Programmatic SEO

Programmatic SEO

One person, a keyword file and a loop. The pipeline below produced 28 English articles and 112 pages across four languages on getseam.app, running at 180,000 impressions and 3,400 clicks over three months. The order matters more than any single step: read where the brand already stands, judge the cluster before entering it, write the three shapes that rank, then feed the result back.

The split that reorders everything

58.5% of Google searches now end without a click to anyone, and an AI Overview costs the number-one page 58% of its clicks. There is no rebound coming: Google answers on the page on purpose, because the alternative is losing the user to an assistant. The numbers, the sources and the HubSpot case that settles the traffic-versus-revenue argument: references/measurement.md.

So every target sorts into one of two jobs, on one question: can an assistant fully satisfy this searcher on the results page? Answer it against a real SERP, never from the keyword's shape. That is the AI filter, and it runs in Phases 2 and 3.

• No, so earn the click. Anything the searcher has to do rather than know: tools and calculators, transactional and commercial queries, and everything local. This lane is intact and traditional SEO works in it exactly as it always did.
• Yes, so earn the mention. Informational queries, question phrasings, long multi-clause asks. You will not get the click. You can be the brand named inside the answer, and that is won off your own site more than on it.

Ranking and being cited are two scoreboards, and the gap is widening

Do not plan the mention lane by planning to rank. Traditional SEO is the foundation and is no longer close to sufficient: ranking first gives roughly a 31% chance of appearing in an AI answer, and most AI citations go to pages outside Google's top ten. One controlled split test moved LLM traffic up and Google traffic down from the same single change. Six independent readings of the overlap: references/geo-signals.md.

Two consequences the whole pipeline runs on:

• The unit is the brand, not the URL. A model is deciding which company to name, not ordering pages. Ranking is an ingredient, and a site can hold position one and never be named.
• Citation is probabilistic. Ask the same question five times and you may be named three, so the metric is share of voice across many runs and a rank tracker aimed at prompts measures nothing.

AEO, GEO and LLMO name the same work. Sell it under the label the buyer already budgets for: a client with a separate GEO line item does not hand it to the supplier who says the category does not exist.

Prerequisites

Search Console access (gcloud auth application-default login with the webmasters scope, per the search-console skill). Nothing else: autocomplete is keyless from Google and volume comes from the Keyword Planner, both free. SERPER_API_KEY is optional. Keep one copy in your own env file and export it before the scripts run. Before writing against a paid SEO endpoint, check what it resells: volume and autocomplete are free at the source. Costs, coverage limits and the rule: references/tooling.md.

Phase 0. Harvest the console

Skip only when the site has no Search Console history. Everywhere else it beats keyword research: Google has already told you which queries it associates with the domain.

GSC="python3 <search-console-skill>/scripts/gsc.py"
$GSC orphans <site> --days 90      # queries you rank for with no page targeting them
$GSC cannibals <site> --days 90    # queries your own pages are splitting

One run on getseam.app surfaced a query split across three of its own pages at 20,448 impressions and position 9.1, plus five variants with no exact page. It pays for itself before a tool is opened.

Run cannibals before Phase 2: a site can be losing the head query to itself, and no new page fixes that. airtabletosheets.com held position 66 to 71 on every head query while its own blog posts held 7 to 31, because three pages split each one and the largest was /setup. A utility page must never carry the head phrase in its title. The read order and why to fix a split before adding a page: workflows.md.

Phase 1. Read where the brand already stands

Diagnosis before production. Ask the assistants the questions a buyer would ask, tally who gets named, and sort what is missing into the six gap kinds (visibility, narrative, topic, format, web mentions, demand), because the fix differs by kind and only one of them is a writing job. The buckets, the fix-build-influence choice and the cadence: references/measurement.md.

Run every prompt signed out, with history and custom instructions off, or the reading is fiction. An assistant tells the account what that account already believes, and the same prompt from two accounts returns opposing answers on questions of fact. A client who has asked ChatGPT about his category twenty times has trained it to agree with him, and will bring that screenshot to the meeting; reproduce it cold before accepting it or arguing with it.

Read three things and none of them is a rank: whether the brand is named, whether the naming is accurate (a confident wrong description is a loss, not a neutral), and which sources the answer cited, because those pages are the real target list for Phase 6.

Phase 2. Judge the cluster before you enter it

The step that saves the most work is the one that stops a cluster from being written at all. Four tests, all of which must pass, and each of them has cost a real cluster when it was skipped: winnable volume rather than total, which slots are winnable rather than whether any are, a qualifier people actually type, and the AI filter. The four with their worked failures, which clusters to own at all, ranking the audience segments before a keyword tool opens, and the prompt research that replaces keyword research in the mention lane: references/content-strategy.md.

Phase 3. Keyword research

python scripts/keyword-research.py -s "your keyword" --depth 3 --brief -o keywords.json

Autocomplete is free, keyless and branched across the alphabet, so a seed returns the tail rather than five phrasings; --no-serper drops related searches and PAA, and the run costs nothing. Supplement through WebSearch for [keyword] reddit, [keyword] alternative and how to [keyword]. Save it to a file: the getseam.app file holds 197 keywords and still plans every article seven months later.

Vet each candidate on business potential, intent and difficulty, then apply the AI filter. To find the click-lane survivors in bulk, filter for transactional intent and include modifiers naming an action: calculator, checker, generator, tool, template, finder, planner, maker. Write in the vocabulary of the forum threads, not of the category: buyers type "sync" and "automatically update" where competitor pages say "ETL", which is why they lose the long tail.

Query length is the cleanest signal of which lane a term sits in, and long qualified questions are mention-lane targets by default; the trigger rates by word count are in references/geo-signals.md.

Phase 4. Intent, then shape

Order the four things a person can be doing by how close they are to paying, deciding then shopping then escaping then learning, and choose the intent before the volume: the four with examples, and why getseam.app's traffic leader converts worst on the site, are in references/content-strategy.md.

Three shapes did the ranking across the 28 articles: 12 comparisons, 6 alternative pages, 10 guides and roundups. The same three are what assistants cite, because a list hands the model a ready-made consensus: 43.8% of pages ChatGPT cites are listicles, and it is the one artefact both lanes want.

• Comparison, You vs Them. Table in the first screen, real measured numbers in it, honest where you lose. One row, CPU during dictation at 12 per cent against 25.6, outperformed the page around it, because nobody else in the category measured anything.
• Alternative, Them alternative. Same reader, one step earlier and angrier.
• Roundup, Best X for Y. Has to list tools you do not own, or it reads as an advert.
• Case study, How we did X for a Y. The shape with no competition. A buyer describes their own situation to an assistant in a sentence no keyword tool has ever seen, and a page carrying that situation with numbers on it is the only thing that matches.

One page per intent. Write only the shape you can support.

Phase 5. Draft, illustrate, publish

python scripts/generate-article.py --keyword "your keyword" --title "Your Title" --word-count 2000

Drafts land in web/content/blog/drafts/. Structure, frontmatter, brief inputs and image guidance: assets/article-template.md.

The draft is the input, never the output. A generated page agrees with everything already ranking, which is exactly what a summariser replaces with one paragraph. The test is whether the page is informationally additive: read the top three results, then name what this page carries that none of them does, your measurement, your customer's words, your screenshot, your opinion. Without that, do not publish. Nothing else here matters more; 96% of the web gets zero Google traffic mostly for failing it.

Then rewrite it so a machine can lift a paragraph out and have that paragraph still be true and still name you. That craft, with the worked before-and-afters, is where most of the measurable citation lift sits: references/write-like-a-source.md. Run references/seo-checklist.md before anything goes live.

Do not answer a thin site with volume. Authority divides across the pages it has to cover, so a small site publishing hundreds of generated pages dilutes what little it has and cannibalises itself; the programmatic lane is for sites with authority to spend. Ramp the cadence, and translate last and only what works: it took 28 articles to 112 pages at almost no cost, and multiplies a mistake as fast.

Phase 6. The off-site half, which is now the larger half

Branded mentions on other people's pages correlate with AI visibility above backlinks, referring domains and domain rating, and most of what a model knows about a brand sits off the brand's own domain. A page nobody else talks about is invisible to the mention lane however good it is. Three tiers in order of value, editorial third parties, then Reddit and forums, then your own other properties, with YouTube as the strongest single factor anyone has measured and a budget line of its own. The tier playbook, the YouTube checklist, the digital-PR campaign that bought 60 national links for about $100, the Reddit dispute, and why a pile of spam links you did not build is left alone: references/earning-mentions.md. Buying links in volume is still the fastest way to lose a site, and white-hat and black-hat practitioners agree on it.

Before it writes, an assistant runs searches of its own. Those fan-out queries brief the page that wins, and the pages it retrieved and did not cite are the sharper half. Read them from a logged-in conversation, never from the paid API, which scored an incumbent at zero in a category it appears in nine times of 24: references/fan-out.md.

Phase 7. Feed the result back

Wait for indexing, filter the console to the new URL, record position and impressions, and run Phase 0 again before the next batch so the cheapest opportunities enter the queue first. Rank tracking alone no longer describes the outcome: the scored per-site advisor report, the three trackable AI signals and the four-layer correlation report are in references/measurement.md.

When the target is local, the playbook diverges completely and the zero-click argument above does not apply: references/local.md.

Boundaries

Structured data mechanics: seo-schema-markup. Console queries and index state: search-console. App stores: aso. Page copy: page-copy, with humanizer for the voice. SvelteKit blog infrastructure: references/blog-setup.md. Auditing an existing brand's citation gaps prompt by prompt is the ai-citation-strategist agent: this skill builds and places pages, that agent diagnoses why a competitor is named instead.

Verification

• Every target sorted into a lane, with the AI filter run against a real SERP rather than guessed from the keyword's shape
• Any claim about AI visibility read from a signed-out session with history off
• The page carries something the top three results do not
• write-like-a-source.md applied, seo-checklist.md passed, no AI crawler blocked
• Off-site mention targets named, not just on-site work
• Each failure mode folded into the phase that would have caught it

Closing a run

This skill appends new failure modes to its own pattern list after each run. A failure mode goes into the phase that would have caught it; if no phase owns it, append it to Learned Patterns with today's date before finishing.

Learned Patterns

• 2026-09-06: Plan the mention lane separately from the ranking lane. Ranking first gives about a 31% chance of an AI citation, and most citations sit outside the top ten. [ask: consolidate GEO/AEO learnings into one skill]
• 2026-09-06: Read AI visibility signed out, history and memory off. A signed-in account returns what it has already been trained to agree with. [ask: consolidate GEO/AEO learnings into one skill]
• 2026-09-06: Where two studies disagree, carry both and say so. A contested number quoted flat becomes a client promise: schema's AI lift is one study for, two against. [ask: consolidate SEO learnings into one skill]