Programmatic SEO in 2026

Two facts about programmatic SEO that both became true this year.

Google's March 2026 enforcement of the scaled content abuse policy stripped somewhere between 50 and 80 percent of traffic from low-value programmatic sites. Whole directories of [keyword] in [city] pages went from a business to a graveyard inside a single update cycle.

At the same time, Zapier is still pulling roughly 16 million organic visits a month from more than 50,000 integration pages. Wise still does around 60 million a month from currency converters. Canva does over 100 million from its template gallery. All three are programmatic SEO. All three are template-generated pages at enormous scale.

So programmatic SEO is not dead. A specific version of it died, loudly, and the version that survived looks so different that people keep confusing the two.

I have been on both sides of this. I have built the thin version, watched it index, watched it do nothing, and quietly deleted it. I have also built the version that works, which is justship.now, a directory where every entry is something I have an actual opinion about. The difference between those two projects is not technical. The code is nearly identical. The difference is what sits inside the template.

Here is what actually separates them.

The Only Question That Matters

Strip away every tactic and programmatic SEO reduces to one test:

Does the page contain something that does not exist anywhere else on the internet, and could a human not reasonably have written all of these pages by hand?

Both halves are required.

If the answer to the first is no, you have generated a page whose only reason to exist is a keyword slot. That is precisely the thing Google's spam policy names: producing many pages primarily to manipulate rankings rather than help users, with generative AI pages that add no value called out explicitly.

If the answer to the second is no, you did not need a program. Write the twelve pages by hand and stop calling it a strategy.

Zapier passes both. Each integration page describes a specific pairing of two products, with the actual triggers and actions that pairing supports, pulled from live data. Nobody could hand-write 50,000 of those, and each one contains information that exists nowhere else in that combination. Wise passes both: live exchange rate data, per currency pair. Canva passes: an actual template you can open and use.

A page that says "Looking for the best CRM in Cluj-Napoca? Here are our top picks for CRM in Cluj-Napoca" passes neither. It has no unique information and there was never a reason for it to exist besides the phrase in the H1.

What Google Actually Changed

The 2026 spam policy defines scaled content abuse as producing many pages primarily to manipulate rankings rather than help users. The named examples are the ones you would expect: generative AI pages with no added value, scraped feeds, stitched content, doorway-style keyword pages, and networks that hide their scale.

Three things about the current enforcement are worth understanding properly.

Intent matters more than method. The policy says nothing about how the pages were produced. AI-generated is not automatically spam and hand-written is not automatically safe. What is being evaluated is whether the pages exist primarily to rank. A programmatically generated page with real data is fine. A hand-written page farm is not.

The classifiers got much better at near-duplicates. The template-and-spin approach, where you swap a variable and change three sentences, used to survive because the pages were technically distinct strings. Near-duplicate detection in 2026 is comparing meaning, not text. If your 400 pages say the same thing with different nouns, they are one page as far as the system is concerned.

Manipulating AI answers is now explicitly in scope. This is the newer wrinkle and people have not caught up to it. Google's spam policy now names attempts to manipulate generative AI responses in Search, not just traditional rankings. So the "spam the AI instead" plan that circulated after AI Overviews launched has been closed off in policy terms, whatever the enforcement reality turns out to be.

The result of March was blunt. Sites built on the thin pattern lost 50 to 80 percent of their traffic. Not a ranking wobble. A structural removal.

Why AI Search Made This Worse for Thin Pages Specifically

There is a second force acting on the same pages, and it is arguably the bigger one long-term.

When an AI Overview appears, the top organic result loses about 58% of its clicks. That loss is not distributed evenly. It falls hardest on exactly one kind of page: the definition-style page whose entire content can be summarized in two sentences.

Think about what that means mechanically. If your page's value is "here is what a CRM is and here are five of them," an AI Overview reproduces that value completely, above your result, for free. There is no reason left to click. The page did not get penalized. It got made redundant.

Pages with real data or actual utility survive that, and they do something better than survive: they get cited. A currency converter cannot be summarized away, because the value is the interaction. An integration page listing specific triggers and actions cannot be summarized away without reproducing the data, which is a citation.

This is the through-line from what I wrote about generative engine optimization earlier this year. The pages that win in AI search are the pages that contain something an answer engine has to point at rather than paraphrase. Programmatic SEO is not exempt from that rule, it is the most exposed to it, because programmatic pages are the ones most likely to be pure summary.

The Data Layer Is the Whole Strategy

Here is where most programmatic SEO projects go wrong, and it happens before a single line of code.

People start with the keyword pattern. "I'll do best [tool] for [industry], that's 40 tools times 30 industries, 1,200 pages." Then they go looking for something to put on those 1,200 pages, discover they have nothing, and fill the gap with generated prose.

Do it the other way around. Start with a dataset you have that other people do not, then ask what page patterns that dataset supports.

The question to answer honestly is: what do I know that is not already published?

Some answers that actually work for indie hackers:

Your own product's data. If you run a tool that processes things, you have aggregate numbers nobody else has. Usage patterns, failure rates, average values, distributions. One page per segment, each with real figures.

Structured comparisons you have actually made. Not "X vs Y" pages generated from marketing copy, but comparisons where you have used both and recorded specifics. This is slower and it is the reason a comparison page survives.

A live feed of something. Prices, availability, statuses, version numbers. The value is in freshness, which is inherently uncopyable and inherently unsummarizable.

Community-generated content with real moderation. Reviews, submissions, entries, each with something specific in it. This is the directory model, and it works precisely to the degree that the moderation is real. I have made this argument at length about directory submissions and what they are actually worth: an unmoderated list of links is worthless in both directions, to the person submitting and to the person running it.

Calculators and converters. The highest-surviving category, because the page's value is a computation and a computation is not a paragraph.

If none of those apply to you, that is genuinely useful information. It means programmatic SEO is not your channel right now, and the honest move is to write twenty good pages instead of generating two thousand bad ones.

The Page-Level Checklist

Once you have a dataset worth publishing, the per-page requirements in 2026 are stricter than they used to be but not complicated.

Unique data above the fold. The specific thing this page knows should be visible without scrolling. Not an intro paragraph explaining the category. The data.

A real reason for this page to be separate from its siblings. If the answer for /compare/a-vs-b and /compare/a-vs-c is "the second noun changed," merge them.

Internal links that follow the data, not a footer. A block of 50 identical "related pages" links at the bottom of every page is a pattern that gets recognized. Links that follow actual relationships in your dataset, three to six per page, are structure.

No page without content. The classic programmatic failure is generating a page for every possible combination, including the 60% of combinations for which you have no data. Those empty pages drag the whole set down. Generate only where the data exists, and let the rest 404 or redirect.

A crawl budget you can actually justify. If you publish 5,000 pages on a domain with no authority, most will not get crawled, and the ones that do will be judged as a set. Publishing in batches and watching what indexes beats dumping everything on day one.

Schema markup that matches reality. Structured data helps AI systems parse what your page contains. It also gets you in trouble faster if it describes something the page does not have.

The Part About AI-Generated Copy

Let me be direct about the thing everyone is actually doing.

Using an LLM to write the prose around your data is fine. It is not a policy violation, Google has said so repeatedly, and pretending otherwise is theater. What is a violation is using an LLM to manufacture the substance, because then the substance is a plausible-sounding average of the internet, which is by definition not unique information.

The distinction in practice:

Fine: you have 800 rows of real data, and you use a model to turn each row into two readable paragraphs of context around it.

Not fine: you have a list of 800 keywords, and you use a model to write 800 articles about them.

The first has an information source and a presentation layer. The second has only a presentation layer, and the thing being presented is nothing.

I use models to write around data constantly. Every one of my pSEO-shaped pages has generated prose in it. The prose is not the product. The data is the product, and if you deleted the prose the page would still be useful, which is the test.

What I Would Actually Build

If I were starting a programmatic play this month for a small product, here is the sequence.

Week one: find the dataset. Not the keywords. The dataset. Go look at what your app already stores, what you could scrape legitimately, what you could compute, or what you could get people to submit. If you cannot name the unique information in one sentence, stop here.

Week two: build twenty pages by hand. Actual hand-written, using the data. This tells you two things: whether the page is genuinely useful, and what the template needs to contain. Almost every template I have built was wrong in ways that only became obvious after writing a few by hand.

Week three: check whether anyone is searching. Now do the keyword work, and do it against pages that exist. If the twenty hand-written pages get zero impressions in Search Console after three weeks, generating two thousand more of them will get you zero impressions two thousand times.

Week four onward: scale in batches. A hundred pages, wait, measure indexation and impressions, then the next hundred. Slower than a bulk publish and dramatically less likely to get the whole domain classified as a content farm.

That timeline is deliberately unexciting. Programmatic SEO in 2026 is a data project with a publishing layer bolted on, and the data project is the hard part. Everyone wants it to be a publishing project with a data problem to solve later, because publishing is fast and data is slow.

The Uncomfortable Conclusion

The version of programmatic SEO that made people money in 2021 is gone and it is not coming back. That version was an arbitrage: producing pages was cheap and evaluating pages was expensive, so volume won. Both halves of that flipped. Producing pages is now free, which destroyed its value, and evaluating pages is now cheap, which destroyed the arbitrage.

What is left is the thing that was always actually valuable and never actually easy: having information other people do not have, and publishing it at a scale a human could not.

This is the same argument I keep ending up at from different directions. Building is easy and distribution is the moat, and the reason distribution is a moat is that it is made of things that do not commoditize. A dataset you built is one of those. A template is not.

If you have the data, programmatic SEO is more effective in 2026 than it was in 2021, because the field cleared out. If you do not have the data, no amount of technique will save you, and the honest thing to do is go get some.

Frequently Asked

Is programmatic SEO dead in 2026? No. Thin, template-and-spin programmatic SEO is effectively dead after Google's March 2026 scaled content abuse enforcement, which cut 50 to 80 percent of traffic from low-value programmatic sites. Data-backed programmatic SEO is working better than before because the low-quality competition was removed.

Will Google penalize AI-generated programmatic pages? Google evaluates intent and value, not production method. AI-generated pages built around real, unique data are fine. AI-generated pages that manufacture the substance itself are exactly what the scaled content abuse policy targets.

How many pages is too many? There is no number. A thousand pages backed by a thousand real data points is fine. Fifty pages that say the same thing with different nouns is not.

Does programmatic SEO still work with AI Overviews? For utility and data pages, yes, and they earn citations. For definition-style pages, AI Overviews reproduce the value above your result and the top organic result loses roughly 58% of its clicks.

What is scaled content abuse? Google's policy term for producing many pages primarily to manipulate rankings rather than help users. The 2026 version explicitly includes attempts to manipulate generative AI responses in Search.

Where do I start if I have no dataset? You do not start. Programmatic SEO is not your channel yet. Write twenty pages by hand about something you actually know, which is what moved the needle for me long before any of this was automated.

Go look at your own database this afternoon. The dataset you need is more often sitting in a table you already have than out on the internet waiting to be scraped.