Programmatic SEO Without a Google Penalty: Five Tests for Every Template Page

Programmatic SEO is fine. Copy-and-swap pages are not.

You have one good service page. A keyword tool says thousands of people a month search "AI for law firms." Real estate agents and accounting firms search too.

So you copy the page eleven times. You swap the industry name. Then you wait.

With AI, that whole plan takes one afternoon. That is exactly why Google wrote a rule about it.

Programmatic SEO means building many pages from one template and a table of data. Google does not ban it. It bans the shortcut: pages that exist mainly to rank and give the reader nothing new.

This guide gives you five tests to run on each page before Google sees it. Then it shows how to keep a big page set honest after launch. You will not have to read every page yourself.

Is programmatic SEO against Google's rules?

Programmatic SEO breaks Google's rules only when the pages exist to rank instead of to help. Two of Google's spam policies cover it.

Scaled content abuse is many pages made mainly to game rankings. Doorway abuse is many near-identical pages built to catch similar searches.

Notice what the rules ignore. Who wrote the page does not matter. Google says the policy applies to pages made by software, by people, or by both.

The page count does not matter either. Purpose and value do.

And Google keeps score. After it tightened these rules, it reported 45 percent less low-quality, unoriginal content in results. It had expected 40.

Our take: a rule with a public scoreboard is a bad rule to bet against.

Google edits its policy pages, so the exact wording sits in the dated box below. Read it once. The tests in this guide do not depend on the wording.

What is a doorway page? The swap test spots one in ten seconds.

A doorway page is a page built to catch one search and offer nothing the main page does not. The fastest way to spot one is the swap test.

Take a finished page. Replace the industry name with a different industry. Read it again.

Any sentence that still reads as true fits both industries. So it says nothing about either one. "We help law firms save time with AI" works for plumbers too.

"Intake screens each caller for conflicts of interest before booking" does not. Plumbers have no conflicts check.

Our rule: ignore the menu, the form and the button. If more than a third of the body survives the swap, the page is not written yet.

How do so many sites end up with doorways? Same way every time.

How to build fifty doorway pages in four easy steps:

1) Write one page about your service.

2) Ask AI for fifty versions, one per industry.

3) Change the headline and the stock photo of two people shaking hands.

4) Publish all fifty on a Friday. Call it a content strategy.

That is not a page set. That is one page wearing fifty name tags.

Programmatic Guardrails: five tests on every page before Google indexes it

The Programmatic Guardrails is our five-test check, run on each page, never just on the template. A page enters Google's index only when it passes all five.

Test 1, demand: people search for this exact page, or a close version of it. No searches, no indexed page.

Test 2, swap: less than a third of the body survives a swap of the industry name.

Test 3, own facts: the page holds at least three facts true only for this industry, like a rule, a workflow or a number.

Can't name three? Then you do not know the industry well enough to publish yet.

For law and CPA firms, our article on where AI pays first in professional firms shows what own facts look like.

Test 4, path: a visitor can reach the page from a hub in two clicks or fewer. A hub is one page that links to every page in the set.

Google finds pages through links it can crawl, and so do people. Our hub-and-spoke guide shows the structure.

Test 5, hold: until a page passes the first four, it carries a noindex rule. Google's policy says to exclude scaled content like this from Search.

Noindex hides the page from Google. It does not delete it. Your sales team can still send it to a prospect.

Build one page per buyer job, not per keyword combination

The demand test is where most page sets should shrink. A keyword tool lists every combination. Buyers do not search every combination.

The chart below is one keyword pull for "AI for [industry]." Law firms led by a wide margin. Several industries with big ad budgets, like med spas, showed no measurable searches.

Those buyers exist. They just do not search that way. A search page for them is a page for nobody.

There is a better way to slice the set. In Competing Against Luck, Clayton Christensen's team describes how one baking soda brand became many products. Each product matched one job: freshen the fridge, clean the carpet, deal with the cat litter.

Same brand. Same base product. A different job each time.

Our reading for search: give each page one job the buyer is hiring you for. "Screen new client calls for a law firm" is a job. "AI for lawyers in Tampa" is a keyword combination.

Pick jobs you can actually win with our guide to choosing searches that pay. Not sure your buyers use Google at all? Read where buyers actually ask the question.

Catch a bad page at the cheapest stage

A page set is a small factory. Every bad page costs more the further it travels.

Andy Grove made this point about real factories in High Output Management. Catch a bad part when it arrives from the supplier, not after it reaches the customer.

For pages, the stages run in this order: the data row, the brief, the first drafts, the live page.

Check the data row first. Are the three own facts for each industry real, sourced and written down before AI writes a word? A wrong fact in the row becomes a wrong fact on every page that uses it.

Then read every word of the first ten pages from any new template. That is a gate. Nothing moves until it passes.

Most teams check last, after the pages are live. That is the priciest moment to find a typo. It is the worst moment to find a false legal claim.

Once the set is live, sample it, and sample harder after a miss

Reading every page forever does not scale. Reading none is how a template rots.

The factory fix is called variable inspection. You check less after a long clean run, and more the moment problems appear.

Here is our version for page sets. Read one page in ten from each new batch, picked at random.

One failure? Read one in three from the next batch.

Three failed samples in a row? Stop the template. Fix it, then recheck every page made since the last clean sample.

Five clean batches in a row? Drop back to one in ten.

Why stop the whole template? Because templates fail quietly. A changed prompt, a broken data column or a model update can spoil every page made after it, and nothing looks broken.

One rule never relaxes. Any page that states a legal duty, a price, a dose or a tax rule gets read by a person before it is indexed.

A wrong claim there is not a ranking problem. It is a trust problem.

Let a script catch the repeats before a person reads

Sampling works better when a script clears the obvious junk first. AI researchers face the same problem with generated training data.

In AI Engineering, Chip Huyen lists the simple filters the Self-Instruct researchers used on generated examples. They dropped repeats, examples too short or too long, outputs that only echoed their input, and one question given different answers.

Those filters were built for training data, not for Google. Our reading: they map almost one to one onto generated pages.

Thin. Flag any page with fewer than 300 words that appear on no sibling page.

Echo. Flag any page whose body mostly restates its own title.

Repeat. Flag any body sentence that appears on more than three pages.

Contradiction. Flag two pages that answer one question differently, like two prices for the same service.

The same book warns that too many examples from one template hurt variety. So split a big set across several templates, one per type of buyer job.

A script costs almost nothing to run on every page. Your attention is the scarce part. Save it for what no script can judge: whether the facts are true.

What if competitors rank with hundreds of thin pages?

Some thin page sets do rank, for a while. We cannot see inside Google, so we will not promise when or whether any site gets hit.

We can tell you what the bet is. Google's own guidance asks whether content looks mass-produced, so pages get little care. That question is about a whole site, not one page.

There is a plainer problem too. A buyer who lands on a page that could be about anyone does not call.

Even when a copy-and-swap page ranks, it does not sell. Ranking without selling is a vanity metric with a hosting bill.

AI is not the problem either. Google's note on AI content says AI helps with research and structure. Use it to draft, then add the facts only you know.

Stuck with old near-duplicates? Fold them into one page with Google's guide to duplicate URLs. Then show who wrote each page with our E-E-A-T checklist.

Indexed, but nobody clicks? The page probably passes the tests and still answers no real buyer job. Our page-two article covers what to change next.

Run the five tests with AI

The prompt below runs the Programmatic Guardrails on one page at a time. It asks for the page, one sibling page and your search numbers. It never invents a number or an industry fact.

This week: pick your three best industry pages. Run the prompt on each and fix every failure.

Then write your sampling rule down before the next batch ships. After that, check how AI assistants treat those pages with what AI engines actually cite.

Eleven copies of one page are still one page. Write eleven real pages, or write fewer.