Or titled, "Search Autocomplete Boosts Sales by 24%, and I Can Prove it."
I'll share the payoff right up front.
Adding or improving typeahead (auto complete) on a search box will increase the average search length by 1.6 words (from 1.7 to 3.3). Longer, more specific searches lead to higher engagement and higher conversion rates.
Specifically, every word added to a search drives a 15% increase in conversion rate. So, adding autocomplete will increase sales by 24% (1.6 words x 15%)
Keep reading, and I’ll show you how I figured it out.
I didn’t make the connection at first. It wasn’t until I got off stage and was back in my room that it hit me.
I gave a presentation at CMC about “How People Search When They’re not on Google + Amazon’s $10B secret and how you can use it to drive 80% more conversions.” It was the first time I’d gotten to use Nacho Analytics to do some original research for an audience.
I was super stoked. The presentation had these two distinct sections, one where I compare the search behavior of the Top 30 search engines (that aren’t Google) like Youtube, Amazon, Netflix, Bing, Ebay, Craigslist, etc. Here’s the slide I used:
I started to drill into really high-level behavior differences on these different sites. First I looked at the number of searches divided by the number of keywords to understand how diverse their searches were.
Then, I looked at the average length of searches, and there were a couple of sites that stood out. (BTW, the search volume on in these is raw and unmodeled - definitely not intended to be an exact estimate, but more of a way to compare behavior.)
Netflix was the first one I looked at , and the reason its search length is so short is that every time you press a key, it updates the search result showing you a new list of shows. It’s pretty clever; their search is actually their typeahead. Here’s what it looks like in action:
The second one I looked at was Craigslist . Craigslist’s auto-complete is utterly awful. This was my slide:
To illustrate what a non-crappy autocomplete experience looks like, I typed the same sequence into Ebay’s search box.
And so, now it makes sense why Ebay’s average search is 3.3 words and Craigslist’s is 1.7 words long.
Then I looked at the top searched keywords from both of those sites and see how they compare. Clearly, Ebay searches are longer and more specific. The Craigslist searches are very general, and even navigational in nature.
So, having good auto-complete changes what gets searched, making the searches more specific. And by making the searches more specific, the quality of the search result becomes better, making people more likely to click on them.
You probably understand this intuitively, because as a non-noob “searcher”, you know that if you are more specific, you’re just more likely to get what you want.
As I’m writing this, I realize that you might want a bit more proof that it’s the auto-complete feature that’s actually causing the length of searches to increase. The best case scenario would be a before and after, or a split test where half of the users see the type ahead feature, and half don’t.
Unfortunately, I don’t have that. But, if we go back to that spreadsheet:
All we’ve got to do is look at a few of the other websites with ~1.7 words per search on here, and check to see if they’ve got decent autocomplete, and then look at the ones with about 3.3.
- Reddit is the first 1.7. Guess what? They don’t have auto-complete! Seriously Reddit? Wow. I’m going to have to do a whole article on that, and why they should and how it’ll improve engagement, time on site, and ultimately ad revenue. Then, I guess I’ll post it on Reddit.
- VK.com is like the Russian FB, and yep – no typeahead.
- Twitter has a suggest, but it’s kind of a special case since the searches are pretty much always going to be a handle (@spyfu) or a hashtag (#seo).
- Baidu doesn’t have type-ahead. Well, not on desktop anyway. It’s also Chinese – and I have no knowledge of Mandarin. Does anyone know if that’s a language thing? Otherwise, wtf? Only major search engine without it.
- Roblox doesn’t. Roblox, if you read this – do what Netflix did. I think it’d work really well for you.
- Tumblr is a bit inconvenient because they have typeahead. But, I think this might be really similar to the twitter thing. Searches are short because the purpose is usually to navigate to one of their user’s blogs. I’m on airplane Internet as I write this, and the suggest is super laggy, so that might also be the issue.
And as for the 3.3ish word sites? All of them have typeahead.
So, basically, if a site has no autocomplete, or if their autocomplete is crappy – the number of words per search is gonna be about 1.7 words, and if it has decent suggest, it’s going to be 3.3 words.
Longer Searches Lead to Higher Conversion Rates
In the second part of my presentation, I showed how Amazon is getting a 6x boost in conversion rate from their search results when all of their competitors only get 2-3x. They make an extra $800M per month doing what they do, and if their competitors did it too, they’d nearly double their current sales.
But, it took me a long time to figure out how they did, and so in the process of getting there, I had a whole bunch of failed hypotheses. One of those was “Maybe Amazon users are just more sophisticated searchers. Maybe they use longer, more precise searches.” It’d be kind of like Google vs AOL searches – I thought maybe that was what I was seeing, but in Amazon vs Walmart or Etsy.
So, in Google Analytics I looked at Amazon sessions with 1-word searches vs 2-word vs 3-word vs 4-word. As you’d expect, as search length increased, so did conversion rate. Here’s that slide:
Then I ran the same analysis on Walmart.com – and saw the same thing. About a 15% increase per word added to the search.
Those were the only two examples I gave in my presentation because the point of what I was doing wasn’t really to prove that longer searches result in higher conversion rates. I was really just trying to get to the bottom of that crazy Amazon conversion rate, thing.
To really prove this, you might want to see a couple more examples. Here’s Etsy:
Maybe the 15% “rule” is a little bit conservative. Etsy is seeing closer to 25-30%.
Here’s REI.com. Follows the upward trend.
As a side note: REI looks like it has a ridiculously high conversion rate. That’s because I have the goal set to “start checkout” rather than the receipt page. That’s actually a trick you can use on smaller sites to get more conversion data – use micro conversions – like add to cart links instead of macro conversions like “receipt page”.
Using custom segments to create reports
Here are the Google Analytics custom segments I used to create these reports:
Site search – 1 word (Custom Segment)
Site search – 2 word (Custom Segment)
Site search – 3 word (Custom Segment)
Site search – 4 word (Custom Segment)
There’s still a part of me that’s skeptical that I’ve really proven this. When I put those segments together, I was looking for a 2-3x difference in conversion rate between sites. I was looking for a really big change, and so I put these segments together really fast, and in a way that I could reuse across sites.
Now that I’m looking to home in on a more precise answer, I think I should create a more controlled comparison. The segments above show sessions where any of the searches are 1 or 2 or 3-word searches. So, the first search could have been 1 word, and the third search in the same session could have been 4 words. In that case, the session would show up in both custom segments.
A more controlled experiment would be to compare users:
Went to the Homepage —> Immediately did a 1-word search vs.
Went to the Homepage —> Immediately did a 3-word search
The Custom Segment looks like this:
And, it all still holds up. In fact, it looks like our 15% bump per word is more like a 16.5% bump.
Causation and correlation
Let’s talk about causation and correlation for a second. One thing I want to make clear: I think that buyers are more likely to use longer searches and “browsers” are more likely to use shorter ones.
So, when we’re looking at longer queries having an increase in conversion rate, I’m not claiming that’s 100% causal. What typeahead/autosuggest does is it helps the search engine deliver higher quality, more relevant results, thus giving the user - whether they’re a “buyer” or a “browser” what they want.
The way to prove that is to look at how many people searched vs how many people clicked on one of the results . That’s how you can understand search quality. To do that, I created a couple more custom segments.
In this segment, we wait for the search to happen and then see if one of the results was clicked immediately afterward.
There’s more than one way to know you’ve landed on an Amazon product page, but the way I used is to look for “sr_1_” in the URL. Incidentally, the number that comes after the “_” tells you the SERP position of the product, which lets you do some pretty interesting things.
That produces a report that looks like this , and if we just divide the number of “clickers” into the number of “searchers”, we get the percent of searches that are click-worthy. This kind of metric is gaining steam in keyword research sites like SpyFu where we recognize the importance of working with metrics like "searches not clicked." Calculating it in Google Analytics helps to tell the story of search quality, of relevance.
Okay, I think I’ve laid out really solid evidence that Autocomplete makes search queries longer, and that longer queries have a higher conversion rate .
Thus, if you add or improve your autocomplete, you’ll see a boost in sales (for the people that search). But, also, keep in mind that in addition to this sales-driven reason to have suggest, there’s also a lot of compelling UX reasons to do it.
Update| A word about our access to the analytics and experiments shown here:
This original article ran in support of a marketing service called Nacho Analytics. At the time, we had access to behavioral data about how people interacted with websites, and we visualized it through the Google Analytics tool. All of that information has since been removed, and the service is no longer available.