Programmatic SEO with AI: How to Scale Pages That Rank

The most complete public snapshot of programmatic scale we found this month is Ahrefs’ guide, updated 2 September 2026 and fetched 8 September 2026 (ahrefs.com/blog/programmatic-seo). Its traffic figures are Ahrefs’ estimates of monthly organic visits from its own index — not Google Analytics numbers — so treat them as directional. What they show is the spread between modules that earned their rankings and modules that merely exist.

The winners share one property: every page answers a distinct, useful question. Canva’s feature directory has roughly 310 pages currently ranking and an estimated 13 million monthly organic visits, up from about 20,000 monthly visits as the directory grew. Notion’s template category pages number about 600 ranking URLs and estimate roughly 204,000 monthly organic visits. QuillBot’s AI-writing-tools library, around 130 pages, estimates about 464,000. Smaller examples behave the same way: NutriScan’s calorie and nutrition directory runs about 4,000 pages for an estimated 182,000 monthly visits, and German news site Ad Hoc News publishes stock-news pages at scale under /boerse/news/ — roughly 72,000 pages and an estimated 1.74 million monthly visits.

The same guide documents the failure side. It quotes SEO consultant Lily Ray as tracking over 70 companies caught in Google’s early-2026 updates after scaling content, and Lars Lofgren as spending months publicly urging teams to stop after watching rankings collapse. It also quotes an opposite data point: Zak Perez’s client scaled AI content across 10,500 pages and saw consistent ranking gains across years and multiple Google updates. The four sites and the two warnings describe the same rule from both sides: pages that give a searcher something specific and verifiable scale; pages that assemble templates around invented or recycled claims get caught once the pattern is broad enough to look like automation for its own sake.

The pipeline below assumes one operator, a spreadsheet, and access to a large language model like Claude or ChatGPT. Steps one to three decide quality and should take most of your planning time; steps four to six are production. We built this sequence after studying how the case studies above are put together, and it is deliberately boring: the leverage is in the columns, not the prompts.

One rule governs the whole pipeline: the spreadsheet is the only source of truth. Every factual claim on a page must trace back to a column, and columns you cannot source must not exist. If the model writes a fact from memory instead of from the row, you have built a machine for manufacturing unverifiable claims at scale — the exact failure mode the enforcement line in section five targets.

Scenario (illustrative example): Aster HR, a fictional HR software vendor, wants to rank for “{competitor} vs Aster HR” queries against the 400 tools its sales team most often replaces. The editorial team cannot write 400 bespoke comparisons, and a purely AI-generated set without a data layer would be 400 pages of the same three paragraphs. The module works because of the spreadsheet.

The data layer holds one row per competitor with eight columns: company, category, headline positioning (sourced from the competitor’s own homepage), three feature differences that matter to Aster’s ideal customer (each from Aster’s published feature comparison), pricing model with a source URL, the switch trigger the sales team hears most often (paraphrased from call notes, marked [INTERNAL]), and review status. The template renders each row as a two-line summary of the competitor, a five-row feature table that differs for every competitor, the pricing-model line with its source link, and three FAQ answers written only from the row.

The generation prompt fixes the boundaries: “Write the page for {competitor} using only the row JSON. Do not research the competitor yourself. Do not add features, prices, or reviews that are not in the row. If the row lacks a fact the template expects, output [VERIFY: missing {fact}].” The flags are the point: when a competitor row is thin, the page gets flagged instead of hallucinated. Aster’s editor reviews the first cohort of 20 pages, fixes the three columns that caused the most flags, and only then generates the remaining rows in batches of 50. No page goes live with an unresolved flag — a fabricated competitor feature is the one mistake that turns a comparison page into a liability.

The honesty constraint applies to this example too: Aster HR is fictional, and we are not reporting any traffic outcome for it. What transfers to a real project is the structure — sourced columns, a template that renders differences, generation that refuses to invent, and a review gate that blocks flagged pages. Run that structure and the pages can honestly claim to be useful; skip any of the four parts and the scale starts working against you.

Google’s spam policies define scaled content abuse as producing content at scale — automated or not — primarily to manipulate search rankings rather than to help users (developers.google.com/search/docs/essentials/spam-policies). Its separate AI-content guidance judges pages on helpfulness, not on how the text was produced (developers.google.com/search/docs/fundamentals/ai-content). Honesty note: Google properties were not reachable from our network on 8 September 2026, so both links point to Google’s published guidance rather than a same-day re-fetch.

Read the policy wording closely and the practical line is clear: the target is the intent of the operation, not the template. A 4,000-page nutrition directory built from a real food database answers a question per page. A content dump where a model rewrote the same generic advice under different entity names answers nothing. AI corrupts programmatic SEO in three specific ways: it invents the facts that make a page specific (hallucinated features, prices, reviews), it paraphrases claims until they drift from what the source said, and it lets teams skip the human pass that would have caught both. All three scale perfectly — which is why they are dangerous at 1,000 pages and merely annoying at ten.

Three audit questions keep a module on the right side of the line. First, could a reader learn something from this page that the template alone would not give them — a specific number, a real difference, a sourced fact? Second, can you defend every factual claim from the data layer, with a source a customer could click? Third, if a prospect who just read the page called your sales team, would the call survive the page’s claims? Modules that pass all three can scale into the tens of thousands without becoming the cautionary tale; modules that fail any of them should stay a prototype.

Q: Is programmatic SEO with AI still viable after Google’s 2026 updates? A: The 2026 enforcement wave hit sites that scaled content without scaling value — Lily Ray told Ahrefs she has tracked over 70 companies caught in early-2026 updates after scaling content. Sites whose pages each answer a real query with sourced data kept ranking through those same updates. Viability is decided by the page, not the technique: a module where every page contains a verifiable difference is viable; a module that mass-produces similar text around different keywords is the definition of the risk.

Q: How many AI-generated programmatic pages should I publish at once? A: Publish in cohorts you can fully review — tens of pages per batch, not thousands. Review the first cohort line by line, fix the template and the data columns that produced flags, then grow batch size as review confidence rises. Launching 10,000 pages on day one is how teams meet the failure pattern for the first time in a Google update rather than in their own review queue.

Q: Do I need to code to build programmatic pages with AI? A: No. The pipeline runs on a spreadsheet, an LLM, and the CMS you already use — most CMS platforms import content from a CSV, and the model outputs copy that a human pastes into the template. Code becomes useful when you want true automation: scripts that render pages straight from the spreadsheet or generate the HTML directly. Start by hand and automate only after the template has proven itself on reviewed cohorts.

Q: Which AI tools do teams use for programmatic SEO pages? A: The common stack is an LLM for drafting — Claude (Pro is $20 per month billed monthly or $17 per month billed annually, per anthropic.com/pricing, fetched 8 September 2026) or ChatGPT — plus a research and rank-tracking tool such as Ahrefs (Lite $129 per month, Standard $249, Advanced $449, per ahrefs.com/pricing, fetched the same day) or Semrush. Our Semrush vs Ahrefs vs Surfer comparison covers the research-tool choice in detail.

Q: How do I keep AI-written programmatic pages from looking duplicated? A: Duplication is a data problem, not a writing problem. If two pages read alike, their rows are alike: the columns do not hold enough real difference to support distinct pages. Add differentiating data — sourced specifications, real screenshots, module-specific FAQs, original numbers — until the template produces visibly different pages. If you cannot find enough real difference across the entities in your pattern, the pattern is too broad and no prompt will fix it.

Three actions, in order. First, audit your product or service for a repeatable query pattern and check the SERP against the three qualifications in Step 1: shared intent, templated competitors already ranking, enough variants. Second, build the data layer for ten rows only, draft the row-to-page template, generate a ten-page prototype with the Step 4 prompt, and review those pages as if a competitor would read them. Third, decide the kill criterion and the cohort size before publishing anything — and if the ten-page review makes you uncomfortable, shrink the plan instead of skipping the gate. A 40-page module that earns rankings beats a 4,000-page module that earns an enforcement action, and the pipeline above works identically at both sizes.