Programmatic SEO means generating many pages from a dataset and a template: every row becomes a page for its own search query, such as a currency pair, a city's weather or a comparison of two tools. It works when every row carries its own, searched-for value that the template puts on display. It turns into spam when the template is supposed to replace that value; Google's spam policies call this "scaled content abuse" and "doorway abuse" (as of November 2026). AI agents don't move that line, because Google judges the page, not who or what produced it. The quickest check: would the page help someone who opened it directly, without searching first?
Sebastian, my AI employee for SEO and GEO, writes a lot of the articles on kevinwelter.com. Everything on the topic is collected on AI SEO for Google and AI Search.
What is programmatic SEO?
Three building blocks:
- A search pattern of head term plus placeholder, like "[city] weather" or "[tool A] vs [tool B]", searched with many different values.
- A dataset that holds the answer for each variation: forecasts, rates, prices, features.
- A template that turns each row into a readable page with heading, table, copy and links.
Programmatic SEO examples: value per row
Google itself names "sports scores, weather forecasts, and transcripts" as helpful automated content. The last row is where the line gets drawn.
What Google's spam policies say (as of November 2026)
Two patterns in Google's spam policies for web search (last updated August 28, 2026) apply directly.
Scaled content abuse: "many pages are generated for the primary purpose of manipulating search rankings and not helping users", with little to no value, "no matter how it's created". Examples include many pages generated with AI without added value, and pages that make little sense to a reader but contain search keywords. Google introduced the policy in March 2024, explicitly for automation, human effort or both.
Doorway abuse: pages created to rank for specific, similar queries that lead users to intermediate pages less useful than the final destination. Two of the four examples fit template projects exactly: pages targeted at specific regions or cities that funnel users to one page, and substantially similar pages that are closer to search results than to a clearly defined, browsable hierarchy.
Three more points from Google's documentation:
- The "Why" question. Google's page on helpful content (last updated October 5, 2026) calls "Why" perhaps the most important question: content should be useful to visitors who come to the site directly. Warning signs include extensive automation for content on many topics.
- Review AI output, metadata included. Google's guidance on generative AI content (last updated October 1, 2026): models predict likely word sequences rather than retrieve facts, so fact-check everything by hand before publishing, including titles, meta descriptions, structured data and alt text. How to test structured data is in Schema Markup Checker: Which Tool Tests What.
- Near-duplicates get clustered. Pages with very similar main content get grouped, one becomes canonical, the rest get crawled less often. That alone isn't spam.
Does programmatic SEO still work?
Yes, under conditions you can check. My working list, derived from the policies:
- Every row has its own value. Your own data, measured values, verified prices, availability. An empty row doesn't become a page.
- The search pattern actually exists. Check demand per variation instead of assuming it.
- The page answers the question itself, instead of sending the visitor somewhere else.
- Pages sit in a hierarchy of overview, groups and detail pages that a person can browse. What that structure looks like is in Pillar Pages and Topic Clusters.
- Whatever adds no value stays out. Google's own advice on scaled content: if you host such content, exclude it from Search.
So is programmatic SEO worth it? If you own a dataset people search for, it can be. If all you have is a list of cities or industries, almost never, the usual case for a small business. What helps small sites more is in SEO Basics for Small Businesses.
Three tests for the line between data and spam
The swap test. Replace the placeholder (city, industry, product) with a different value. If the page is still true afterwards, the row carries no value of its own. An "SEO agency in Austin" page that reads just as well for Dallas is the textbook doorway page.
The direct-visit test. Someone gets the link from a colleague, not from a search. Do they stay, or do they search again?
The similarity test. Measure how much text two pages share. On November 6, 2026, Sebastian compared three industry articles on this site, using five-word sequences in the German text: between 0.9 and 2.7 percent of one page's sequences also appear in another, mostly a shared intro sentence and link text. A template version of the trades article with three terms swapped, such as "tax firm" for "trades", shares 91.7 percent with the original. It's a heuristic: Google publishes no threshold, and low overlap alone doesn't make a page useful. It does show quickly where only a few words were swapped.
What AI agents change
AI agents make text nearly free. The bottleneck moves to the dataset and the checking, which is where agents help.
An example of the typical mistake: an agent gets 300 cities and the task to write one page per city. It does that cleanly, and every page is a doorway page. The brief was wrong, not the agent.
How an agent performs with and without a paid data tool is in Sebastian's own case study, AI Employee With and Without an SEO Tool. How an audit finds existing template pages and duplicates is in How to Run an SEO Audit With AI, and whether your pages make it into the index shows up in Search Console (Setting Up Google Search Console). How a single article gets written with AI without sounding generic is in SEO Content Writing with AI.
A lot of posts isn't programmatic SEO
On August 27, 2026, kevinwelter.com had 10 blog posts; today the blog index lists more than 400 per language. In the second build-out stage, September 6 to 7, 2026, 76 new posts went up in German and English. That's volume.
The difference is the process. Every post is written to a guide and a fact sheet, with a search phrase from real Google suggestions, the answer first and sources, then checked by a test script and an editor in a separate role. I proofread, answer Sebastian's questions and approve; I proofread all 76 posts of that second stage and answered 42 questions along the way. Nothing goes live without me.
Honesty includes what went wrong. On September 18, 2026, an independent cross-check found factual errors the internal editing had missed; since then, large deliveries get an outside check. On September 19, 2026, Sebastian found that 205 of 4,770 search queries pointed to more than one of my pages. Similar posts compete when nobody maintains the internal links. How Sebastian finds missing links is in Internal Linking: Finding the Links You Miss.
Google's warning sign about "lots of content on many different topics" applies to this blog too. My answer: the topics follow what I do, Kubernetes, which I wrote a book about, Claude and my AI workforce. Whether Google agrees is Google's call.
The counterexample: industry pages
The obvious template would be "AI for [industry]": one text, twenty industry names, twenty pages. The overview AI by Industry: Examples for Your Business goes the other way: one page per industry, with what applies only there.
That's what a template can't do: it would give every industry the same case study.
Frequently asked questions
Is programmatic SEO against Google's guidelines?
No. Google's spam policies target many pages made mainly to manipulate rankings with little value, and doorway pages for similar queries. Dataset pages that each deliver their own answer aren't what those policies describe.
How many pages is too many?
Google gives no number; what matters is whether each page helps people. A thousand pages with real data can be fine; twenty city pages with the same text aren't.
What should I do with existing template pages?
Run the swap test, then decide page by page: fill it with real content, merge it into another page and redirect, or take it out of Search.
Where to go from here
Pick five pages on your site that follow the same pattern and run the swap test on each. If you'd rather not build the process yourself, Sebastian is in my community AI Employee Framework, as a ready-made employee folder in the Premium tier. For an assessment of your site, get in touch via the contact page. How visibility in AI answers works is in What Is Generative Engine Optimization?. How Sebastian tracks pages in AI answers is in AI Visibility Tracking: How I Measure It. How the numbers become a fixed weekly report is in SEO Reporting With AI: A Weekly Report. What an llms.txt does and doesn't do is in llms.txt: Hype vs. Evidence.
