Experiment - using nofollow to save PR

Some time ago I published an article: Using nofollow to save PR is still working.

If you read this article, I conducted a test that showed that nofollow still works. I sincerely believed in the results of this test, but as it turned out, I made a mistake!

The test results, as it turned out, were not convincing, since the examples were simple and not in sufficient quantities .

And having done that, I unintentionally misinformed all of you. For which I apologize. Now my task is to improve and put everything in its place.

In the meantime, what should I do to save PR?

You can forget about the results of the first test. This means that I and my colleagues from SEO recommend how not to remove nofollow and not to add. And we say this because there is no convincing information.

Spend time writing content or loot backlinks

Remember, the most valuable information that Google has ever given to optimizers.

Provide the information that Google wants to make available to everyone - these are the long-term prospects in SEO to be successful.

So, let me remind you what was the last test.

A full description (with all the conclusions and results) of the past test can be found here: Using nofollow to save PR is still working.

What went wrong?

I can say that the experiment passed without problems. As it turned out, the problem could not necessarily arise during the experiment, but rather it was a mistake in interpreting the results . I used an incorrect metric in several examples.

New test

Than testing, which affects the preservation of PR better, I'm going to check if nofollow works at all.

We worked with the numbers (below) and found out that we could conduct an experiment in two ways. In the first of these, it was enough to use 40 examples - and that would be enough for accurate results.

But the second path had less chance of success, because it required much more examples. Below are the test schemes using nofollow and test with regular links.

These charts are simplified versions of the pages that were used in our test. The real pages were more natural and human.

In this new test, both "normal links" led to two different pages using a unique text anchor. Variation of the link "x" led to a separate page. The variation of the link "y" was in nofollow and led to a completely different one.

For each group we needed to find out which page will be ranked above all.

Our assumption was that the page to which the page links to the nofollow link (variation of the link "y") will have the highest result . Because we believed that on a page with regular links, the link weight would be distributed equivalently and they would give less weight to the page to which they lead.

This test was duplicated 20 times , as in the picture below:

We think that 20 tests and 40 different domains is enough . In order to say with confidence that nofollow really works when conducting our second test, we had to see in 15 examples out of 20 that the PR was saved.

If this does not happen, we will be forced to conduct a third test with even more examples.

What should warn you against repeating the same mistake?

In processing the old dough, I was assisted by Ben Hendrickson , who was nearby. Please do not be lazy to check our calculations before we begin the test.

Calculations for the nofollow test of 168 pairs.

This test consists of several samples. In each trial, the page with either nofollow or no will be quoted higher.

Thus, the number of victories will be distributed according to the binomial distribution. Where, n is the number of samples, and p is the probability that nofollow will win. Here is the approximate binomial distribution:

And here is the formula for the number of victories, where W is the number of victories, and z is the number of deviations from the norm:

The assumption that nofollow will win at p = 0.5 means that nofollow is not working.

In order to make sure that nofollow works in 95% of cases, we should get at least the following: W (n, 0.5, 1.645).

How many victories will we see?

We are 95% sure that we should get at least W (n, p, -1.645), where p is the chance of winning nofollow in each trial. If we get a low p = 0.625 for what we are trying to prove, then we get a lower W (n, 0.625, -1.645).

Now we can use this minimum number, which should work out if we expect nofollow to work at 95%. Then we can calculate everything for our number of samples:

So, we calculated that we need 168 samples . Even if these calculations do not prove that nofollow works in 95% now, we can say that it works in 62.5% .

Calculations for the nofollow test of 20 pairs.

Why don't we test for 20 pairs?

The answer is simple. We have already conducted a test of 168 pairs - this is a huge number of different domains. And we have already received the exact answer that nofollow works at least at 62.5%.

In order to prove that nofollow works in 95% of cases, we need to achieve 15 nofollow victories out of 20.

You can do the calculations again, but we got these numbers using the binomial distribution calculations.

Substitute p = 0.5 (as this is the minimum for a page with regular links), n = 20 and any number of victories until you find the smallest number whose chances to approach are less than 5%.

You get 15. And this is what we wanted! We proved that at least in 15 cases out of 20 nofollow works !

What did I learn from the first test?

Many, after reading my very first test with nofollow, believed me. That was until Darren Slathen showed me a mistake.

For me, this meant that I had already hurt someone, but I did not have the opportunity to contact each of the readers.

Those few who noticed a mistake were disappointed in me and all the misinformation spread through their blogs, twitter, etc. And this lesson showed how information quickly spreads over the Internet.

(de) information on the Internet is always there.

As soon as information appears on the Internet, it becomes uncontrollable from both its creator and people who want to hide it.

A small addition ... Why do not you post the real addresses of the tested sites, so that we can calculate everything ourselves?

I'll do it, but not now. Having laid them out now, I will put both this experiment and the future at risk.