How to Use Data Science for SEO

The applications are well documented. Group a keyword list by meaning. Read a log file too long for any person to open. Forecast next quarter out of the last eight. Catch the week a number quietly broke. All of it works, and none of it gets repeated on this page.

Three further questions get settled elsewhere. Whether you need to write code. Then whether your business holds enough data to model anything at all. And whether a correlation ever proved a cause. Take all three as answered.

Something survives all three. It sits underneath the whole discipline, and it rarely gets said out loud on an SEO page. Programming skills for machine learning, the first of those, are covered on another page.

What Data Science Is Actually Made Of

Strip the software away. One idea is left, and the rest is built on top of it.

Observations have to be independent.

That assumption sits under a regression. Under a test as well. Reach for a statistical instrument without thinking. Almost certainly you have reached for one that assumes it. Each case has to be able to surprise you. Learn what happened in one, and you should still be genuinely unsure about the next.

It gets stated on the first day, then treated as furniture. SEO platforms built on data sit on that same assumption, dashboards and all. Same fate meets the levers behind PPC performance, stated once and then ignored.

Ask One Person Four Hundred Times

Put a question to four hundred people. Now you have learned something.

Then put the same question to one person, four hundred times over. You have learned one thing. Four hundred times.

Both exercises hand you four hundred answers. Only one of them handed you four hundred pieces of information. So the power was never sitting in the count. It was sitting in the independence.

A Website Is the Least Independent Thing You Own

Now look at what the mathematics is being pointed at.

Your pages share a domain. They came off one template. Two or three people wrote every one of them. Same couple of years. Also the same view of what a good page looks like. They carry the same navigation and the same footer. Meanwhile they link to each other. On a few queries they even compete with each other.

Pick two of your pages at random. They resemble one another far more closely than two pages pulled off the open web. Then notice when that resemblance started. It was there before anybody optimized anything.

So four hundred pages is not four hundred cases. Instead it is one website, counted four hundred times.

You Do Not Have a Small Sample

Here is where the argument parts company with the usual complaint.

The usual complaint says a business holds too little data to work with. Often true, and it has a page of its own. Yet this trouble does not lift when the pile grows. A site carrying forty thousand pages has it worse. Some machine generated that data, off one template, out of one database.

You do not have a small sample. What you have is a large one that keeps agreeing with itself.

Nothing in the Software Objects

Fit a model to that data and a number comes back. No warning appears anywhere. Nothing turns red.

The number will look like a finding. It arrives carrying a confidence, and that confidence got manufactured by counting one website over and over. No error message exists for it. None is ever going to.

So the failure is silent, which is much of why it survives. A broken script announces itself before lunch. This one publishes.

The Big Studies Have the Same Shape

The published ranking studies run into a softer version of the same thing.

A million URLs sounds enormous. Only a million URLs is not a million websites. Pages inside one website are alike for reasons that have nothing to do with how well they rank. Crawl more data and the number gets more confident. Meanwhile much of that confidence arrives from websites turning up over and over.

None of which makes any particular study wrong. It does mean the interesting part is the method rather than the headline. Almost nobody reads that far.

Somebody Will Point at the Statistician

Fair enough. It is a real objection with real answers behind it.

Cluster the errors. Fit a mixed model. Better still, treat the website as the unit rather than the page.

Take that last one seriously for a moment. Treat the website as your unit, then go and count what you are holding.

One.

So the correction never hands your four hundred back. Rather it tells you the truth about the one. That is worth knowing. It was never the number anybody hoped for.

The Half of Data Science That Survives

None of this touches the better half of the job.

Much of the work is description instead of inference. Counting. Joining two records that have never met. Sorting a thing far too large to look at by eye. Reading a year of server log data, then finding the pages nobody has ever visited.

Not one of those operations needs independence. A log file has never once needed a p-value. Which of my pages went a year without a visitor is arithmetic. Then the answer is a fact rather than an inference.

So the rule comes out short. Counting is safe at any size. Concluding is not.

Use data science to see. Then stop before you use it to prove.

Where the Data Science Argument Stops

Nothing here says your numbers lie. Search traffic really did fall last month. Your clicks are your clicks.

Independence is a matter of degree, as well. A varied website behaves better than a uniform one. Meanwhile a study drawing on many sites stands on firmer ground than one drawing on a handful.

Then it cuts both ways. Anybody announcing from their own data that links have stopped mattering has made the identical mistake, only pointing it the other direction. The trouble was never optimism. It was the arithmetic.

How To-The-TOP! Uses Data Science

SEO Company To-The-TOP! has been doing this work in Calgary since 2007. So the data gets used the way it can be trusted. We count. Looking comes next. Then we go and find the pages nobody reads, which is a good deal of what a website audit turns up.

What nobody here will do is hand you a law about search extracted from one website. So keyword research answers to the market rather than to your own history. A Google Ads campaign answers faster, since the feedback arrives quickly enough to argue with. The portfolio shows the keywords this work has reached. No guarantees, and the usual three to six months before anything meaningful moves.

Take the finding you believe most about your own website. The one that came out of your own data. Then put a single question to it. How many separate websites is it standing on?

Several, and you are holding something. One, and you are holding a description of your own site. That is worth having. It was never a rule.

Common Questions About Data Science and SEO

Do I need a data scientist for SEO?

Almost certainly not. Counting covers most of what a business needs. Joining and sorting cover the rest. A competent SEO already does that work. Hire the specialist when the data genuinely outgrows a person.

Can data science tell me what Google’s ranking factors are?

Not from your own website. The pages inside one site are too alike for the arithmetic to carry much weight. Published studies meet a gentler version of the same problem, so read how they were built.

What is data science actually good for in SEO?

Seeing at scale. Log files, unvisited pages, a keyword list grouped by meaning, the week a number moved. Description needs no independence at all, so nothing in this argument touches it.

Is my website too small for any of this?

Size is a separate question with a separate answer. Notice that a bigger website does not solve the trouble described here. A large templated site repeats itself more, never less.

Does this mean SEO cannot be measured?

No. Counting is measurement, and your content and traffic numbers stay perfectly good. What the argument warns against is one operation only. Turning a count into a claim about cause.