"Programmatic SEO is dead" has been said every year since the category existed, and it's wrong in the same way every time: it conflates the production method with the output quality. Generating pages from a template didn't die. Generating the same page a thousand times with one variable swapped did — and AI Overviews just made that failure mode visible faster than Google's algorithm ever did.
Here's the test that actually separates the two, and how to apply it before you publish the next batch.
The test
For any page in a programmatic set, ask one question: what does this page know that its siblings don't?
Not "what's different about the URL" — the city name, the integration name, the year. What does this specific page know — a number, a named entity, a claim — that the template alone couldn't have produced for every other page in the set.
If the honest answer is "nothing, it's the same paragraph with a find-and-replace," the page shouldn't exist. If the answer is real (this integration has a rate limit the others don't, this city's ordinance changed in a specific way, this tool's free tier caps at a specific number), the page earns its slot.
This isn't a stylistic preference. It's the mechanical reason AI Overviews expose thin programmatic content faster than search rankings ever did.
Why AI answers make this test non-optional
Roughly 40% of searches now surface an AI-generated answer before any organic link. Those answers need a source to cite, and an LLM extracting a citable claim is doing something closer to fact-checking than to ranking: it's looking for a specific, verifiable statement it can lift and attribute. A templated paragraph that could apply to any of 500 sibling pages gives it nothing to extract — there's no claim in it that's true of this page and not the others. It gets skipped, and skipped silently; you don't get a penalty notice, you just get zero citations and don't know why.
A human skimming a search results page will tolerate a mediocre page if it ranks. An extraction pass has no equivalent tolerance — it either finds a citable claim or it doesn't.
What passing the test looks like in practice
The pipeline is unchanged: keyword research → topical map → page generation. What changed is where you put a quality gate.
Topical map stage — before generating anything, write down, per planned page, the one fact that page will carry that its siblings won't. If you can't fill that in, don't create the page. This turns "how many pages should we make" from a volume decision into a data-availability decision — you're capped by how much real per-entity information you actually have, not by how many rows are in your keyword export.
Brief stage — put that unique fact in the brief as a required element, not a nice-to-have. A brief that just says "write about X for city Y" produces exactly the interchangeable page the test is designed to catch.
Generation stage — check the output against the brief mechanically: is the required unique claim actually present, verbatim-checkable, in the first few sentences? This is exactly the kind of pass/fail check that's easy to automate, because "is this specific fact present" doesn't require judgment — it's a lookup.
Citability formatting, on top of the uniqueness check — once a page has something worth citing, make it extractable: state the specific claim in the first 40-80 words rather than burying it in paragraph four, use FAQPage schema so the Q&A structure is machine-parseable, and name concrete entities (numbers, dates, product names) instead of vague language. A page can pass the uniqueness test and still fail to get cited if the unique claim is buried under three paragraphs of throat-clearing.
What this doesn't mean
It doesn't mean smaller programmatic sets are automatically better, and it doesn't mean hand-writing beats generation. A hundred pages built from real per-entity data will outcite a thousand pages built from one template with a variable swapped — but a thousand pages each carrying a real, distinct fact are fine too. The gate is uniqueness, not headcount, and it applies identically whether a human or an agent is doing the writing. Google's own helpful-content guidance backs this structurally, if not by name: it penalizes unhelpful pages regardless of who or what produced them.
If you're running this pipeline with an agent doing the generation step, the uniqueness check is the one part worth keeping firmly in human (or at least human-reviewed) judgment — an agent can competently verify a brief was followed, but "is this fact actually interesting and true" is a call that benefits from someone who knows the domain. We built the citability half of this into our SEO, AI & Syndication Kit, but the test above works whether or not you ever look at what we sell — it's just the question worth asking before the next batch of pages goes out.
Full answer with the FAQ breakdown: https://agentkitworks.com/answers/is-programmatic-seo-dead
