since 1996 using almost every major forum software that has existed. I'm currently running my own software (which we tried and failed to commercialise), and the largest forum I run is https://www.lfgss.com/ with 250k monthly unique visitors, but the long tail is very long.

We actually now survive more on donations from forum users than any advertising. About 2/3 revenue is donations, 1/3 from affiliates.

Affiliates work really well, but the rise of cash back sites mean that the revenue per sale is dropping. Affiliates work well for delivering new customers, but cash back is leading to retailers paying out for retaining customers which means the cost is shifting from acquisition to retention, which is traditionally where they make their profit. Hence cash back is leading retailers to make the terms far more strict, shorter cookie periods, lower %, etc.

Donations have been the saviour. As I'm running the sites as a platform, the larger sites effectively subsidise the young and growing sites.

Also, as a platform: 1 volunteer (me) runs all of the sites. Moderation is by communities, and has proven very effective whilst not requiring any additional cost. Though I did invest in some good lawyers for the T&Cs to make sure that approach kept the platform free of liability from individual sites on the platform.

Step 1: Run hundreds of forums

As someone who has founded and runs an adserver for a living (we deal only with legitimate publications, nothing sketchy), I found this thought provoking.

Since I deal mostly in niche/local advertising, I might have a unique perspective. It turns out that the more relatable a banner ad is, the less offended the reader is.

That may sound obvious, and it probably is. But the general effect is that users engage with and are sometimes genuinely interested in banner ads for their favorite restaurants, venues, and other spots. Local and niche advertisers are part of the community. A Ford or Chevy ad is generally much less relatable.

So my overall point is that the nature of the publication and its advertising clients is going to impact whether some readers are turned off - so it's difficult to derive a rule of thumb for banner advertising as a whole.

I think a most of the problems with ads are just caused by how Google and other big players run online advertising.

Somewhere along the way "let's add an image of some company our readers might be interested in" turned into "let's serve a huge Javascript that perceivably slows the browser in order to serve an ad from the highest bidder who might or might not run a criminal enterprise. Also make sure the ads show off the vast amounts of data we collect, we have to appear as creepy as possible".

Small ad networks or sites that organize their adverts themselves sometimes have great ads. But I am not sure if the reputation of online ads isn't already damaged beyond repair.

Add to that the small revenue from ads (the author makes $30/month from ads and $700/month from patreon)


I run Adsense on https://arcade.ly. The pages are heavy anyway because they implement games, and much of the payload is music and, to a lesser extent, sound effects. As an experiment I tried loading the pages without music and was nothing short of appalled by the amount of crap Adsense downloads.

Thing is, it's not Adsense itself that's the problem: it's the adverts themselves and the lack of control/moderation Google chooses to exert over ad providers. They serve up tons - I mean really a lot - of JavaScript, which tries to track you in N different ways, all so they can display a hyperlinked image with the most basic of animations and mouse over effects. This payload, often running to well over 1MB, measurably slows load and execution performance on the page.

The creepiness infuriates me, but it's the overall shittiness of the development practice that produces so much bloat for so little value I find most offensive. I mean these people go in to work, half-ass throwing together a bunch of code and assets in the worst way imaginable, and then carelessly spooge their effluent all over the web... and then they have the audacity to start whingeing about people running ad-blockers. What do they expect?

My site gets next to no traffic at the moment but if/when it's grown a bit I'm going to start looking at non-ad ways to earn money from it. I'm even considering ditching them at the moment to see if that might help grown traffic in itself.

I don't remember it being like this before they tied it in with DoubleClick?

This is the biggest problem I have with Adsense. Most of the time it seems like they couldn't find an ad that fits me but their system decided to just throw me a random ad for Forex Trading or whatever. This turns their supposedly sophisticated ad system into a random generator as far as I'm concerned.

They could fix this by:

1. Just showing blanks if they can't get something targeted to my preferences

2. Not allowing advertisers to be too general in their targeting if they are doing performance campaigns(CPC and CPA). Only allow CPM advertisers to target broad categories.

Your concerns are typically addressed by:

1. Reserve pricing - if no ad is 'good enough', e.g. no advertiser is willing to bid above a certain threshold for you, then no ad is shown

2. ePCM pricing - Adsense predicts what the likelihood of clickthrough/conversion is for a given ad and user, and includes that in its auction algorithm. Roughly, eCPM bid = cost per click / probability of click. This addresses the typical problem of high CPC/CPA bid but low engagement probability ads crowding out otherwise good CPM-based ads.

In general responsible advertisers have aligned incentives with viewers. Intrusive or obnoxious (or malicious) advertisements hurt viewers, who are less likely to return.

I think a high reserve price is actually a really neat idea - you'd think that might have the knock-on consequence of reducing the attractiveness of adblockers, and furthermore of reducing viewer fatigue (so that the ads you do show have more impact). If you're smart enough, and take into account the "damage" to your relationship with the viewer when choosing an ad (i.e. aiming primarily for the ad the viewer is most likely to appreciate, and only secondarily other effects such as how likely they are to click and how much the advertiser is willing to pay), you might even be able to make ads the viewer actually wants to see; which in the long run is probably a very good place to be.

I wonder if part of the reason that never happened is simply that ads are in a sense "reserved" layout space that need to be filled consistently and cannot be filled later on (due to layout jumpiness).

I wish I had a better way to tell Google to raise my reservation price than blocking most ads I see, maybe pay for a higher one?

I just wanted to say that it’s actually nice not to be the only one who is harassed by forex trading ads around the web. Them and Mobile Strike are pretty much all of the ads I see. (I do have Adblock but can’t block for YouTube mobile etc.)

There are other apps for youtube. E.g. newpipe.

As a user,I am most offended by ads that target me. I would rather see a completely unrelatable ad. It feels like learning new information that way,as opposed to being stalked and preyed upon.

The other option is to target the ads to the publication. As long as the publication is on a niche topic or geographically localized, you can choose ads that will likely interest people (local restaurants) without needing to track them. Tracking is only necessary when the ad venue is incredibly broad across topics and geography.

As someone who used to work trafficking ads and spends a lot of time blocking them myself, I'm of the opinion that this is the approach to take.

ReadTheDocs and sites like BikeRadar are my favourite examples: If I go to a bike site, I expect the ads I see to be about bike stuff. If I'm on a technical site (like Stack Overflow or ReadTheDocs) I expect to see ads about related technical topics.

Unfortunately the advertising industry has gone down the route of hyper-invasive, user-data focused advertising, for what I consider to be very little benefit to either publishers or users. The only people who benefit are those who pedal the tech that enables this market.

With all the ML tools available to us now, I'm not sure why we couldn't build tools that accurately match ad content to site content. You don't need to invasively track users users this way either, because you only ever need to be concerned with the topic of the site/article and not who's seeing it.

We could build these things, but part of the "advertising host" optimisation metrics is "time to ad display" so the algorithms in use are frequently "deliberately dumb" for reasons of optimising the time involved in deciding what advertising should be displayed.

Somehow, televisions had the miraculous ability to display content-targeted ads with no user knowledge and zero latency...

It's much easier to target the publication than the individual user.

Web pages are far more static than users, your adserver/ML/analysis tool need only analyse the page once (not including updates for content changes) and then pass this information straight on to ad buyers.

Even if advertisers analysed the page content on seeing it, rather than the publisher, it would still be net faster than doing lookups/analysis per adserve (as you would when you target a user).

I don't think it would be hard at all to build something which analysed a page and made an adserving decision with a lower "time to ad display" than a similar process that used user data.

That's how old media has been doing it for centuries. But media is hard: you have to invest in personal relations with advertisers.

New media just outsources all the hard work to Google and other big ad placers. Easier but you lose editorial control over your work.

More like "for about 150 years".

Advertising, in the modern sense, is, well, modern. You pretty much simply did not have goods to advertise (or the means to advertise them) other than for very specific items -- think classifieds rather than bulk and mass advertising.

The rise of the factory system, of transport networks, and the capacity to create far more of a good than local demand could absorb created the first modern-style advertising in about 1860. The first prominently-featured adverts were the 19th century equivalents of Viagra ads: mostly-ineffective patent medicines.

Hamilton Holt has a delightful contemporary account of this, written (and presented as a lecture, it's short) in 1909, looking back on the previous 50 years or so of publishing history. I cannot recommend it highly enough.


In fact, that’s how Google started out with AdSense as well: It was called “content targeting advertising”, because Google scanned your webpage and then served ads that related to the content of your webpage.

I don’t know what happened along the road to lead us to the current tracking and spying mess.

What happened is the ad landscape shifted from targeting placements and content to targeting audiences and where they are in their journey to purchase. It simply performs better for advertisers and it's often more scalable. In turn, this makes the platform more money because sites that may have low value inventory can suddenly be sold for higher CPMs if an advertiser for example wants to bid on their retargeting data in the auction. It let's them price based on people.

I have been wondering why that isn't done more often myself. I mean,it isn't like tv ads didn't work.

Advertising is very much beset by the mindset of being "so preoccupied with whether or not they could that they didn't stop to think if they should" (Jurassic Park.)

Tracking and microtargeting just seem too attractive for advertisers to pass up. Same goes for large size, autoplay, multimedia ads.

These mechanisms would have to fail to perform in obvious ways (not long-term, subtle, whole-market ways) before people will stop using them.

I do not know you, so I won’t presume (certainly not on HackerNews, the least representative sample of the Internet) but I have seen many experiments proving that most people who think they don’t like targetted ads actually prefer them -- until they make a very informed but very wrong guess. Typical examples include:

- advertising for something shameful (underwear and dildo are my go-to examples) or that you otherwise want be kept secret (engagement rings);

- repeat ads for specific and unusual big-ticket items. I have told friends working at Amazon that no one I know ever started a vacuum cleaner collection, but they only response is “You would be surprised…” to which I can only assume the world is either weirder than I thought, or their analytics function has missed something.

What you (and most people) probably would prefer is the ability to edit what the ad engine knows or assumes about you: gone are dildos, rings and vacuum cleaners; come books about programming patterns!

I keep telling anyone close to any of those recommendation engines to allow me to do that, but they mumble things about design, feasibility and hacking that genuinely don’t add up to me.

Not true,sure I would prefer programming related ads but I still want it page-relevant and not based on my browsing history,habits or sensor data.

I would love to see more ads that instead target the content of the page. When I'm reading about machine learning, it should be books close to that topic, and other related things. I can't believe little that happens.

It has the added benefit of having a high probability that you'll be interested in the product, since you wouldn't be on the page if you didn't find the topic interesting. Instead we get tracking networks showing you ads for TVs because you bought a TV last week.

I think the problem with that is that most browsing is done on pages that do not target a valuable topic: news, weather, memes, etc.

For that reason, it's better to remind you about that Amazon product still in your cart while you're browsing for random stuff.

I used to to that with Amazon. About a decade ago, I added an Amazon banner, and the link included a search term. So I would add an appropriate search term (related to the top blog entry on the page, which is where the Amazon banner appeared). Sometime in the past few months, Amazon removed that and only allows generic sales banners to be presented.

Considering the sprawl of sponsored content, maybe because it's cheaper/easier to just buy the author?

The other side to this is users see the ads, relate to the ads, click the ads, and thus leave your site. Your pageviews will drop.

Whether or not one agrees with the outcome, if you have a large site you definitely should be doing independent testing. Google's ad reps, much less reps from other networks, will tell you to do things that harm your site.

Well, most ads open in a new window. And most serious publishers barely have the resources to survive let alone do A/B testing. The publishers I work with thankfully shy away from network advertising and only deal with clients they have a direct relationship with. Regardless, it really isn't easy or very lucrative running a digital content based business. There are outliers, sure, but the long tail paints a much different story in my experience.

> It turns out that the more relatable a banner ad is, the less offended the reader is.

This is of course true, and kinda obvious, but I am not sure this is workable as business model. I mean, here are not a lot of ads that are immediately relevant for me. Maybe once or twice per month there's something that I buy that ad could sway my decision at least potentially (i.e. expose me to some provider I wasn't aware of, or some brand I haven't considered). The rest is basically wasted on me. How much 2 clicks per month (with unknown conversion rate) are worth to advertisers? I have no idea, but I wonder whether it is possible to support a site on that revenue?

I, of course, could be wildly atypical case (for one, I use adblockers and rarely impulse-buy) and maybe typical visitor would be much more engaged with ads.

Or it could be that ad buyer industry (not ones that make the ads, but ones that order them - Fords and Chevys of the world) do not have proper tools to see how effective their ads are and just throw a chunk of budget to the wind, get some traffic from it and feel good about it.

It's true — but advertising has two sides. You can of course be running a campaign that you're hoping for immediate ROI on, like a flash sale. In that case, you really want clicks and some sort of immediate indicator that the campaign is working, like conversions.

But sometimes as a business owner you're also campaigning for visibility / mindshare. At the minimum you want to distribute a message and let people know you exist. That's why people still buy billboards. It's a type of advertising that will probably never go away - especially because it's relatively low cost to implement for all sides - just need a designer to create the ad and you're good to go.

It doesn't seem like the most dependable business model but it's pretty timeless even in the digital age. The advertiser has an ego and instinct and is ultimately going to buy what they feel is going to work for their business.

Maybe a site like gwern.net is generally less suited for A/B experiments like that because of the unique look and speed 'HTML only' sites have. Sites like cnn.com or yahoo.com are generally much more crowded and colorful, so banner ads blend in much better. This might in return result in a higher user acceptance rate for banner ads and probably a different engagement rate too.

I imagine the worst advertising is where it feels both unfeelingly-impersonal yet also spyingly-specific.

Where do your ads appear? On your site I do not see a publisher signup page.

I used to roll my eyes every time I saw someone on HN talk about "signal to noise ratio," but I'm getting more and more swayed. Humans deal in abstractions: more clean an abstraction is, the more we can hold onto it.

So when I read a ~10% difference in traffic, I'm not surprised. Art, in absolutely any form, is always about focusing on your theme. And ads, by their very nature, blur that focus.

The author's approach to estimating the value of traffic, even by subjective criteria, strikes me as puzzling, somewhat arbitrarily valuing ad-free page views at $0.02, despite the fact that they generate no direct revenue.

> My website is important to me because it is what I have accomplished & is my livelihood, and if people are not reading it, that is bad. How bad? In lieu of advertising it’s hard to directly quantify the value of a page-view, so I can instead ask myself hypothetically, would I trade ~1 week of traffic for $360 (~$0.02/view, or to put it the other way, $18720/year)? Probably; that’s about the right number - with my current parlous income, I cannot casually throw away hundreds or thousands of dollars for some additional traffic and so I feel comfortable valuing page-views at ~$0.02 or a bit less.

If your income is as parlous as you say, then it seems bizarre to value any amount of unmonetized traffic over cash flow.

People call page views a "vanity metric" with good reason. You can't put food on your plate with page views. By valuing a single page view at $0.02, you're literally doing work for free, "for exposure."

We all make fun of businesses who ask professionals to work for $0 for exposure. It's no better when you're making yourself work for unfair terms!

To put this another way, I think the author can probably arrange to buy relevant traffic via Google AdWords at or less than $0.02 per page view. If traffic ("exposure") is that valuable, then, yes, stop running ads on your site, and instead buy ads of your own.

But if that sounds outrageous to you, if you would never reach into your pocket to pay Google $360 for a week's worth of website traffic, then you should rethink whether unmonetized page views are really as valuable as you think they are.

One can have enough money to value traffic but not enough to spend thousands of dollars for small amounts of additional traffic. As you have less money, the remaining money becomes much more valuable...

> it seems bizarre to value any amount of ad-free traffic over cash flow.

Also, while I do not know for certain how much each page-view is worth, I do know for certain that it is not exactly $0! Errors made on the assumption it is a cent or two are surely less than errors made on the assumption it is literally worthless.

(As it happens, the introspection was pointless because the effect came nowhere near the threshold, and if it had, I would've had to reintrospect anyway since my Patreon has been working better and I have more money since I began the A/B test in January. There's no reason to not revisit thresholds if circumstances have changed, after all.)

> One can have enough money to value traffic but not enough to spend thousands of dollars for small amounts of additional traffic. As you have less money, the remaining money becomes much more valuable...

If you have very little money, then every cent counts, including AdSense cents. Page views count only to the extent that they convert into something: Patreon patrons, consulting gigs, or what have you.

And I'm not saying your page views convert at $0 to those things, but I think it's plausible that you're off by an order of magnitude, e.g. that an ad-free page view is worth only a tenth of a cent.

He has Patreon. Ads are not the only way to turn traffic into money.

I honestly couldn't follow most of the math but found this really interesting. I used to do UX analyses at a marketing agency and would constantly link clients to [1]. Page speed is huge, especially on desktop where its easier to navigate away. OP mentions this, but I imagine the effect is massive on a site where users are actively researching a topic and have tens of tabs open and so can easily switch to a similar website.

I do think this test could have been done without banner ads involved. Looking at page load speed per user and how that effects traffic would likely have the same results.

From my link below:

> A 1 second delay in page response can result in a 7% reduction in conversions.

[1] https://blog.kissmetrics.com/loading-time/

In my experience that kissmetrics blog is a load of nonsense in practice.

We massively reduced load time on the all the key pages from 4 secs to roughly <400 ms avg. No 25% bump in sales for us. At the time the client sold about 50-80 things a day on thousands of daily visits. A month later, they sold about 50-80 things a day. This was a fairly standard website similar to booking.com or any search/book site, basically location aware searches with a landing page for each venue and a "book" CTA.

I did a similar thing with their mobile site a year or two later, massively reducing the page-weight and reducing loading times on the devices I tested from seconds to again sub-second load times. Again no discernible bump.

Right - the pageload time only matters for certain sectors. We sell sports science equipment and our competitors have 10s load times; I'm sure they have no issue selling their stuff. If they got their load times down to ~1s I bet they'd sell more... but not that much more. We are in the 1.5s-2.0s load time and I'm happy with that; given we use Woocommerce to drive our traffic I highly, highly doubt I can get it much lower given our current site design which does very well engagement-wise. That's fine for us; our conversion metrics and related analytics show that we're doing very well for our industry.

For niche equipment you're either gonna buy it or you aren't. If it isn't Amazon, Google, Wal-Mart, or places where you are preying on impulse buys (Woot, LowEndBox/VPS, etc), then the page load times matter a hell of a lot less than something like user design.

Did you try to see what happened if you increase the load time above 4 seconds?

I have a similar experience - after a redesign and significant improvement of page load time, the bounce rate and few other metrics improved but conversion rate stayed almost exactly the same. Kissmetrics are probably measuring this across thousands of sites and averaging the results as opposed to taking a slow site and tracking conversion rate after it gets speedier. So, it could be a correlation kind of thing - better sites, with more resources, simply invest more into every aspect of the business, including site speed.

I am wondering how much of that lack of sensitivity could be due to having two types of visitors:

- people who know what they want, know this is not ads or JavaScript and have an anti-advertising or anti-scripting tool; when KissMetric ran their experiment, those tools were not widely available and those visitors went elsewhere at the time;

- people who don’t know and enjoy the browsing and don’t mind the delay; didn’t then, don’t now and have kept on using your competitors’ website.

These kinds of stories don't get made into blog posts because they don't help sell any kind of narrative. Novice eCommerce store owners will focus on the "top 10 things you can do to increase conversion" from blogs like the one mentioned above.

At a large scale page load time would have a discernible impact on your bottom line, but for most it is a premature optimization (as long as you aren't trying to deliver 10mb webpages).

I'm aware of those results, but I assumed that they weren't an issue because I was using the best-practices 'async'/background loading AdSense JS snippets. The ads never block the rendering or loading of the rest of the page, they pop in a second or two later.

He seems to not show how the ads looked like on his page. He seems to only attribute it to load time. But maybe the ads were visually annoying? Why no screenshots?

I have adblock so I don't usually see them and didn't make a point of screenshotting them. When I did (usually checking in Chromium that some CSS change was working), they were fairly ordinary AdSense ads, IMO. Like on DNB FAQ, they would usually be Lumosity or college ads. They weren't flashing or visually offensive, so I never saw them as a problem.

I looked at some of your pages on archive.org but saw no ads either. Can you give me a hint how to find the pages with the ads?

It was on almost all of the HTML pages (except the index & Modafinil) up to September. Perhaps the JS doesn't run when loaded from IA or something?

Good question. Even if it does not load, it should still show the adsense code in the html. But looking at some random post, I don't see it:


Adsense is usually a script loaded from googlesyndication.com which is not in the source here.

Maybe archive.org rewrote the html so much that it is not there anymore?

After a few moments of searching I'm going to assume that whoever specifically scraped and uploaded the page may have removed the adsense javascript. Doesn't look to be a normal archive.org thing.

You can upload pages to archive.org? Wouldn't that open the door for spam and other abuse?

I checked some other site and the Adsense code is not changed by archive.org. So something is strange.

Maybe Gwern experimented with different types of ads and on that page at that time did not use adsense?

I do see the adsense code in a snapshot from a year later:


I am yet to see any ad setup that guarantees they are not visually annoying. Most (reputable) ad networks guarantee they filter ads and remove the evil ones. I haven't seen one that delivered. I do not see how they could - manual review of each ad is not scalable and advertisers would not abide by it, they'd just move to network with less stringent rules.

I believe Project Wonderful is the closest thing I know - it's a manually-reviewed ad network used for webcomics, with the downside that it's accordingly much less valuable for site owners.

There are ways to have comments that are a lot less awful than Disqus.

Which are and which work with a static Hakyll site?

That seems to require you to host your own server and doesn't look much better in terms of complexity.

See my comment nearby about EffectiveDiscussions — it's a new commenting system, open source, & there's hosting. (https://www.effectivediscussions.org/blog-comments if you didn't find my other comment)

(What did you mean with "look ... in terms of complexity"? You mean how-complicated-it-is-to-install?)

I don't like Disqus mostly because the UX is clunky. Are there other things people dislike about it?

Being tracked by a single third party on every site that uses it comes to mind.

Disqus shows ads, unless you pay like $10/month. (& tracks your visitors, like someone else mentioned.)

Maybe you would instead like the looks of EffectiveDiscussions — it's a new commenting system I'm developing. See my other comment nearby (or, if you didn't find it, go to: https://www.effectivediscussions.org/blog-comments ).

This is UX-ish, but I regularly have login/posting issues and very slow loading times off Disqus.

Commenting system for static site please? Meaning I don't have to host anything dynamic myself and the price won't kill you.

EffectiveDiscussions is a new alternative this year 2017. It's open source, or hosting for $2/month. (I'm developing it.)

Demo: https://www.kajmagnus.blog/new-embedded-comments (That's a static website built with Gatsby.)

Read more about it: https://www.effectivediscussions.org/blog-comments (landing page, not finished)

Since it's new, it's a bit risky: maybe some bugs. & has some novel features, inspired by HackerNews actually: https://www.effectivediscussions.org/-32/how-hacker-news-can... (but that's not totally ported to embedded-comments yet). Currently implementing single-sign-on.

I scanned the article to see if it was interesting enough to dive into the details, but when I saw the Disqus plugin I closed the tab. What's the point of dropping Adsense if you keep the Disqus plugin?

Simple, they value the discussion capabilities provided by Disqus. Disqus was worth keeping, Adsense was not.

That’s my point exactly. Did they do the same exercise for Disqus? Or did they get all scientific on evaluating Adsense, and go with a “hunch” for Disqus?

That’s like saying “Marlboro cigarettes are harmful, here’s proof, let’s grab a pack of Lucky Strikes and discuss”.

It's more like smoking two brands of cigarettes and then deciding to stop half your smoking; that is surely an improvement...

This is just whataboutism. AdSense should be judged on its own merits, and it fails badly: it does nothing I want (aside from delivering quite small amounts of money) at an unacceptable cost. Maybe Disqus is also bad but I would have to ponder how much I want commenting and how to deal with the extant corpus of comments and in any case, at the moment I am gearing up for another ad test to try to figure out what about AdSense is so harmful (and help convince the doubters that there is an effect at all), so I can't test Disqus without complicating things a lot.

The content is characterized as "highly technical longform static content in a minimalist layout optimized for fast loading & rendering catering to Anglophone STEM-types in the USA", which is definitely true, but I think there's an important subset not mentioned. According to http://www.gwern.net/ the most popular three articles on the site are about darknet drug markets, and #4 and #5 are reviews of specific controlled substances in the US. I don't know what the ads for these pages looked like, but here's an example from a Google search for "modafinil" in incognito mode with my adblocker turned off: https://i.imgur.com/3dRssOy.png ["the silk road" and "lsd microdosing" didn't turn up any ads, but it's possible Google does more filtering on its own site than on others... I couldn't easily find any LSD review sites that used AdSense for a fair comparison]

If I saw an ad like that in a banner at the top of the page, it's going to cause an immediate, visceral reaction: the site is sketchy, maybe trying to scam me, and any content is going to have a big conflict of interest. To be fair, since I use an ad-blocker, I'm probably more ad-sensitive than the sample set... but who hasn't gotten their fair share of similar email spam before?

The CSVs don't have page information, unfortunately, but I'd expect most of the harmful effect comes from the drug-related pages.

The modafinil page didn't have AdSense during the experiment, and if it had, I think the content would have not involved modafinil; Google was burned badly years ago by mega-fines for allowing greymarket pharmacies to advertise and I assume it's totally filtered them out, so any served ads would be targeted at other personal characteristics, presumably. (I also vaguely recollect trying to buy some adwords once while experimenting with advertising gwern.net pages, and the drug categories being totally out of bounds.)

During the experiment, roughly, the most popular pages were a mix: https://www.gwern.net/About#january-2017---july-2017 The top 5 were modafinil (excluded from AdSense), spaced repetition (psychology), LSD microdosing, DNB FAQ (also psychology), and "Story Of Your Life" (physics/literary criticism).

I am curious, what constitutes “traffic”, and if it is page loads, why is it the right thing to measure? Users who load the page are already visiting, regardless of whether there is an ad on the page or not. (Or does it include a delay to filter visitors who immediately close the page?) And they did not know in advance whether there was going to be an ad, so their decision to visit the site is not based on the presence of ads. (Unless perhaps if they had a particularly bad experience a previous time — bad enough to remember the domain and not follow the link.) I don’t understand how turning ads on and off every two days can be used to understand long-term effects. There appears to be the assumption that there is a relation between a user seeing an ad or not, and subsequent traffic. There might be a small correlation because of visitors browsing the site, or likeliness to share, but if outside of the spikes, the majority of visitors arrive from an old external link or search engine, and they only read a single page, then how does turning ads on and off every two days measure impact on these users?

It is probably a combination of performance impact, lower page views per visit, and lower return rate.

What do you think the mechanism is on the connection between ads and pageviews? Or the mechanism for the autocorrelation of pageviews? Have you been able connect the dots on something like sharing to sites like HN?

> What do you think the mechanism is on the connection between ads and pageviews?

The mechanism that makes the most sense to me is via influence in social media resharing. This requires affecting only a few people (power users/the '1%') but can plausibly deliver the necessary effect size. The logical next step in the analysis would be to go back into Google Analytics and pull out another set of time-series and try to split between search engine, direct, and referrals; the decrease should be concentrated as search engine < direct < referrals. I'm not looking forward to figuring out how to code that up as a time-series model, though...

> Or the mechanism for the autocorrelation of pageviews?

Well, some of the mechanisms there are obvious just from the referring URLs. Many people don't read HN via the front page, they use a third-party interface or a daily newsletter; so while OP is getting substantial traffic today, it'll also get traffic tomorrow from people catching up with today's 'best of HN' etc. And from HN it gets auto-posted to Twiter, and of course many people will only see the tweets later when they log into Twitter the next morning or evening. Then you have timezones - perhaps someone is asleep now even if they do read the HN frontpage directly. It might get included in blogger link roundups or sites like The Browser weeks or months from now - more autocorrelation. What about search engines? They'll notice, and will update. No one has mentioned any comparable experiments so right now my A/B test appears to be _sui generis_, so people debating the costs and benefits of advertising might link it over the next few months. And so on. Lots of possible ways.

> ~60% of visitors have adblock

This is much higher than I'd expect. Of course, it is a special kind of site, but still. Can we say a model is in serious trouble yet?

I think the right metric for AdBlock (like every service) is retention. If most people who try it like it enough to remember how to install it on their following browser that model has a sigmoid-shaped problem.

I'm amazed they can even tell. A lot of people who use adblocking also browse as ghost traffic that can't even be analyzed.

Ad blocking means that your browser does not send HTTP requests for items that match the profile of ads and tracking beacons. A website can detect this by creating a local asset linked to from the HTML that is never rendered, but which looks like an ad from the ad blocker's perspective. If a browser loads your page, but subsequently won't load that dummy asset, it is probably using an ad blocker.

As always, thanks for the very thorough write up. My immediate thoughts were highlighted in the "Discussion" section. Namely, the demographic of your website might respond more negatively to the presence of ads than the general population. This seems to be substantiated by the unusually high use of ad blockers (~60% vs ~18% [1]).

I also wonder if some of the drop off could be explained by the use of Patreon and ads. For instance, I would never pay for Hulu because they still show ads. Did you receive any Patreon cancellations or negative feedback from pledged users?

[1] US ad block usage as of Feb 2017 https://pagefair.com/blog/2017/adblockreport/

Patreon-wise, I had ads long predating Patreon, so presumably anyone signing up was already OK with that? I don't remember anyone ever complaining about both ads & Patreon, unless there's a complaint form on Patreon.com I haven't been checking. (Looking at the signup numbers, it goes from $619 for August to $645 for September/October, which at least is in the predicted direction, but too few datapoints to tell since the usual trend is upwards.)

How does this experiment methodology account for cross device mixed experience? Because of the multiple devices use, There will be lot of users that have seen both variation.

I think changes over time in device should be like any other unobserved variable like age or country or gender which might affect traffic: averaged out by randomization. The point of randomization is that on average, the two groups will be balanced. (If the intervention itself changes the mix and the changed mix then yields more traffic, then that changes what one might speculate about how it works but the simple result of ads->less-traffic is still true.)

I would love to remove ads from all but one of my sites. But I have not been very effective at monetizing my writing some other way. I removed them from a bunch of sites and saw a decrease in ad revenue, but did not see some noticeable difference in either traffic or tips. (I only later added Patreon.)

There are myriad contributing factors that I can't isolate out. People don't like ads, but many of the people who loudly complain about ads will also very aggressively tell you that they will not tip, will not be a Patreon supporter and expect high quality content for free. They will aggressively say that your business model is your problem, not theirs, while actively shooting down any way to monetize content.

These are often people who make good money as, for example, programmers. Another thing such people have told me is to get a real job, like they have, and write for free. It is a kind of classism in a sense: Writers who provide good quality web content aren't seen as valuable or serious contributors the way programmers are. Yet, at the same time, people decry both the lack of good content and the trend towards content done as content marketing.

I get paid to write content done as content marketing. I rarely write anything I feel bad about. I work on projects I can believe in. But I find it frustrating because I would rather spend more of my time producing high quality, independent content for my own projects. These are projects intended to put out useful information and/or to be entertaining.

I am doing some unique things that add value to the world. I run a homeless site with useful information for homeless Americans. Most homeless services have terrible websites. They are almost always donor facing, not client facing. It is very challenging for a homeless person to find useful information online. I had a college class on internet search and I had a hard time finding information I needed.

The other thing is that a lot of homeless services are a very negative experience and can help keep you trapped. Services are often designed with the idea that homeless people are incompetent losers who can't be trusted to make good decisions. They often come with strings attached that can help keep a person trapped in poverty.

I want to promote market based solutions that can work for homeless individuals, but can lead to a future because they continue to be viable after you get off the street. Programs to help the homeless or help poor people often actively encourage failure.

I think I have worked out a couple of models that work, but I am not finding a path forward. I continually come up against the idea that I should just stop caring about marginalized people because there is no money in it. I should just go where the money is, and never mind that the world is going to hell.

And I can't quite bring myself to do that because there is scary stuff happening and I feel like if I don't do something, no one will. Because rich people aren't going to solve it. The attitudes of so many of them that poor people just need a handout (UBI), not earning opportunities, is part of the problem.

We have the technology to allow people with disabilities, like me and my sons, to work in a way that works for us. But I don't know how to overcome the attitudes that we shouldn't bother to try, we should just accept that we are losers unworthy of any respect.

The inability to get taken seriously is a big barrier to getting engagement and traction and that is a big barrier to turning it into money. The whole thing makes me crazy. It is a case of the key is in the safe.

> I think I have worked out a couple of models that work

You should write these up in detail and post it to HN. I would read it!

I have written about them and posted stuff to HN. It usually doesn't get much attention.

Nutshell version:

1. Flexible online means to earn money.

2. Genuinely affordable market based housing, such as SROs.

I have several blogs where I talk about such things, like Mic Digs, Project: SRO and Write Pay. They don't get much attention.

Edit: just to prove my point, here is a recent piece on working for pay while homeless. I have submitted it to HN. You can go watch it be ignored:


Maybe you could do something like indie hackers [1] and highlight the stories of people that are trying to get off the streets by doing online gig work. That would be inspiring for people looking for help and interesting to everyone else.

[1] https://www.indiehackers.com/businesses

I did spend some time looking over Indie Hackers yesterday. I am unclear whether anything I do at all fits there. But thank you for suggesting it. I am trying to figure out how to get taken more seriously, in spite of it being an uphill battle for various reasons.

I'm curious if any ad networks or SSPs are differentiating based on load times. I wouldn't think so because big sites that drive lots of impressions tend to load slower and aren't as concerned with page speed as someone like OP.

I'd also be curious to see his user to patreon donator conversion rate, and if removed ads and 'gaining' 10% of traffic would be worth it from a revenue stand point.

Unless I misread this it sounds like the assignment is on 2 day chunks rather than some unique user ID. Just doing this at impression rather than user level makes me question the applicability of the results, but really interesting read nonetheless.

Per-user level doesn't capture the effects I am interested in. Imagine you click on a website and the banner ad makes you subconciously less happy and so you don't submit it to HN for thousands of other users to read; how would randomizing on you and thousands of others' user IDs help? There's no way to link your user ID's randomization to the subsequent visits (or non-visits) of thousands of other people; it just averages everything out to 0 ('500 visitors were exposed to ads, 500 were not; conclusion: ???'), when in fact the real answer is something like '-1000 hits because user yellow_postit was exposed to ads and didn't like them'.

But if you did it on a per-day or per-week or per-month basis, then you could see effects like that: "on ad-free weeks, there are twice as many submissions to Reddit which drives +100 hits per day" etc. Then you don't need to try to track each individual user's activities or reshares - whatever the effects are, they'll show up at the population level.

Isn't this particular test susceptible to confounding effects? Traffic fluctuates day-to-day, week-to-week, and month-to-month, so how can you be sure it was the presence-or-absence of ads and not something else? If you randomize at the visitor level, you are sampling from both high-and-low traffic days, and control for any external fluctuations.

Because you're randomizing at the 2-day level, on average there will be just as many advertising/high-traffic days as advertising/low-traffic days, and as many no-advertising/high-traffic days as no-advertising/low-traffic days. The randomization is unaffected by traffic and uncorrelated with it. The unit of analysis is each day, not each visitor. This is why it has to be run for several months, otherwise you don't wind up with a decent n=50 pairs.

That's the tradeoff here: it lets you look at the totals, but it takes a lot longer than if you randomize per visitor in which case you could finish the test in a few days, often.

The stats went well over my head but as a web analyst I thought maybe you could have asked a simpler question. Pick a segment of your site such as visits from Google Search who landed on your homepage (most likely people who searched "gwern") which should reduce a lot of those spikes.

Subsetting will also increase the variance of each datapoint (consider the extreme case of picking a subset which was 0 or 1 visits per day), so is probably not a win. It's also hard to imagine what subset properly reflects all sources of traffic and so is informative about the total effect of advertising. Search queries definitely is not it.

I don't know how much research has been done on the subject,but personally I don't mind ads in general,I mind ads that target (stalk) me and intrusive ads.

I can tolerate ads that target the page content or random ads just fine.

> I don't know how much research has been done on the subject,but personally I don't mind ads in general,I mind ads that target (stalk) me and intrusive ads.

That used to be my position. I think I still don't mind honest-to-goodness advertising: this product/vendor exists and has these features. I actually do still see this kind of ad on sites connected to electrical engineering; presumably other competitive technical product areas like photography or scientific/industrial equipment have it too. But the more I've experienced the dissonance of being hit with bursts of mainstream advertising after moving to a generally lower-advertising lifestyle (ad blockers/NoScript, cord-cutting, more judicious selection of news sources, etc.), the more distinctly I've felt the reality that we're almost a century into the general trend of "advertising" being a euphemism for psychosocial manipulation. In other words, modern advertising is for-profit propaganda. In my view, it's not enough for the media of advertising to be more passive. The technology is amplifying it, but the overall strategy is what's fundamentally exploitative. Many people think that merely by being cynical about this, they're immunizing themselves. It's tempting to think that if you study the techniques and make snarky remarks like, say, "oh, look how cool this Lexus is; if I buy it that means I'm powerful and important", then you've somehow defeated it. Make no mistake: that's just another strand in the web. The idea that the advertiser wants to push is still being normalized in your mind, even through the filter of irony.

I don't intend to be alarmist here. Humanity has just about survived a century of this crap, after all. Life goes on. Revolutions, both figurative and literal, haven't been rendered obsolete by the sheer marketing budgets of incumbents. But I think we have evolved a warped view of what qualifies as normal, honest communication.

> A decision analysis of revenue vs readers yields an acceptable damage of ~3% total traffic loss. Power analysis of historical ... traffic data demonstrates that the high autocorrelation yields very low statistical power with standard tests & regressions but acceptable power with ARIMA models.

Was all that math worth your time, or would you have made more money selling tamales on the corner?

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact