Our Agentic SEO Pipeline: 1,559 Clicks, 0 Activations

Loadout and a coding agent can run an evidence-backed SEO operation from live research through a reviewed pull request, publication, and measurement. Our current system does that. It also exposes the part that is not working yet.

Across equal, complete 28-day Google Search Console windows, Aident blog clicks rose from 40 to 1,559 and impressions rose from 20,512 to 101,008. In a separate matured product cohort, only 3 of 1,878 blog visitors signed up after a blog touch, and none produced an observed installation, activation, successful Action, or payment within seven days.

This is Part 1 of an open case study. The useful story is not that an agent can publish a lot of pages. It is that a coding agent can operate a source-backed system with cost, quality, approval, and measurement gates, then tell us plainly where the funnel still breaks.

The Current Result

The search comparison uses page-level Search Console aggregates for blog URLs on aident.ai. The periods are equal and do not overlap.

This is observational evidence, not a causal experiment. We published many pages during the period, the number of blog URLs appearing in Search Console increased, search demand changed, and indexing had more time to compound. The comparison establishes that the operating system can capture materially more organic discovery. It does not isolate which part of the system caused the increase.

The product result comes from a different cohort because downstream behavior needs a complete follow-up window. For visitors first seen on the blog from July 22 through August 11, 2026:

The same-person sequence is evidence of association, not proof that a page caused the later action. It is still enough to reject the flattering but false conclusion that more search traffic already means business growth.

The Four Parts of the System

Our pipeline separates responsibilities instead of asking one model prompt to do everything.

This division matters. The agent can work broadly inside a reviewed contract, while consequential actions remain visible and stoppable.

The Pipeline, Step by Step

1. Reconcile Reality Before Choosing Work

Every run begins from current code and compares three states:

• the articles tracked in the repository;

• the items actually present in the Framer Blog collection; and

• open or recently merged automation pull requests.

At the start of this run, the repository tracked 255 valid posts, Framer contained 247 published items, and 8 merged posts were ready for publication. That difference could look like an invitation to publish immediately. A read-only Framer schema check found duplicate field definitions for every core Blog field, so the agent stopped all CMS writes instead of guessing which field ID was correct.

That stop is part of the pipeline. Automation that cannot refuse an ambiguous write is not production automation.

2. Collect Evidence Through Loadout

The coding agent searches for the current Actions that match the job, inspects their input schemas, and preflights the exact inputs before execution. For this case study it used connected, read-only Search Console data for discovery and existing PostHog evidence for downstream progression. It also inspected the available Ahrefs research contracts.

The Ahrefs inputs validated, but their estimates remained unbounded above our 100-credit approval gate. The agent did not execute them. Search volume, keyword difficulty, and traffic potential therefore remain unknown rather than being filled with invented numbers.

This is the practical advantage of Loadout for coding agents: the model can discover and use live systems, but price, permissions, schemas, and account state remain explicit parts of the decision.

3. Build a Candidate Set, Not a Keyword Conveyor Belt

The agent must consider several kinds of work before drafting:

• an awareness or problem-education opportunity;

• a solution-consideration or evaluation opportunity;

• an integration-led workflow using a capability available now;

• a keyword-compounding refresh for an existing canonical; and

• an answer-engine mention or citation gap.

Each candidate is checked against the existing catalog before it can become a new URL. If a page already owns the reader job, the correct action is usually to preserve or refresh that page. This canonical gate keeps the agent from manufacturing near-duplicates simply because a keyword variant exists.

Our scoring then considers reader urgency, evidence of demand, Aident's firsthand expertise, the distinctness of the answer, product progression, distribution potential, and portfolio balance. A draft needs both a high score and a supported content archetype. Volume alone is not a pass.

4. Draft an Answer With Its Limits Attached

A selected article must give the direct answer early, use primary sources for claims that can change, and keep observations separate from conclusions. Every article also declares:

• one canonical URL;

• one funnel stage and content lane;

• the primary distribution channel;

• zero to three contextual next-step articles;

• one measurable call to action; and

• a day-7, day-30, or day-90 decision window.

This case study is a solution-consideration article. Its primary outcome is attributable product-page visits at day 30. We are establishing a baseline rather than inventing a target before the first comparable window exists.

5. Turn Quality Rules Into Files and Checks

The coding agent does not hand a prose document directly to a CMS. It produces four reviewable artifacts:

content.md
artifacts.json
stats.json
cover.jpg

The content stays CMS-neutral. The publishing artifact records Framer fields, SEO metadata, canonical ownership, distribution, next steps, and the tagged CTA. The stats artifact declares the measurement contract before publication. The cover must be original, topic-specific, and safe for both landscape and square crops.

Repository checks validate the JSON schemas, internal links, image dimensions, funnel metadata, search archetype, canonical action, CTA, and publication state. The pull request remains the approval boundary. The lifecycle itself never merges its own work.

6. Publish Only a Merged, Exact Artifact

Publication is a separate lifecycle action. The agent may publish only an article that is already merged into the main branch, marked ready for the CMS, and still matches the reviewed slug, body, metadata, and cover.

After a write, the agent must verify the public canonical, expected title, body markers, contextual links, tagged CTA, and durable image. A successful provider response without a working public page is not completion.

The current duplicate-field condition in Framer fails this gate. This article and the existing ready queue will remain unpublished until the schema is unambiguous.

7. Measure Discovery and Progression Separately

The system preserves immutable day-7, day-30, and day-90 snapshots. Search impressions, clicks, position, answer-engine visibility, and referring sources measure discovery. Next-step clicks, product-page visits, setup starts, connections, activations, users, payments, and revenue measure progression.

Missing attribution stays null. An observable event count of zero stays zero. Those are different claims.

The deeper method is documented in Measure Blog-to-Product Attribution With PostHog and Codex. Our earlier research on AI agent SEO and product growth explains why exact-problem pages won discovery before the product bridge was instrumented well enough.

What the Coding Agent Does Well

The current system is strongest at work that benefits from breadth plus strict bookkeeping:

• querying several current sources without copying credentials into the workspace;

• preserving exact date windows, query scopes, and caveats;

• checking every new idea against hundreds of existing canonicals;

• applying the same metadata and measurement contract to every article;

• generating a complete review packet instead of a loose document;

• stopping when cost, permissions, or a target schema is ambiguous; and

• returning later to the exact artifact when its measurement window is due.

That is a better use of a powerful coding agent than asking it for twenty keyword articles. The leverage comes from letting it operate the whole evidence and review loop while retaining explicit boundaries.

What Still Needs Human Judgment

The agent can enforce a strategy, but it cannot decide the strategy's values for us. People still own:

• which audience and product outcome matter;

• what evidence is strong enough to publish;

• when an observed association is worth a new experiment;

• whether a draft represents Aident accurately;

• whether an external write or spend should proceed; and

• whether the system should favor traffic, trust, activation, or revenue in the next cycle.

Our current judgment is clear: protect the pages that already rank, reduce traffic-only expansion, and focus the next experiments on the bridge from exact technical answers to useful product workflows.

What We Will Report Next

This series will keep the denominators visible. Future updates will report:

1. whether the publication queue and Framer schema blocker were resolved;

2. this page's day-7, day-30, and day-90 discovery and progression windows;

3. changes in portfolio search clicks and concentration across equal windows;

4. next-step clicks, product-page visits, setup starts, connections, and activations;

5. first-touch and assisted signups as separate measures; and

6. payments or revenue only when a connected source supports attribution.

We may update this canonical with longitudinal evidence or publish another installment when the next result answers a genuinely different question. We will not create a thin recap merely because a calendar date arrived.

Run a Bounded Version With Your Coding Agent

Follow https://aident.ai/SETUP.md

Then ask your coding agent to use Aident Loadout for one read-only opportunity cycle: reconcile your current content catalog, query a complete Search Console window, inspect product progression separately, propose at least three distinct actions, and stop before any write. Require the agent to return its sources, date bounds, canonical decision, current price quote, and unknowns.

Set up Aident Loadout and run one evidence-backed SEO cycle.

For the general discovery, schema, preflight, and execution model, continue with How to Use Aident Loadout.

Sources and Method Notes

• Google Search Console Search Analytics API

• Google: About Search Console data

• Google: Creating helpful, reliable, people-first content

• PostHog HogQL documentation

• GitHub: About pull requests

• Aident repository-tracked Framer, Search Console, PostHog, strategy, and milestone evidence, inspected August 25, 2026

Refresh this case study when the article reaches a declared measurement window, the publication contract changes, the product attribution contract changes, or a comparable cohort shows that the current discovery or progression pattern has materially changed.