Google told us in 2010 that page speed adds Googlejuice, but never filled in the details. They keep mum so SEO blackhats don’t game the system, but nobody likes being kept in the dark.
Hacking Google’s algorithm
Zoompf ran an intricate experiment to figure out what Google is really doing with search results. It’s beautiful – go check it out.
They set up servers to test speed on 100,000 pages that scored different ranks for various search terms.
Zoompf showed their results graphically. If they ran heavy-duty stats, they kept it quiet so as to not frighten the children.
Zoompf was kind enough to post the source data, so I went back with a statistical chainsaw. My analysis saw similar results, but found details that seem important in teasing out deeper knowledge. They invited me to elaborate, so here goes …
The statistical rubber room
I regressed search rank against possible factors from Zoompf’s data, pulling out the predictors that weren’t statistically significant. Let’s boil down the results, ranked by importance:
- Time To First Byte (TTFB) is positively correlated to Search Engine Rank Position (SERP) – faster correlates to better rank. TTFB is a pretty raw measure of server horsepower and distribution. TTFB’s importance is surprising because it’s typically only 20% of the entire load time.
- SERP decreases with faster Full Render Time (FRT). Really? FRT is how long the whole page takes to appear. Wow, this means slower pages rank better.
- Nothing else rises above the noise. Other variables have p-values above 0.05 (good chance we’re seeing randomness).
Stats Xanax. Stat.
Faster pages don’t deserve worse SERP. No way. So what’s actually happening? Consider the possibilities:
- Google messed up. Not buying it. They’re too smart for an obvious goof. Next.
The strange number for FRT is because of correlation between First Byte and Render Time. More plausible. Multiple regressions get difficult to interpret when explanatory variable correlate with each other.
However, a single variable regression shows the same relationship. This confirms the strange data and makes us even suspect First Byte reliability (which points in the right direction). Next.
- Hidden relationships affect both speed and rank. Some people suggested this in the Zoompf comments. Bingo.
- Popular sites like Amazon are more likely to get good ranking. They can also afford the best servers and TTFB.
- Popular sites serve richer content and so take longer to download. Speed’s important, but it isn’t as important as good content.
What it means
Zoompf hacked halfway into Google’s page ranking algorithm, but we’re still in the dark.
Sometimes, stats are nasty. Hidden problems lead you down dead ends.
Similar problems crop up in my world. It’s difficult to provide helpful numbers connecting speed to dollar value, but the analysis is valuable for websites. Confounding factors add another layer of work for Web Speed Ledger, but would otherwise send people in the wrong direction.
Zoompf’s study is an innovative approach to reverse-engineering Google’s algorithm, but was a lot of work.
If they have the energy to build another dataset, I’d love to see them try again in a month. Some websites will speed up or slow down without much changing other search ranking factors. The differential in time vs the difference in rank for these sites should knock out most of the confounding factors.
Go for it, guys.