Templated pages are not banned in 2026, and generating them with AI is not the offence. The test Google applies is value per page — whether each generated URL carries a fact a reader could not have gotten from the template itself. Get that wrong at scale and the likely outcome isn't a manual action; it's 5,000 URLs that Google crawls once and never indexes. The gap between programmatic page sets that work and ones that don't is enormous: on Ahrefs' estimates, Wise's programmatic pages average about 314 organic visits each per month while Zapier's average about 0.38 — same technique, same search engine, more than 800× apart.
So if you're sitting on a spreadsheet of 5,000 comparison-page permutations and a generator that could ship them by Friday, that spread is the number that should decide your Friday. The penalty risk is real but small and slow. The wasted-quarter risk is large and immediate.
Most programmatic SEO advice still argues from the 2024 spam policy update. The most specific thing Google has ever published about generating a page per query variation is newer than that, and sits in its guide to optimizing for generative AI features, last updated 2026-07-10:
"While it might be tempting to create separate content for every possible variation of how people might search (for example, by focusing on other queries that people have asked, or fan-out queries), doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy."
Read that twice, because it names the mechanic much of 2026's tooling is built on: pull the AI Mode fan-out set, mint a page per branch, catch the long tail. Google now says doing that primarily to manipulate rankings or AI responses is the violation — and that qualifier is the same one that has always governed this policy. Purpose, not method. The guide adds a line worth pinning next to your page-count dashboard: "This is also an ineffective long-term strategy, as a high quantity of pages doesn't make a website higher quality or more relevant to users." That's Google saying the tactic fails on its own merits before enforcement enters the picture.
The same document hands you a usable vocabulary, too. It contrasts commodity content — its example, "7 Tips for First-Time Homebuyers," is "often based on common knowledge, which could originate from anyone" — with non-commodity content, which "provides unique expert or experienced takes that go beyond common knowledge." Swap "article" for "generated page" and you have your template test. "Best CRM software in Denver" rendered from a city list is commodity. Response times, integration coverage and pricing tiers you measured yourself are not.
So what does the policy itself say — and what would land in your Search Console if you got it wrong?
Google's spam policies page, last updated 2026-05-15, defines the offence in one sentence: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." Listed examples include "Using generative AI tools or other similar tools to generate many pages without adding value for users" and "Stitching or combining content from different web pages without adding value."
Notice the phrase doing the work in both: without adding value. Volume alone isn't the trigger, and neither is generation. The practitioner reading that best matches the documentation is that risk needs volume, manipulative intent and low value present at the same time — and that large legitimate directories with unique data per page sit outside it.
Two neighbouring policies on that page catch directory and comparison generators at least as often:
/alternatives/[competitor] set exists mainly to funnel clicks to signup, that's the policy you're nearest.Now the nuance that saves an afternoon of panic-Googling. Plenty of 2026 SEO writing asserts that sites "receive a manual action labelled Scaled content abuse." That label isn't its own action type in Google's manual actions documentation — it appears inside the "Major spam problems" description as an example of aggressive technique. The realistic notice for a thin programmatic set is "Thin content with little or no added value": "low-quality pages or shallow pages," with thin affiliate pages, scraped content and doorways as the examples. The policy name and the notice name are different things, and knowing that is the difference between diagnosing your own Search Console and pattern-matching a blog post.
One thing genuinely did change this year. On 15 April 2026, Google updated its documentation to say it may use spam report submissions to take manual action, reversing its long-standing line, and that it forwards "whatever you write in the submission report verbatim to the site owner." Publish 5,000 pages targeting named competitors and any one of them now has a documented path to putting your set in front of a human reviewer. Not a reason to skip comparison pages — a reason to build ones you'd be happy for a reviewer to read.
That's the loud risk. The quiet one costs more.
Here's the outcome founders rarely model. You ship the set and nothing happens — no notice, no traffic, no signal, just a growing pile of URLs sitting in "Crawled — currently not indexed" and a hosting bill.
Google's crawl budget guidance for large sites, updated 2026-07-22, explains the mechanism. Crawl allocation depends on "popularity, overall user value, content uniqueness, and serving capacity," and the first optimisation Google lists is: "Eliminate duplicate content to focus crawling on unique content rather than unique URLs."
That's the whole argument in seven words. Google budgets attention by unique content, not unique URL, and a template varying a city name across 3,000 pages produces one of those and three thousand of the other. Low-uniqueness sets don't get punished; they get throttled. An un-crawled page can't rank, can't be cited by an answer engine, and still costs you to build and host.
The base rate makes it worse. Ahrefs' study of roughly 14 billion pages found 96.55% get zero organic traffic from Google, with another 1.94% getting one to ten monthly visits. That's the population your generated pages join. For a 5,000-page set, the honest default isn't "some will underperform" — it's that almost all of them produce nothing unless something on each page earns its place.
Then there's the 2026 weather, which has been unkind to this page type specifically. Aleyda Solís's Sistrix analysis of the March 2026 core update, covering 26 March to 11 April, found gains concentrated in "official and institutional sites, specialist and niche platforms, established brands" — while "losses were more common among aggregators, directories, and comparison-driven sites." Job aggregators, broad travel and real-estate platforms and dictionary sites all lost visibility, in an update where SE Ranking measured nearly 80% of top-three results shifting position against 66.8% in December 2025.
None of that is a spam-policy story. It's a core update reweighting what counts as a useful result, and it landed on exactly the category a directory generator produces. Whether the demotion reverses is unknown — Google's core update documentation warns recovery "could take several months" and may mean "waiting until the next core update." A spam update shipped on 24 June 2026 too. If that rolling-update cadence is the part that worries you, we've written separately on planning a blog around continuous core updates.
So: manual action, unlikely. Algorithmic demotion of the category, demonstrated and recent. Quiet non-indexation, near-certain unless each page carries something. Which raises the obvious question — what does "carries something" actually look like?
Ahrefs' programmatic SEO guide publishes estimated page counts and monthly organic traffic for four famous programmatic programmes. Do the division:
| Site | Estimated pages | Est. monthly organic visits | Visits per page |
|---|---|---|---|
| Wise | 14,888 | ~4.67M | ~314 |
| Nomadlist | 25,873 | ~41,200 | ~1.6 |
| Webflow | 31,516 | ~27,600 | ~0.9 |
| Zapier | 800,632 | ~306,000 | ~0.38 |
(Ahrefs estimates published in 2023, not company-reported figures — treat the magnitudes, not the decimals.)
Wise earns roughly 800 times more per page than Zapier while running 54× fewer pages. All four get held up as models to imitate; three are, per page, close to noise. If your business case reads "5,000 pages × 20 visits each," you're underwriting a number only one of these four reaches — and it reaches it because of what's on the page.
What's on a Wise page? A real SWIFT/BIC code. Ahrefs' case study on Wise counted nearly 12,000 landing pages targeting SWIFT/BIC combinations for US visitors alone. The variable on each is a verifiable bank identifier plus a live conversion rate: delete the database row and nothing is left. That's the point. The data exists in the world independently of the page, a reader needs it, and it's annoying to find elsewhere. Value per page, in one sentence.
Zapier is even more instructive, because the cliff appears inside a single template. Ahrefs' Zapier case study found integration pages drove 16% of Zapier's entire organic traffic — an enormous channel. But the two-app page for Google Sheets + Dropbox drew an estimated 1.9K monthly visits across 444 keywords, while the three-app chain page for Google Sheets + Trello + Slack drew zero. Same template, same rendering logic, same company.
Quality didn't change between those two pages; demand and data ran out while the permutation engine kept going. Real people search for how to connect Sheets to Dropbox. Approximately nobody searches for a specific three-app chain, and the page has no extra fact to offer if they arrive. That is the exact moment a template stops being useful and starts being scaled content.
The rule generalises: page sets should expand along demand and data, not along permutations. Your generator can render every cell in the matrix; your business case only survives the subset where a query exists and a fact exists.
If you're still sceptical that this is about data rather than about AI authorship, the evidence there is unusually good.
Google has been consistent to the point of repetition. Danny Sullivan, Search Liaison: "We don't really care how you're doing this scaled content, whether it's AI, automation, or human beings." He paired that with a legitimate example — a retail site using AI to summarise real customer reviews, where the automation adds value instead of manufacturing pages for search traffic.
The rater guidelines say it more precisely. The January 2025 revision assigns the Lowest rating when main content is "auto or AI generated... with little to no effort, little to no originality, and little to no added value," and defines scaled content abuse for raters as volume produced "with little effort or originality with no editing or manual curation." Effort, originality, curation: three human inputs no generation pipeline supplies by default, and all three of which a pipeline can be designed to include.
The measurements agree. Ahrefs ran its detector across 600,000 pages — the top 20 results for 100,000 random keywords — in July 2025 and found 86.5% of top-ranking pages contained at least some AI-generated content, with a correlation between AI share and ranking position of 0.011. Not a weak relationship; no relationship. Their conclusion: "Google probably doesn't care how you made the content. It simply cares whether searchers find it helpful."
Graphite's October 2025 study is usually served up as the rebuttal, and it isn't one. Across 31,493 keywords, 86% of articles ranking on Google's first two pages were classified human-written, and only 7% of #1 results classified AI. Read together, the two studies say something precise: AI assistance is now the norm at the top of results and carries no penalty, while fully unedited output is under-represented there. Assistance is fine; abdication underperforms — and that applies to a generated comparison page exactly as it does to a blog post.
The figure usually quoted in the other direction deserves a caveat. In March 2024, Originality.ai analysed 79,000 websites and found 1,446 — about 1.9% — had received manual actions, all of them showing signs of AI use. But that panel was sites already monitored by an AI-detection company, in a period when most new web content showed AI signals. It establishes that enforcement is real and rare, not that AI causes it.
The clearest cautionary tale isn't statistical anyway. In 2023, marketer Jake Ward used a GPT-4-based tool to publish 1,800 articles in a matter of hours for Causal, cloning a competitor's content structure. Traffic reached roughly 610,000 visits in early October 2023, then fell to about 190,000 by mid-December — against a pre-campaign baseline near 200,000. Months of effort to land back where it started. The mechanism wasn't "AI wrote it"; it was that the pages restated facts already better served elsewhere. Our closer reading of Google's scaled content abuse policy covers that intent-and-value distinction in more depth.
Every question traces to a Google document, and each is answerable per template, not per page.
That last one is the cheapest insurance here. Skipping the pilot is how a set reaches 5,000 URLs before anyone learns whether page one deserved to exist.
None of this argues for publishing nothing. Conductor's 2026 survey of 250+ executives found 94% of enterprise organisations planning to increase AEO/GEO investment, with scaling AI content ranking as the number-one content priority. The question is what shape that scale takes.
The answer-engine data points toward depth. DeltaV Digital tracked 21,075 AI-engine responses and 25,337 citations between 14 April and 13 July 2026 across ChatGPT, Perplexity, Gemini, AI Overviews and AI Mode. Comparison pages took just 4.1% of citation share but posted a 1.87 citation rate — citations per retrieval, roughly 45% above the portfolio average and the highest of any page type measured.
Read that as an instruction: when an answer engine does retrieve a comparison page, it cites it more readily than almost anything else — so the value sits in being retrievable and differentiated, not in having many of them. A small, genuinely researched comparison set beats a permutation grid on the surface where your buyers increasingly research. Industry variance is severe, though (listicles took 61% of citations in B2B technology services and 0% in healthcare), so check your vertical. If citations are the goal, the on-page patterns that earn ChatGPT and AI Overviews citations matter more than page count.
The economics explain why judgement is now the constraint. Page-generator pricing currently runs $149/month for up to 1,000 pages, $399 for 5,000 and $899 for 20,000, with overage between $0.25 and $0.05 per page. At eight cents a page, production is effectively free — and anything that cheap is not a moat. The scarce inputs in 2026 are the data on the page and the judgement about which pages deserve to exist.
One founder put that trade-off better than any framework. Launching an AI-tool directory into a saturated market in April 2026 — automated per-tool analysis, semantic pain-point tagging instead of generic categories — they wrote: "I'd rather have 100 deeply categorized tools than 5,000 dead links." That's a day-zero post with no results attached, so treat it as strategy rather than proof — but it's the right instinct.
Practically, for a founder deciding this week:
Where a page needs research rather than a database row, the honest version of automation is one researched page at a time. That's the idea behind Magic Share's research-and-verify drafting workflow: each run reads around 20 sources, every cited URL is resolved before the draft is saved, and nothing publishes until you approve it. That's the opposite of a permutation grid — it's the manual curation Google's rater guidelines say scaled content lacks, built into the process.
Almost certainly not — and if it did, the notice probably wouldn't say what you expect. "Scaled content abuse" is a spam policy name, not a listed action type in Google's manual actions documentation; the closest label is "Thin content with little or no added value." The dominant failure mode is quieter: pages get crawled and not indexed, or demoted at the next core update. Note too that since 15 April 2026, spam reports can trigger manual reviews.
No Google policy names a page-count threshold, and asking for one is the wrong frame — risk needs volume, manipulative intent and low value together. Wise runs roughly 14,888 pages at about 314 visits each; Zapier runs 800,632 at about 0.38. The useful ceiling is the number of rows in your dataset with both real search demand and a fact worth publishing — usually far smaller than what your generator can render.
No. Google's position is method-agnostic, restated repeatedly, including Danny Sullivan's "we don't really care how you're doing this scaled content." Ahrefs found 86.5% of top-ranking pages contain some AI content with essentially zero correlation to position. What underperforms is unedited output: Graphite classified only 14% of ranking articles as predominantly AI-generated, and 7% of #1 results.
Harder than they were. Aggregators, directories and comparison-driven sites were the loser category in the March 2026 core update, while specialist and owned-data sites gained, and whether that reverses is unknown. A narrow directory built on data nobody else has assembled is still viable; a broad one assembled from public blurbs is competing directly with the category Google just demoted.
Google isn't testing whether a machine made the page. It's testing whether the page carries something a reader couldn't have gotten from the template — and at five cents a page to generate, that judgement is the only expensive part of programmatic SEO left. Build the 200 pages that pass the test, watch what gets indexed, then decide about the next 4,800.
And if the pages that need real research are the ones you never get to, that's the gap Magic Share fills: fact-checked, link-verified drafts held for your approval. Plans start free with three posts — no permutation grid required.
Magic Share researches, writes and fact-checks posts like this for any site — point it at your URL and review your first draft today.
Plant your first post