Programmatic SEO: How It Works, Examples and the Line to Spam

On 9 September 2026, like every morning, I went through the SEO news over coffee. That is how I ran into a post by John Mueller about programmatic SEO. It stayed with me.

Not because of the basic message. That Google takes a dim view of pages produced in bulk has been common knowledge in the industry for years. What stayed with me was a spot in the middle of a paragraph that slips past on a quick read: Mueller is not talking about pages but about the site.

This overview grew out of that. The first part explains the term, how it works, examples and implementation. The second part shows where Google draws the line and what Mueller said about it in September 2026.

A day later I read the 16 top pages for „programmatic seo“ in Germany and the US in full. 15 of them could be retrieved. All of them mention risks. The rule „scaled content abuse“ is named by two of them, the doorway rule by three, the Rater Guidelines by none. That is exactly where this article comes in.

Programmatic SEO is generating many pages from three ingredients: a keyword pattern that selects the pages, a dataset with the content of each individual page and a template built the same way for all of them. It is an industry term: it does not appear in Google’s spam policies as of the version of 28 August 2026.

The pattern consists of a fixed part and a changing part, called head term and modifier in the industry. „Wetter“ plus a place name, „Elektriker“ plus a city, „EUR to USD“ plus the next currency pair: every combination becomes its own URL.

The difference from a classic article is not quality but the ratio of effort to page. An article is written individually at a desk. A programmatic page is created as a row in a table. A hundred rows turn into a hundred pages.

Programmatic SEO and classic SEO compared

This is how I classify the two approaches:

Technically programmatic SEO is simple. You take a list of cities and a list of industries and get one page per combination by multiplying: 50 cities times 20 industries make 1,000 URLs without a single written line of text.

That is its strength and its problem at the same time. The template makes sure all pages look equally tidy. It does not make sure there is something of its own on each page. That comes from the dataset alone: from rates, forecasts, listings, prices or ratings for exactly this one combination.

Three building blocks decide the quality of a page family:

• The dataset. Own, licensed or public data in a sensibly prepared form. The less a row contains, the thinner the page becomes.
• The template. It defines which fields appear where. It can generate structured data along the way. What structured data can and cannot do is covered in Is structured data your advantage.
• The hierarchy. Overview pages that organise the individual pages by region or category. In its spam policies Google lists among other things substantially similar pages as doorway abuse that are closer to search results than a clearly defined and browseable hierarchy. What such a structure can look like is described in my hub-and-spoke model.

Naming examples is easy. A direct comparison says more. That is why I took two sibling pages from each of five page families and noted what stays the same and what changes.

Zapier explicitly lists its app directory as an example of programmatic SEO on its own blog. For the other four I only describe the openly visible structure.

What all five have in common: each page has something of its own compared with its sibling page, such as a different currency pair, a different location or different listings. The benefit sits in the data of the row, not in the template.

I check three conditions before building a page family at all. The list is my own assessment and not guidance from Google.

• There is data with its own value. Rates and forecasts, prices and ratings, availability or your own measurements. A place name alone is not a dataset.
• There is demand per combination. Do people search for exactly this combination of head term and modifier? A cross product reliably also produces combinations with no measurable demand.
• Every page passes the neighbour test. Two pages side by side and everything they share deleted: what is left has to give its own reason to visit.

The advantages follow from the construction: a system built once serves many queries at the same time, updates run through the dataset instead of hundreds of individual texts and new combinations cost little technically. Maintaining the data is still work.

The disadvantages are the flip side of the same property. An error in the template ends up in all pages at once, thin data becomes thin a thousand times over. A bad dataset also cannot be compensated for by more pages.

1. Define the pattern. Head term and modifier determine how many pages there can be. Write down the full list of modifiers before you build anything. Then you see the size of the project in black and white.

2. Check demand. Get search volumes for a sample of the combinations and delete everything with no measurable demand. For this article I did that with the head term itself: according to my query through DataForSEO, „programmatic seo“ sits at around 90 searches a month in Germany and at around 590 in the US. Both are estimates produced by that tool.

3. Build the dataset. The quality of every page is decided in its row. So run the neighbour test on the dataset already: two random rows side by side show you whether you end up with two different pages or the same one with a different name.

4. Build template and hierarchy. One template for all pages plus overview pages for browsing the family. A page reachable only through search lacks exactly this hierarchy.

5. Publish in stages. Start with part of the pages instead of all of them. Submit them via sitemap and watch Google’s reaction in the Page indexing report. According to Google, submitting a sitemap is merely a hint and does not guarantee that Google uses it for crawling. According to Google a single sitemap holds at most 50 MB uncompressed or 50,000 URLs. Larger amounts you split across several sitemaps.

Google aims its crawl budget guide mainly at three groups: sites with a million pages or more and weekly changes, sites with 10,000 pages or more and daily changes and sites with a large share of URLs in the Search Console status „Discovered - currently not indexed“. In my view the third group in particular concerns page families. For everyone else, according to Google, an up-to-date sitemap and a regular look at the Page indexing report are adequate.

6. Monitor and clean up. Measure the ratio of submitted to indexed pages per group and look at the impressions. Groups without indexing or without impressions you merge or delete. How I go about that is described further down under What to do if it does.

Tools by task

I deliberately do not recommend specific products here. In my view the choice of tool is secondary as long as these four tasks are covered:

• Holding data: a spreadsheet or a database in which every row is a page
• Generating pages: a CMS with templates and data collections or a static site generator
• Keeping data in sync: an automation that transfers changes from the dataset to the pages
• Checking: a crawler for titles, duplicates and status codes as well as Google Search Console for indexing and performance

Google’s spam policies describe practices, not methods. Two of them matter most for programmatic SEO. A third source shows the standard Google uses to have such pages assessed.

Scaled content abuse

The rule arrived in March 2024. In the current version of the spam policies from 28 August 2026 the definition reads:

„Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users.“

Among other things, Google lists these five practices as examples:

• using generative AI or similar tools to generate many pages without adding anything for users
• scraping feeds, search results or other content and turning it into many pages, also through automated rewording or translation
• stitching together content from different pages without adding value
• running several sites to hide the scaled nature of the content
• generating many pages that make little sense to readers but contain search keywords

How the content is produced plays no role. According to Google the rule applies „no matter whether content is produced through automation, human efforts, or some combination of human and automated processes“. If you host such content on your site, it should be excluded from Search according to Google. How Google classifies AI texts beyond that is covered in Does Google penalise AI content.

A question from Google’s guide to helpful content fits here: „Are you using extensive automation to produce content on many topics?“ In the same guide Google asks whether you write to a particular word count. The answer comes right with it: „(No, we don’t.)“

The consequences are also in Google’s announcement from 2024: sites with violations of the spam policies may rank lower or not appear at all. With a manual action a message arrives in Search Console and a reconsideration request can be filed afterwards. How a spam update shows over time can be read in the August 2026 spam update.

Doorway abuse

The second rule matters at least as much for programmatic SEO. In my review of 10 September three out of 15 pages name it, all only as a keyword. Google’s definition reads:

„Doorway abuse is when sites or pages are created to rank for specific, similar search queries.“

Two of the examples in the rule read like a description of badly built page families:

• „Having multiple domain names or pages targeted at specific regions or cities that funnel users to one page“
• „Creating substantially similar pages that are closer to search results than a clearly defined, browseable hierarchy“

In my view, city pages with the sole purpose of forwarding to the same contact form come very close to the first example. A page family of substantially similar pages without overview pages moves close to the second.

What the Rater Guidelines say

In its documentation on AI content Google itself points to section 4.6.5 of the Search Quality Rater Guidelines. The version of 11 September 2025 says:

„Pages and websites made up of content created at scale with no original content or added value for users, should be rated Lowest, no matter how they are created.“

The sentence after that is remarkable. Raters are to give the lowest rating even without knowing how the content was created, as long as they strongly suspect scaled content abuse „after looking at several pages on the website“. The view therefore goes across several pages of a site and not across a single one.

The trigger is a public question on Bluesky: an owner describes his site, launched in April, a directory for tech stacks modelled on BuiltWith. By his own account many pages were indexed at first and then dropped out again, after which he set them to noindex. Since then, according to his account, the core pages have been sitting on „Crawled - currently not indexed“ for months, the plain about page among them.

Mueller’s reply opens with a diagnosis of the construction:

„it sounds like a programmatic SEO play, where you iterate through a large list of domain names, technologies, and attributes - finding all possible combinations.“

That is the description of a cross product: list A times list B times list C, and out falls one URL per possible combination. Technically that is a loop. Editorially it is nothing.

Then comes the sentence everyone quotes:

„Programmatic SEO like this often leads to a site that’s either spam, borderline spam, or low quality. It’s easy to spin something up with many pages, it’s hard to provide real value to users. Our systems have possibly lost faith in your site providing good value to users based on the old pages.“

Read the first three words again. Programmatic SEO like this. Mueller is not condemning the method as such, he is condemning the one implementation from his own description two posts earlier. I read three write-ups of this. SEO Südwest translates the qualifier („Programmatic SEO wie das beschriebene“), Seroundtable leaves it out of its own framing and quotes the post in full afterwards, and Search Engine Journal turns it into a point of its own: „Mueller did not say all programmatic SEO is spam.“

This position is not new for him, by the way. Back on 25 July 2023 Mueller wrote on X that programmatic SEO is „often a fancy banner for spam“. Seroundtable documents a second, qualifying post from the same day („To be fair, programmatic SEO is not always spam“). On 9 September I looked for that second post through four different routes: directly at X and through two proxy APIs and finally in the Internet Archive. It is no longer reachable through any of those four routes. The first post comes back through exactly the same route without trouble, so the method is not the issue.

This is the actual finding and it is the most expensive point for anyone with a large generated directory. The usual picture of thin pages is a calculation per page: a weak page does not rank, a strong one does and both sit side by side independently of each other.

Mueller’s wording describes something fundamentally different from that familiar page against page calculation. He writes „Our systems have possibly lost faith in your site“ based on the old pages: the object of the assessment is the site, the trigger is pages in a form with no present day equivalent.

That is my reading of the wording, not Google describing a mechanism. The thread describes no stored site rating, which is how Search Engine Journal reads it too. Mueller names no system anywhere in it.

That fits the case described. The generated pages are on noindex according to the owner and have disappeared from the index. What stays out anyway is the about page with no automation of any kind.

My conclusion for practice: the effort spent on the good content may not pay off in isolation. A strong guide on a domain with twenty thousand combination pages then does not start at zero but behind the impression that domain has made. If an objectively good text on your domain fails to take off, its neighbourhood is therefore a lead. How Google decides what makes it into the index at all is something I took apart in Crawling and Indexing.

This section matters to me more than the one before it, because this is where the write-ups tip over. A careful sentence turns into a mechanism and by the next retelling that mechanism sounds like documentation.

Four things are not in the thread:

• No name. Mueller names no ranking system, no signal, no classifier.
• No value. There is no score and no rating and no threshold for a tipping point.
• No penalty. A manual action is something else and comes with a message in Search Console. Nothing of the sort is mentioned anywhere.
• No duration. He writes „tends to take time“. That is not a deadline, that is an expectation.

On top of that the word „possibly“, written by Mueller himself. He is describing a possibility for a case he knows from four short posts. Search Engine Journal puts it plainly and correctly: the thread describes no stored site rating and it also does not say the effect is permanent.

Why I keep hammering on this: the idea of a secret domain rating is seductive because of how much it explains. It also still explains something when the cause is somewhere else entirely. A text with no demand does not rank because nobody is searching for it. A page with no internal links pointing at it is harder for Google to find. Both look from the outside like lost faith. The self-test for replaceable content asks exactly the same question from the other side: what is on this one page that is nowhere else on the web in this form.

Four checks in this order, about an hour of work in total: the neighbour test, the indexing ratio, the demand probe and a look at how crawling is distributed.

1. The neighbour test. Open two random pages from the same generated group side by side. Delete everything they have in common. What is left is the value of the page. If all that is left is the title and two numbers and one line of body text, you have your answer.

2. The indexing ratio. Pull the Page indexing report in Search Console and look at the ratio. If 900 out of 12,000 submitted URLs end up indexed, that points more to a selection by Google than to a technical glitch. On its own it is not yet a diagnosis. The status „Crawled - currently not indexed“ is the most useful one here, in my view, because Google has to have fetched the page for it.

3. The demand probe. Take twenty keywords from your generated inventory and get real search volumes for them. Does the query exist at all? A cross product reliably produces combinations with no measurable demand.

4. The crawl view. Look at how crawl activity is distributed across your areas. If most of it goes to the generated directories and the editorial pages rarely get a turn, that may have the same cause. The levers for that are in Optimize crawl budget.

Let me say up front what I cannot say: Whether a particular measure brings a particular site back is something nobody knows. Mueller explicitly does not claim it either. He points to the effort others put into spam and core update problems and makes it clear that he is not calling this site spam. That is a yardstick for the order of magnitude, not a set of instructions.

What I consider sensible, in my order of priority:

• Decide instead of hiding. Noindex conceals the pages. Deleting or merging removes them. If a group of 8,000 combination pages offers nothing beyond what 40 good pages offer, then it is 40 pages. I described the procedure for that in Content Audit and Pruning. Before deleting, take a look at the other causes for „Crawled - currently not indexed“ above under „What Mueller did not say“.
• Check demand before rebuilding. Pages for queries that do not exist are worthless in small numbers too.
• Create real evidence of value. Your own data or your own analysis or a comparison with no precedent. Something out of reach of a cross product.
• Plan for patience. Mueller’s „tends to take time“ is the only time reference in the whole thread and he names no more precise deadline anywhere.

The technique is quickly explained: pattern, dataset, template and the page family is done. The hard part is not the building but the question of what is unique to each page compared with its sibling page.

Google’s rules answer that question not with a number but with purpose and benefit. Mueller’s reply from September 2026 shows the possible level for it: he writes about the site and not about the individual page. The thread does not contain more than a „possibly“ on that, though. System, score and deadline are missing.