Blog · Content Systems

Programmatic SEO: how to build page families without thin content

Generating pages from a database is not spam. It becomes spam at the exact point where the template stops carrying information the reader could not have assembled from a search box, and that point can be designed, measured and enforced before anything is published.

What is programmatic SEO?

Programmatic SEO is the practice of generating many search-targeted pages from one template and a structured dataset, where each page answers a variation of the same query pattern, such as “[tool] integration with [app]” or “[metric] benchmark for [industry]”. It works when each page carries real information from the data, and fails when only a substituted noun differs.

The mechanics are simple. A query pattern has a fixed head and a variable modifier. A dataset holds one row per modifier and one column per fact the page needs. A template turns each row into a page. The difficult parts are everything around that: whether the data deserves a page per row, whether search engines want that page type for the query, and what happens to the rows with thin data.

Common programmatic page families
Family Query pattern Where the value has to come from
Integration pages [product] + [app] integration What the integration actually does, triggers, actions, setup steps
Template and example galleries [document type] template The template itself, usable on the page
Comparison pages [product] vs [competitor] Verified, current feature and pricing differences
Location pages [service] in [city] Genuinely local facts: prices, availability, providers, rules
Data and benchmark pages [metric] for [segment] The numbers, with the method and sample stated
Calculator or tool pages [calculation] calculator The computation, run for the reader
Glossary pages what is [term] A definition worth reading, which is hard to template
The third column is the whole question. If it is empty for a family, the family is not ready

Several of these are familiar from large sites. Shopify’s guide to the subject names Zapier’s app integration pages and Tripadvisor’s localised destination pages as examples, and Ahrefs’ guide includes Notion’s template category directory and Canva’s feature pages. The pattern in each is the same: the page is useful because the row behind it holds something specific.

Programmatic SEO versus a content cluster

A content cluster is a set of hand-written pages on related subtopics, linked to a pillar, each answering a different question. A programmatic family answers the same question many times for different modifiers. The distinction matters for planning: a cluster is limited by writing capacity, a family is limited by data quality. Most sites that get programmatic SEO wrong treat it as a cheaper way to write a cluster, when it is a way to publish a dataset.

Is programmatic SEO spam? What Google’s policies say

Programmatic SEO is not spam in itself, but Google’s spam policies apply to generated pages exactly as they apply to written ones, and the deciding test is whether each page adds value for users. The policy that matters most is scaled content abuse, and it is explicitly indifferent to how the pages were produced.

The policy’s listed examples include using generative AI or similar tools to generate many pages without adding value, scraping feeds or search results to generate pages, stitching content from different pages without adding value, and creating many pages that make little sense to a reader but contain search keywords.

The same document defines doorway abuse as pages created to rank for specific, similar queries that lead users to intermediate pages less useful than the final destination. One of its examples is pages targeted at specific regions or cities that funnel users to one page, which is a description of the weakest kind of location family.

Google widened the scaled content policy in March 2024. Its announcement, New ways we are tackling spammy, low-quality content on Search, said the policy now covers producing content at scale to boost rankings whether automation, humans or a combination are involved, and an update to the post said the accompanying changes reduced low-quality, unoriginal content in results by 45 percent, against an expected 40.

What changed in March 2024
45%
less low-quality, unoriginal content in results, by Google’s account
Any method
automation, humans or both are covered by the scaled content policy
Google, March 2024 Search update announcement, updated 26 April 2024

Practitioners are wary for good reason. Ahrefs’ guide quotes Google’s John Mueller describing programmatic SEO as often a fancy banner for spam. The useful reading of that remark is the word often: most of what gets published under the label fails the value test, and the label itself is not the problem.

The test to apply to every page

Could a reasonably motivated reader have produced this page themselves in under a minute with a search box? If yes, the page is a wrapper around a query and adds nothing to the index. If no, because the page contains a comparison, a calculation, a normalisation across sources or genuine editorial judgement, it is a real page that happens to have been generated. A second check takes ten seconds: swap the modifier on a finished page for a neighbouring one, a different city or a different app, and read the page again. If every sentence is still true, the page carries no information specific to its row. Google’s helpful content guidance asks the same thing in other words: its self-assessment asks whether content provides original information, reporting, research or analysis, and its list of search engine-first warning signs asks whether you are using extensive automation to produce content on many topics.

Where the information on a programmatic page comes from

Every defensible programmatic page draws its value from a source beyond the template, and there are four: proprietary data, computation, normalisation of messy public sources, and an editorial layer. If a proposed family rests on none of them, it does not yet have a reason to exist.

Four sources of information, and how each scales
Source Example Defensibility Scales
Proprietary data Usage figures, prices or test results only you collected Highest Poorly, and expensive to collect
Computation A calculator or model run over public inputs High Well
Normalisation Several inconsistent public sources reconciled into one format Medium Well
Editorial layer A generated skeleton with human-written judgement on top Highest Worst
Name the source for a family in one sentence before any build work starts
  • Proprietary data is the strongest position because nobody can copy the page without the data. Its cost is collection and upkeep.
  • Computation means the page does arithmetic the reader would otherwise do in a spreadsheet. The inputs can be public; the output is not sitting anywhere else.
  • Normalisation is unglamorous and real. Pricing pages from twenty vendors use twenty formats; one comparable table is work a reader would not do.
  • An editorial layer adds a judgement per row: which option suits which buyer, what the data does not show. It scales worst and defends best.

When programmatic SEO is the right choice, and when it is not

Programmatic SEO is the right choice when a query pattern repeats across many modifiers, the results for that pattern reward a structured page rather than an essay, and you hold data that makes each row worth a page. If any of the three is missing, hand-written pages or a single comparison page will usually do better.

Programmatic SEO go or no-go
Does the query pattern repeat across more than a few dozen modifiers with real demand?
No
No-go. Write the pages by hand. The build overhead is not worth it for a short list.
Yes
Do the top results for ten sample rows, head and tail, reward a structured page?
No, they are long guides
No-go. A template page is the wrong format. Write the best single guide, or one comparison page, instead.
Yes, tools, directories or comparisons
Can you name the information source in one sentence?
No
Stop. The family has no reason to exist yet. Find the data, the computation or the editorial layer first.
Yes
Go. Define the coverage threshold, build a test batch of ten to twenty pages from the thinnest rows, and read every one.
The tree covers the three conditions. The rules below apply once a family is a go.
Three more rules once a family is a go
If Then
A modifier shows no measurable demand in your keyword tool Do not generate it. List it on the hub page instead
Fewer than half the rows clear the coverage threshold Publish the rows that clear it, or fix the data first
Two modifiers produce the same results page They are one page, not two. Merge the rows
Each one is cheaper than the build it prevents

The second question in the tree is the one most often skipped. Search ten sample rows by hand, including some from the tail of the list, and look at what ranks. If every result is a 2,000-word guide from an established publisher, a 300-word template page will not compete however clean the data is. If the results are directories, marketplaces, tool pages or structured comparisons, the format is right.

The last rule in the table prevents a family from cannibalising itself. Modifiers that look different in a spreadsheet often mean the same thing to a search engine: “for startups” and “for small businesses”, singular and plural forms, an app’s old and new names. The diagnosis and fixes for that are covered in the guide to finding and fixing keyword cannibalization, and the cheapest place to apply them is the source list, before generation.

How to do programmatic SEO, step by step

Programmatic SEO done properly is a sequence of ten steps, from choosing the query pattern to pruning the family after launch. The order matters because the expensive steps, building and publishing, come after the cheap checks that decide whether they are worth doing.

  1. Choose the query pattern. Find a head term with a repeating modifier in your keyword research tool, such as Ahrefs, Semrush or Google Keyword Planner, and export every modifier with its volume.
  2. Validate the page type. Search ten sample rows by hand, head and tail, and confirm the results reward a structured page.
  3. Name the information source. One sentence: which of the four sources each page rests on.
  4. Build the dataset and field inventory. One row per modifier, one column per fact. Record where each column comes from and how often it changes.
  5. Define the coverage threshold. List the fields a page must have to be worth indexing. Rows that fail are held back.
  6. Design the template. Separate fixed copy, variable fields, computed blocks and the editorial block, and design for the hardest row, not the richest.
  7. Plan the internal links. Hub pages by category, links between related rows, and a link from the family to the commercial pages it supports.
  8. Build a test batch of ten to twenty pages, chosen deliberately from thin, odd and messy rows, and read every one as a reader would.
  9. Publish in waves, each with its own XML sitemap, and watch indexing for the first wave before releasing the next.
  10. Monitor and prune. Review the family monthly: indexing, impressions per page, a sample of the tail, and data freshness.

Tools for each stage

A typical tool stack, by stage
Stage Common tools What to check
Keyword pattern Ahrefs, Semrush, Google Keyword Planner Modifier list, volumes, sample results pages
Dataset Google Sheets, Airtable, a database Field completeness per row, source per column
Page generation Webflow CMS, WordPress with an importer, Next.js or another static site generator Template output for the thinnest rows
Quality review Screaming Frog or Sitebulb crawl Duplicate titles, thin word counts, missing fields, orphan pages
Indexing Google Search Console, one sitemap per family Submitted against indexed, by family
Structured data A schema validator Markup matches what the page shows
Any equivalent tool works. The checks in the third column are the point

Step six, designing for the hardest row, is the one that separates good families from bad ones. A template designed around the richest row looks excellent in the mock-up and turns into a mostly empty shell at row four thousand. Design around a row with the minimum fields that clear the threshold, and the template holds up everywhere.

Programmatic SEO launch checklist

The checklist turns the ten steps and the controls in the sections below into the checks to tick before each wave goes live. Work through it for wave one, then again for every later wave, because a family that passed in March can fail in September when its data changes.

Programmatic SEO launch checklist
Before the build
Before wave one
After launch
Ticks are saved in your browser. Copy as Markdown to share with the build team.

How to set a coverage threshold for generated pages

A coverage threshold is the minimum set of data fields a generated page must have before it is published and indexed. It is defined once per family, enforced at build time rather than discovered at audit time, and it is the single control that most reliably keeps a programmatic family out of thin-content territory.

The threshold is a list of required fields plus a rule for what happens to rows that fail. Pages that clear it publish. Pages that do not are held back: not published with gaps, and not published with a placeholder apologising for the gaps. Holding a row is cheap and reversible. Publishing a thin page and then removing it costs a crawl, a removal and some trust in the rest of the family.

A worked example

Take a hypothetical SaaS company planning integration pages, one per app its product connects to, with 1,200 apps in the dataset. The required fields are: what the integration does, written and verified; at least three supported triggers or actions; setup steps; a screenshot; and a last-verified date.

Hypothetical: 1,200 integration rows against the threshold
Row status Rows Share Action
All required fields present 310 25.8% Publish in wave one
Everything except setup steps 540 45.0% Write setup steps, then publish
Name and category only 350 29.2% Hold. Do not publish
Total 1,200 100%
Hypothetical figures for illustration

The arithmetic decides the plan. Wave one is 310 pages, 25.8 percent of the list. The 540 rows missing only setup steps are the real opportunity: at a hypothetical 15 minutes per row, writing them takes 540 times 15, or 8,100 minutes, which is 135 hours of work for a 174 percent increase on wave one. The 350 rows with a name and a category are exactly the pages Google’s scaled content policy describes, and they stay unpublished until there is something to say.

Without the threshold, this family launches as 1,200 pages, of which 350 are empty shells and 540 are incomplete. With it, it launches as 310 good pages and grows as the data does.

Coverage threshold calculator

The calculator runs the same arithmetic on your own dataset. Wave one is the rows that clear the threshold, divided by all rows, capped at 100 percent. Hours to fix the near misses is near-miss rows multiplied by minutes per row, divided by 60. The growth from fixing them is near-miss rows divided by complete rows. Held rows are whatever is left: all rows minus complete rows minus near misses.

Coverage threshold calculator

Count your rows against the threshold you defined, then estimate the fix time.

Wave one, share of the dataset25.8%
Hours to fix the near misses135.0
Growth on wave one from the fixes174%Shows 0 when no rows clear the threshold
Rows to hold back350Not published until there is something to say
Held rows, share of the dataset29.2%
Hypothetical defaults from the integration example above. Change any number.

With the hypothetical defaults, wave one is 25.8 percent of the dataset, the near misses take 135.0 hours, fixing them grows the family by 174 percent, and 350 rows, 29.2 percent, stay held. My working rule, a judgement rather than a measured threshold: if the held share is above half, treat the dataset as the project, not the template.

How to design a programmatic template that is not thin

A programmatic template avoids thin content when most of each page’s words come from fields that genuinely differ by row, and the fixed copy is limited to what a reader needs to use those fields. The design question for every block is what varies, and whether the variation means anything.

Template blocks, and whether they make a page unique
Block Varies by row Counts as unique value
Title and H1 Modifier only No, but must be accurate and distinct
Opening answer Computed from the row’s data Yes, if it states the row’s specific facts
Data table or comparison Every cell Yes. Usually the core of the page
Computed result or chart The calculation Yes
Editorial note Written per row or per segment Yes, and the strongest
Method and sources Little No, but it is what makes the page citable
Related rows and hub links The neighbours No, but it is how the family is crawled
Generic explainer copy Nothing No. Keep it short or move it to the hub
If the last row is the longest block on the page, the template is upside down

Three design rules follow from the table. First, write the opening as a statement of the row’s facts, not as a sentence with a slot: “The integration syncs contacts and deals both ways, with three triggers” beats “Connect X with our product to boost productivity”. Second, put the shared explainer copy on the hub page rather than repeating it on every row, so each row page is mostly the row. Third, when a field is empty, drop the block entirely. A heading followed by “No data available” is the most recognisable sign of a thin family.

Variation for its own sake does not help. Spinning the same sentence five ways across rows produces pages that differ in wording and not in information, which is the obfuscation that Google’s policy names when it lists synonymising among scraping techniques.

Crawling, indexing and site architecture for large page families

Large programmatic families need deliberate crawl and index management: a hub structure that reaches any page in a few clicks, one XML sitemap per family, accurate last-modified dates, and controls that stop filter and parameter URLs multiplying. Without them, the family competes with the rest of the site for crawling and its indexing problems stay invisible.

Crawl budget is a real constraint only at scale, and Google gives rough figures for where scale starts. Its crawl budget guide is written for sites with roughly a million or more unique pages changing weekly, or ten thousand or more changing daily, and for sites with many URLs in the Discovered, currently not indexed status. Below that, indexing problems on a family are usually about quality and linking, not crawl capacity.

Sitemaps and monitoring

Google’s sitemap documentation limits a single sitemap to 50,000 URLs or 50MB uncompressed, allows a sitemap index file, and says Google uses the lastmod value when it is consistently and verifiably accurate. Two practices follow. Give each family its own sitemap, so Search Console’s page indexing report can be read per family. And set lastmod from the date the row’s data actually changed, not the date the page was regenerated, or the signal becomes noise.

Search Console’s page indexing report then becomes the family’s health check. Crawled, currently not indexed means Google crawled the page and did not index it, and it may or may not be indexed later. Discovered, currently not indexed means Google found the URL and has not crawled it yet. A family where a growing share of pages sits in either status, while the rest of the site indexes normally, is telling you something about the template or the linking.

Reading the page indexing report for one family
What the family sitemap shows Likely cause First thing to check
Most submitted pages indexed The template and linking are working Impression coverage, the next ratio
A growing share in Crawled, currently not indexed Google read the pages and passed Read ten of those pages; compare them with indexed members
A growing share in Discovered, currently not indexed Found but not yet crawled Click depth from the hubs, and whether members are linked from anywhere but the sitemap
The family indexes far worse than the rest of the site A family-level problem, not a site-level one The coverage threshold: rerun it against today’s data
Status definitions from Google’s page indexing report help; causes and checks are practice

Google’s link guidance says that every page you care about should have a link from at least one other page on the site. For a family, that means category hubs that link to every member, members that link to their nearest neighbours, and no reliance on paginated index pages a crawler has to walk through a hundred deep. The general principles are in the piece on treating the internal link graph as a product decision, and the Internal Link Mapper shows click depth and which pages receive only navigation links, which is how orphaned rows in a family are usually found.

Filters and parameters

Families built on a filterable directory create their own crawl problem. Google’s faceted navigation guidance warns that crawlers will access a very large number of filter URLs and have less time for new, useful ones, and recommends blocking filter parameters in robots.txt when those URLs do not need to appear in Search. It describes rel=canonical on filtered URLs as generally less effective in the long term. Decide which filter combinations deserve to be pages, give those clean URLs, and block the rest.

Structured data and AI search for programmatic pages

Structured data on programmatic pages must describe only what each page actually shows, and the same discipline that makes a generated page useful to a reader makes it usable by AI search: specific facts, stated plainly, in sections that make sense on their own.

Templates make structured data easy to add and easy to get wrong at scale. Google’s structured data policies say markup must be a true representation of the page content and must not mark up content that is not visible to readers.

Generate the markup from the same fields as the visible content, omit it when the field is empty, and check a sample with the Schema Validator before each wave.

For AI assistants and answer engines, a generated page is retrieved and quoted in passages, not as a whole. A row page whose opening states the row’s specific facts, and whose table carries labelled values, gives a retrieval system something precise to quote. A page whose distinctive information sits below 400 words of generic explainer gives it nothing. The underlying mechanism is set out in what AI search actually retrieves from a page.

How to govern a programmatic family after launch

Governing a programmatic family means catching quality decay in the tail before it spreads, through a monthly sample review, data freshness rules and a pruning rule for pages that never earn anything. Families rot quietly because aggregate metrics are dominated by the head, where data is richest.

The first hundred pages usually cover popular entities with complete data. Page four thousand covers an entity with three fields populated. Aggregate traffic looks healthy because the head carries it, while the tail grows into a large set of pages that would fail the threshold if it were rerun today.

Most of that governance can run as automated checks: field coverage per page, data age, and pages with no impressions after a set period, each with an alert and an owner. What to automate, and what to keep for a person, is covered in the guide to SEO automation.

  1. Sample the tail monthly. Pull twenty pages at random, weighted towards the tail, and read each as a reader would. Thin pages are invisible in dashboards and obvious within thirty seconds of reading.
  2. Rerun the threshold on every rebuild. Data changes. A row that passed in March can fail in September when a field goes stale.
  3. Prune on a schedule. Each quarter, list pages with no impressions in the last 90 days and decide for each: improve the data, merge into the hub, or remove.
  4. Track freshness per field. Pricing and feature data age fastest. A comparison page with last year’s pricing is worse than no page.
Pruning rules for pages in a family
Page state after 90 days Action
Indexed, impressions, clicks Keep. Refresh data on schedule
Indexed, impressions, no clicks Improve the title and opening. Check the row’s facts are visible early
Indexed, no impressions Check demand for the modifier. Merge into the hub if there is none
Not indexed after repeated crawls Rerun the threshold. Improve the data or remove the page
Data stale beyond the freshness rule Hold for review until the data is refreshed
Return 404 or 410 for removed pages; Google’s crawl budget guide calls this a strong signal not to recrawl

Writing these rules into the build, rather than leaving them in a reviewer’s head, is what content systems work consists of: the decisions are made once, encoded, and applied to every row on every rebuild. A family reviewed as a family, rather than page by page, is also how a content audit should treat generated pages.

How to measure programmatic SEO

Programmatic SEO is measured per family rather than per page, using four ratios: the share of submitted pages that are indexed, the share of indexed pages earning impressions, clicks per indexed page, and conversions per family. Each ratio points at a different fault when it is low.

Four ratios, and what a low value points to
Ratio How to compute it If it is low
Index rate Indexed pages divided by pages submitted in the family sitemap Quality or linking problem in the template
Impression coverage Pages with any impressions in 90 days divided by indexed pages Modifiers without demand, or cannibalisation inside the family
Clicks per indexed page Family clicks divided by indexed pages Titles and openings do not match the query
Conversions per family Sign-ups or leads attributed to the family No path from the family to commercial pages
Compute from Search Console, filtered by the family’s URL folder and sitemap

These figures have no universal benchmark. Their value is in comparison: against the next wave, against other families on the same site, and against the same family next quarter.

Read the ratios in order. A low index rate makes the others meaningless, so fix it first. The same habit, judging a change by whether the fault stops recurring rather than by one good month, is how the SEO audit checklist measures whether an audit worked, and it applies to every rebuild of a family.

A first week, month and quarter

  • Week 1. Choose the pattern, validate the page type on ten sample rows, name the information source, and build the field inventory. Most families should be stopped or reshaped here.
  • Month 1. Define the coverage threshold, design the template for the hardest row, build and read a test batch of ten to twenty pages, and publish wave one with its own sitemap.
  • Quarter 1. Watch the four ratios for wave one, fix the template before releasing wave two, run the first tail sample and the first pruning pass, and decide whether the family has earned a second wave.

A free tool is often a better first programmatic asset than a page family, because it puts computation, the second information source, on one page instead of spreading thin data across many. The case for that, and how to choose which tool, is in choosing which free tool to build first.

The short version

Programmatic SEO is safe when every page carries information the reader could not assemble themselves, and Google’s scaled content policy applies however the pages were made. Name the information source before you build, enforce a coverage threshold at build time so thin rows are held back, give the family a hub structure and its own sitemap, and sample the tail every month. Families fail on thinness in the tail, not on volume at the head.

Operator note

I ask one question before any programmatic build: which of the four information sources is this resting on? If the answer takes more than a sentence, the family is not ready, and building it anyway means discovering that at page four thousand instead of at the whiteboard.

The coverage threshold is the control that saves these projects, and it is the one most often cut for launch. Hold pages back rather than shipping them with gaps. A family that launches small and grows with its data is easier to defend than one that launches large and has to be pruned.

Frequently asked

What is programmatic SEO?
Programmatic SEO is generating many search-targeted pages from one template and a structured dataset, with each page answering a variation of the same query pattern, such as an integration page per app or a benchmark page per industry. It works when each page carries specific information from the data, and fails when pages differ only by a substituted noun.
Is programmatic SEO against Google’s guidelines?
Generating pages from data is not itself a violation. Google’s scaled content abuse policy covers many pages generated mainly to manipulate rankings without helping users, and it applies no matter how the content is created, including by automation, humans or both. The deciding question is whether each page adds value a reader could not get from a quick search.
Does programmatic SEO still work?
It works for families where each page holds data, a computation or a judgement the reader cannot easily assemble, and where the results for the query reward a structured page. It works poorly for templated pages that swap a city or product name into otherwise identical copy, which is the pattern Google’s spam policies describe as scaled content and doorway abuse.
How do you avoid thin content with programmatic SEO?
Set a coverage threshold of required fields and hold back every row that fails it, design the template on the thinnest row that passes, and drop any block whose field is empty. Then sample the tail of the family every month, because thinness appears there first. Page count is the wrong unit: a thousand pages each carrying a real calculation is fine, and two hundred that differ only by a noun are not.
What is a coverage threshold in programmatic SEO?
A coverage threshold is the minimum set of data fields a generated page must have before it is published and indexed. It is defined once per family and enforced at build time. Rows that clear it publish, and rows that do not are held back rather than shipped with visible gaps or placeholder text.
Do programmatic pages hurt crawl budget?
On most sites they do not, because Google’s crawl budget guide is written for sites of roughly a million or more pages, or ten thousand or more that change daily. Where a family is large enough to matter, give it a hub structure, its own sitemap and accurate lastmod dates, and block filter URLs that do not need to be in Search.
Can I use AI to write programmatic SEO pages?
Google’s guidance says using generative AI to create many pages without adding value for users may violate its scaled content abuse policy, and that AI-generated content and metadata should be fact-checked before publishing. AI can help draft an editorial layer over real data. It cannot supply the data, and a generated paragraph restating an entity’s name is not information.
How do I measure programmatic SEO results?
Measure each family on four ratios from Search Console: indexed pages against submitted pages, pages with impressions against indexed pages, clicks per indexed page, and conversions per family. Compare waves and families against each other rather than against a universal benchmark, and fix a low index rate before reading anything else.
Related reading
Keyword cannibalization: how to find it, fix it, and stop causing it Internal linking strategy: your link graph is a product decision AI search optimization: what actually gets retrieved from your page
Working on this yourself?

I have 3 client slots open for Q4 2026. If this raised a question about your own situation, or you want help with it, send it over.

Say hello