The Programmatic SEO Starter Mini-Guide

Programmatic SEO means creating a large number of pages automatically from a template and a structured dataset, rather than writing every page from scratch. Think of a site like Zapier, where every page follows the format 'Connect [App A] to [App B]'. There are thousands of those pages. Nobody wrote them one by one. They built a template, fed it a list of app pairs, and let the system generate the pages. The reason this works is that people search for very specific things. Someone does not search 'CRM integration'. They search 'how to connect HubSpot to Google Sheets'. If you have a page that answers that exact query, you rank for it. Multiply that across hundreds or thousands of specific queries and the traffic adds up fast. What it is not: a way to flood the internet with thin, useless, copy-pasted pages. That used to work in 2013. Google has had years to get better at detecting it, and the penalty when it catches you is severe. The entire site can be de-indexed. Programmatic SEO only works when the pages being generated are specifically useful to the person who lands on them.

Before you touch a template, you need a dataset. This is the list of variables that will make each page different from the others. The quality of your dataset is the quality of your programme. If the dataset is shallow, the pages will be shallow. Good datasets have two qualities. First, they are specific: each entry is a distinct, nameable thing, not a vague category. A list of 300 UK towns is a good dataset. A list of 'various industries' is not. Second, they carry enough real information per entry that a page about that entry can say something meaningful. A list of UK towns paired with real statistics (average broadband speed, number of registered small businesses, median marketing spend for SMEs) gives you something to work with. A bare list of town names gives you nothing. Where do you find datasets? Public sources are the cleanest starting point. Government open data portals, Companies House, ONS, industry associations, and export-ready data from tools you already use (Google Search Console, your CRM, LinkedIn Sales Navigator exports) are all fair game. You can also build your own dataset by scraping publicly available structured data, by surveying your audience, or by aggregating published reports into a spreadsheet. The dataset does not need to be massive to start. One hundred well-chosen entries is enough to test whether the approach works before you invest in scaling.

A template only works if it maps to a real search pattern. The search pattern is the sentence structure people type, with one or two variables that change across queries. 'Best AI tools for [industry]' is a pattern. 'HubSpot vs [competitor] for [company size]' is a pattern. 'How to do [task] in [city]' is a pattern. To find yours, start in Google Search Console if you have traffic already. Look for queries where you rank between position 5 and 30, that follow a clear pattern, and where the search volume per variation is modest but the number of variations is large. That sweet spot, low competition per variation but many variations, is where programmatic SEO wins. If you are starting from scratch, use a keyword research tool (Ahrefs, Semrush, or even free tools like Keyword Surfer) and search for your core topic. Look for the 'also searched for' and 'questions' sections. When you notice the same question repeating with a different variable swapped in, that is your template hiding in plain sight. One critical check: make sure the intent behind each variation is informational or commercial, not navigational. Navigational searches (where someone is looking for a specific brand or site) will never lead them to your page regardless of how good it is. You want queries where the person is researching or comparing, and where no single dominant answer exists yet.

The template is the frame that stays the same across every page. The variables are what change. The rule is: the variable data must do enough work that each page feels written for that specific entry, not like a generic page with a name swapped in. A weak template looks like this: '[City] is a great place to do business. Here are some AI tools that can help businesses in [City].' Then a generic list of tools follows. The city name appears twice but does nothing. The page is identical regardless of which city you pick. Google sees through this immediately. A strong template looks like this: '[City] has [X] registered SMEs, with [top industries] making up the majority of the business base. The most common operational bottleneck in this type of market is [specific challenge]. Here are the three AI tools that address that bottleneck most directly, with notes on which business sizes they suit.' Now the city data is doing real work. The page about Bristol says something different from the page about Edinburgh, because the data behind it is different. When you write your template, mark every sentence as either 'static' (same across all pages) or 'dynamic' (pulled from the dataset). Aim for at least 40 percent of the meaningful content to be dynamic. If you cannot hit that, your dataset is not rich enough yet. Go back and add more fields before you generate the pages.

Google's Helpful Content system, updated repeatedly since 2022, is specifically designed to catch scaled content that adds no real value. The signals it looks for include: pages that are nearly identical apart from one swapped variable, pages that do not answer the query they appear to target, pages with no original analysis or perspective, and pages that exist purely for the search engine rather than for the reader. Before you publish a batch of programmatic pages, pull five random samples from different parts of your dataset. Read each one as if you are the person who searched that query. Ask: does this page answer the question I had? Is there anything here I could not find by Googling for ten seconds? Does it tell me something I would not know just from reading the page title? If the answer to any of those is no, the template needs more work. Also check for variation at the top of the page. The opening paragraph is where Google's crawler makes its first judgement. If every page in your batch opens with the same sentence structure, with only the variable swapped, that is a flag. Write two or three different opening structures and rotate them. Even a small difference in how the introduction is framed signals to the crawler that these are not machine-generated duplicates. Finally, set your pages to noindex while you test. Publish a batch of fifty, monitor them for four to six weeks in Search Console, and check for crawl anomalies before you scale to hundreds. A noindexed page that proves itself can be switched to index. A spammy page that gets indexed at scale can take months to recover from.

A common mistake is generating hundreds of pages and then not linking to them from anywhere. Orphaned pages do not get crawled reliably, do not inherit authority from your main site, and tend to rank poorly even when the content is good. Structure matters as much as the content itself. The cleanest structure for a programmatic programme is a hub-and-spoke model. Your main pillar page (the hub) covers the broad topic and links out to the individual programmatic pages (the spokes). Each spoke links back to the hub. If your pages cluster naturally into groups (by region, by industry, by use case) add a middle tier: a category index page that collects all the spokes in one cluster and sits between the hub and the individual pages. For internal links to work, the anchor text needs to be descriptive and varied. Do not link every spoke page with the text 'click here'. Use the specific variable as the anchor: 'AI tools for accountancy firms', 'AI tools for construction SMEs', and so on. This signals to Google what each page is about and helps distribute relevance across the structure. Keep the URL structure simple and consistent. A format like /ai-tools-for/[industry-slug]/ or /[city-slug]/ai-marketing-guide/ is readable, crawlable, and scales cleanly. Avoid deep nesting. Pages more than three clicks from the homepage rarely get crawled regularly on smaller sites.

AI writing tools can accelerate the production of programmatic content significantly, but they work best as a formatting and variation layer, not as the source of the underlying data or analysis. The mistake most people make is asking the AI to generate both the data and the content. The result is pages that sound plausible but contain nothing verifiable, nothing specific, nothing a reader can act on. That is the exact failure mode Google has trained itself to detect. The right split is: you supply the structured dataset with real data points, and you use AI to help draft the narrative sections that connect and contextualise those data points. For example, if your dataset includes the number of AI tool adopters per industry sector, the average ROI reported, and the top three use cases, you give those figures to the AI and ask it to write a two-paragraph summary for each industry. The AI is handling the prose. The intelligence is the data you assembled. For quality control at scale, build a review prompt into your workflow. After the AI drafts each page section, run a second AI pass with a prompt that checks for: banned phrases, sentences that make claims without sourcing them to the data, and any passage that reads identically to another page in the batch. Flag and fix before publishing. If you are using a tool like Airtable or Notion as your dataset, you can connect it to a generation pipeline using Make or Zapier. The flow is: new dataset row triggers template fill, AI fills the narrative sections, output is pushed to a staging environment for your review, then you publish approved pages in batches. This keeps the human review step in the loop without slowing the scale.