AI can write ten thousand pages before a founder has learned whether one page deserves to exist.

The cheap part is now multiplication. The expensive part is still truth.

Programmatic SEO works when a template gives many people a useful answer from distinct, reliable evidence. It fails when a keyword and a city name are poured into the same empty mould.

Earn the template with one page. Earn the scale with twenty.

Define the repeated user question

Do not begin with a spreadsheet of keywords.

Begin with a question your intended customer asks repeatedly, where the answer changes in a meaningful and knowable way.

Good candidates include:

  • compatibility between a real product and many integrations;
  • documented requirements across jurisdictions or platforms;
  • public benchmarks derived from a stable dataset;
  • directories with verified attributes and useful comparison;
  • calculators whose inputs produce materially different outcomes;
  • reference pages generated from the product’s own evidence;
  • status, release, or support information tied to real objects.

Weak candidates usually replace one noun while leaving the answer unchanged:

  • “Best launch checker in Zurich”;
  • “Best launch checker in Geneva”;
  • “Best launch checker in Basel.”

If geography does not change the product, evidence, availability, or advice, the city is decoration.

Write the repeated question in one sentence. Then explain why each answer must differ.

Make one page by hand

Choose a representative query and produce the best answer you can without a template.

Include:

  • the direct answer;
  • the facts required to trust it;
  • the method or source behind those facts;
  • important exceptions and limits;
  • the next useful action;
  • links to the primary evidence;
  • the date or version when freshness matters.

Give the page to three people who perform the job. Watch what they search for next. If they still need another result to finish the task, the page is incomplete.

This manual page becomes the specification. A template written before the answer is understood will automate the founder’s ignorance.

Google’s current guidance on helpful, reliable content asks whether a page offers original information or analysis, serves an existing audience, and leaves the reader able to achieve their goal. Those questions are more useful than a target word count.

Find the variable that carries value

Every scalable page needs a variable that changes the answer, not merely the title.

For an integration directory, the useful variables might be:

  • authentication method;
  • supported actions;
  • required permissions;
  • sync direction;
  • setup steps;
  • limitations;
  • last successful test;
  • current availability.

For a launch-readiness benchmark, they might be:

  • technology category;
  • observed public checks;
  • sample size;
  • collection date;
  • methodology;
  • distribution of results;
  • limitations in the evidence.

The variable should come from a source you can name and maintain. If the only variable is a keyword phrase, you do not have programmatic content. You have a thesaurus attached to a deployment script.

Create a source-of-truth contract

For every field on the page, record:

Source: Where does the fact come from? Owner: Who or what updates it? Freshness: When does it become unsafe or misleading? Fallback: What appears when the value is missing? Proof: How can a reader verify it? Removal: When should the page stop being indexable?

Do not let the language model invent a missing price, compatibility claim, location, rating, founder biography, or customer result. Missing data should produce an honest omission, a visible “not verified,” or no page at all.

This matters for agent discovery as much as search. An agent can repeat a clean false fact with great confidence. Consistency makes information retrievable; provenance makes it worth retrieving.

Use the SEO for AI search guide to keep names, claims, dates, and canonical sources consistent across the website.

Set a publication gate

A database row should not automatically become an indexable page.

Require enough evidence to answer the user’s question. A page might need:

  • a unique entity with a stable identity;
  • three or more verified attributes;
  • a meaningful answer distinct from sibling pages;
  • a source or observation date;
  • no unresolved contradiction;
  • a working destination or next action;
  • an internal path from a useful hub;
  • a canonical URL that will remain stable.

If the record fails the gate, keep it unpublished, combine it into a stronger page, or expose it only where users already have context. Do not publish “Coming soon” pages for every possible integration and call the count an SEO strategy.

The gate should be code, not a reminder in a launch document.

Do not build doorways

Google’s current spam policies define doorway abuse as substantially similar pages created for similar queries that funnel people toward another destination. The examples include regional pages with slight variations and pages that sit between search results and the actually useful part of a site.

A useful page must do the job it promises.

If “Launch checklist for marketplace apps” contains only a generic paragraph and a button to run the same scan, it is a doorway. If it explains marketplace-specific trust, supply, demand, payment, cancellation, moderation, and cold-start checks with evidence and practical steps, it can stand on its own.

The product may be the next action. It cannot be the missing answer. If the template compares named products, use the honest comparison-page standard to test both sides, source changing claims, and state who should choose each option.

Keep the template smaller than the evidence

A scalable page should have a stable structure. It should not feel as though the stable structure is all there is.

A useful template might contain:

  1. A direct answer unique to the entity or case.
  2. Verified facts with dates or versions.
  3. The method used to obtain them.
  4. A comparison or interpretation based on those facts.
  5. Known gaps and exceptions.
  6. Related entities chosen by a real relationship.
  7. A next action that follows from the answer.

Do not add paragraphs merely to create length. Google explicitly says there is no preferred word count in its people-first content guidance.

Use normal language. State “This integration can read projects but cannot update them” instead of “Unlock seamless project-management synergy with our robust integration ecosystem.”

The template should reveal the evidence, not smother it.

Generate analysis only where the facts support it

AI can help explain structured evidence, but bound the task.

Give the model:

  • the permitted facts;
  • the reader and question;
  • the claims it may and may not make;
  • required uncertainty language;
  • links or identifiers for sources;
  • examples of acceptable and unacceptable inference;
  • a schema for the output;
  • a validation step that rejects unsupported claims.

Do not ask, “Write an SEO page about this company.” Ask, “Using only these verified fields, explain which launch checks are present, which are absent, and which cannot be determined. Cite the observation behind each claim. Do not infer private security or compliance.”

This is the difference between assisted interpretation and synthetic fact production.

Google’s guidance on AI-generated content does not ban AI as a production method. The purpose and result matter. Automation used primarily to manipulate rankings can violate policy; automation used to create useful content still has to meet the same quality standard.

Build hubs that reflect a real hierarchy

Do not leave thousands of pages in a sitemap with no human path to them.

Create browseable hubs based on how users think:

  • category;
  • use case;
  • platform;
  • audience;
  • problem stage;
  • region, when region genuinely changes the answer.

Each hub should explain the category, expose useful distinctions, and link to the strongest child pages. Each child should link back to the hub and sideways only when the relationship helps the reader.

Breadcrumbs should match the visible hierarchy. Canonical URLs should be stable. Pagination should be usable without relying on a crawler to guess hidden state.

YC’s interview on SEO and growth emphasizes cross-linking and writing around topics the intended audience actually searches for. Internal linking works when it expresses subject structure, not when every page links to every other page.

Make each page legible to agents

An AI agent benefits from the same qualities as a hurried human:

  • a descriptive title;
  • a direct answer near the beginning;
  • explicit entity names;
  • stable definitions;
  • visible dates and versions;
  • primary-source links;
  • clear tables or lists for repeated fields;
  • limitations beside the claim;
  • server-rendered HTML;
  • one canonical URL;
  • consistent facts across pricing, docs, policies, and product pages.

Structured data can reinforce what the visible page already says. It cannot turn a weak page into a trustworthy source.

If a field matters enough to put into JSON-LD, it matters enough to show and verify on the page.

Launch twenty pages, not twenty thousand

Choose a sample that contains variation:

  • common and uncommon entities;
  • complete and incomplete records;
  • fresh and older sources;
  • cases with exceptions;
  • mobile and desktop layouts;
  • short and long values.

Review every page manually. Then ask intended users to complete the job.

Check:

  • Does each page answer a distinct question?
  • Are all factual claims supported?
  • Can a missing value create nonsense?
  • Do titles and descriptions describe the visible page?
  • Are canonical URLs and status codes correct?
  • Can users reach the page through a hub?
  • Can they continue to a useful next step?
  • Does the page work without client-side rendering?
  • Does an agent extract the correct entity and claim?
  • Would you publish this page if it were the only one in the set?

Fix the system from the sample. Do not promise that defects will average out at scale.

Measure usefulness below impressions

Track the sequence:

  1. Eligible pages published.
  2. Pages discovered and crawled.
  3. Pages indexed.
  4. Queries and impressions.
  5. Qualified visits.
  6. Completion of the page’s information job.
  7. Useful next actions.
  8. Product activation, return, or revenue where relevant.
  9. Corrections, complaints, and support burden.

Indexing is not success. Traffic is not success. A page that ranks for the wrong query and sends confused visitors into the product can create cost without traction.

Read search terms beside conversion and support evidence. The startup traction guide separates attention from activation, retention, payment, and referral.

Maintain, merge, and remove

Programmatic content creates an operating obligation.

Monitor source failures, stale records, broken destinations, unexpected template output, and sudden changes in page counts. Sample pages after every template or model change.

Give time-sensitive pages an owner and review interval. Preserve a truthful “last verified” date. Do not change dates merely to look fresh; Google’s people-first guidance explicitly warns against changing dates when the content has not substantially changed.

When a page no longer clears the publication gate:

  • refresh the evidence;
  • merge it into a stronger parent;
  • remove it from internal links and sitemaps;
  • return an appropriate redirect, unavailable status, or noindex based on the real state.

Do not keep a misleading page alive because it once received impressions.

Use a founder-sized proof plan

Before scaling, produce these artifacts:

One manual reference page: the answer worth repeating. One source contract: provenance, ownership, freshness, and fallback for every field. One publication gate: the minimum evidence required for an indexable page. One template: direct answer, facts, method, limits, relationships, and next action. Twenty varied pages: reviewed by humans and tested with intended users. One monitoring view: page eligibility, freshness, indexing, usefulness, and errors. One removal path: a safe way to merge or retire weak records.

Only then increase the count.

Scale what is true

Programmatic SEO is not a loophole in the need to understand a customer. It is software applied to a repeated information job.

YC’s discussion of startup content strategy returns to focus, audience, and original data. Those are exactly the inputs a useful scalable page needs.

Find a question people in your market repeatedly ask. Build one answer worth keeping. Identify the evidence that makes each instance different. Refuse to publish when the evidence is missing. Test a small varied set. Measure whether readers finish the job.

Then scale.

Ten thousand thin pages are not an asset. They are ten thousand public claims you must defend.