How This Website Was Built and Is Run with PolyDraft: The Build Log
A dated record of how polydraft.ai was built by PolyDraft's agent, how the site is run week to week, what went wrong, and where its search numbers stand.
PolyDraft's own agent built this website in summer 2026 from a plain-language brief: the research, the keyword plan, the page structure, the copy and the code. Since October we've run the site the way we're building PolyDraft to help run one: AI agents plan and draft from search data, separate agents fact-check, and a script checks every release. This log records both, with dates, the mistakes, and the search numbers so far, which are still small.
Key facts, as of November 6, 2026
| Built | Summer 2026, by PolyDraft's agent, from a written brief; our earliest CMS records date from August 12, 2026 |
| Stack | Astro pages rendered on the server, deployed to our own Cloudflare account |
| Size | 78 URLs in the sitemap: 40 blog posts, plus product, audience, comparison, tool, showcase, author and policy pages and the blog index and category pages |
| Run since October 13 | 6 release records: 20 new posts and 4 refreshes, each through the workflow below |
| People involved | Vincent Chen approves releases; AI agents do the research, drafting and fact-checking |
| Search results so far | 49 clicks and about 5,700 impressions from Google in the three months to November 5; average position 52.8 |
How PolyDraft built the site

The brief, not a template
The build started with a plain-language brief: what PolyDraft is, who it serves, and what the site must never say. The agent treated that brief and the knowledge base built from it (company details, plan prices, product capabilities) as its only source of business facts. If a fact wasn't there, it didn't go on the site. Two of its rules: no ranking guarantees, and no invented testimonials or statistics. The agent applied them on every page.
One primary keyword per page
The brief called for exactly one primary keyword per page, so no two pages compete for the same query, and the agent built an internal-link dictionary that points each keyword at the page that owns it. We still work this way. Here are some of the page owners today, with the search data behind them:
| Page | Primary keyword | US monthly searches | Difficulty | Data date |
|---|---|---|---|---|
| Homepage | ai website builder | 27,000 | 48 | Sep 29, 2026 |
| /ai-visibility | AI visibility | 4,200 | 31 | Oct 31, 2026 |
| /seo-automation | SEO automation | 2,200 | 15 | Oct 2, 2026 |
| /website-localization | website localization | 700 | 0 | Sep 28, 2026 |
| /blog/lovable-vs-base44 | lovable vs base44 | 700 | 1 | Sep 17, 2026 |
| /blog/lovable-alternatives | lovable alternatives | 700 | 1 | Sep 23, 2026 |
| /blog/framer-vs-webflow | framer vs webflow | 700 | 7 | Oct 29, 2026 |
Figures are Ahrefs estimates for the US from our keyword library, with the date each was captured. A difficulty of 48 is hard for a new site to reach, which is why most of our blog posts target low-difficulty keywords first.
Writing, then checking
Each page was written from its own research brief: sections, target length, structured-data type, internal links and the questions the FAQ should answer. Key sections open with the answer, because that's the part AI assistants quote. An automated quality gate then checked the whole site for content depth, SEO completeness, accessibility and design consistency, and the issues it found were fixed before launch.
The technical layer came with the build; it's on every PolyDraft plan:

You can check two of these on this site right now: our robots.txt, which lets AI search crawlers such as GPTBot, ClaudeBot and PerplexityBot in, and our llms.txt, a plain-text summary for AI systems. We take that file apart, including a charset problem we found, in our llms.txt example and validation guide.


How we’ve run the site since October
A website isn't finished at launch. Most of the work is deciding what to publish next, writing it well, linking it, checking what performs and refreshing what doesn't. That's the work PolyDraft exists to do, so we run it on our own site first.

The release workflow
Every release in our records since October 13 has followed the same steps:
- Pick the keyword from data. Each post starts from a row in our keyword library: search volume, difficulty and the page that should own it. If another page already owns the keyword, we refresh that page instead of writing a new one.
- Brief and draft. An AI agent writes a brief (reader, angle, what the post must answer) and then the draft, using only sources it can cite.
- Automated QA. A script checks structure, links, sources, image alt text, metadata and wording, and blocks the release on any error.
- Separate fact-check. Another AI agent, given the draft but not the writer's reasoning, checks every claim against the original source. It doesn't fix anything itself. It reports, and the writer fixes.
- Approval. Vincent reads the result and approves the release before it goes live.
- Publish and verify. The post goes live through the site's content API. Then we check the live page: status code, canonical tag, one H1, structured data, the mobile layout, and that every image loads.
- Link it in. We add links from related posts, each with a dated update note, and notify search engines through IndexNow.
| Release | Date | What shipped | What the checks caught |
|---|---|---|---|
| W42 | Oct 13 | 5 posts, an audit of 17 existing posts, 18 inbound links | A five-agent fact-check corrected claims after approval |
| W44 | Oct 26 | 5 posts, 6 inbound links | Mobile tables overflowing the screen; fixed on all 11 posts from W44 and W45 |
| W45 | Oct 26 | 6 posts, 15 inbound-link edits | Overlap with an older post on the same topic; FAQs swapped |
| Refresh | Oct 29 | /blog/lovable-alternatives rewritten | 4 blocking errors in the rewrite, including a competitor's export options and a claim about PolyDraft's own rendering |
| Batch | Nov 5 | 4 posts | A draft implied hands-on testing we hadn't done; rewritten |
| Refreshes | Nov 6 | 3 comparison posts, with screenshots of each vendor's pages | Out-of-date prices and plan limits; claims about PolyDraft tightened to what has shipped |
Eight October posts have no release record: five published on October 8 and three on October 22. We don't count them here. Our records also have two gaps: in W42, fact-check corrections landed after Vincent's approval, and the October 29 refresh has no written approval in its record.
What went wrong
- This log overclaimed. Earlier versions said the code you were reading was entirely the agent's and that "nothing in this process is specific to us". Neither stayed true: people and AI agents have edited the site since the build, and our release workflow is something we run on our own site, not yet a product feature. This version says which is which.
- A post we cared about wasn't indexed. In late October Search Console showed our Lovable alternatives guide as not indexed. It was likely too thin and had factual errors, so we rewrote it with a sourced table and added links to it from related posts.
- Our auto-linker made bad links. It linked the word "pricing" in a paragraph about another company to our own pricing page, and linked a negative phrase on an earlier version of this page to our About page. We fixed the pricing links; the About-page rule, flagged on October 13, still needs an engineering fix.
- Search-engine pings stopped being automatic. The site's IndexNow submissions started failing because of rate limits on shared Cloudflare addresses, so URLs now queue until we send them by hand.
- Unknown blog addresses redirect. A mistyped blog URL redirects to the blog index instead of returning a "not found" page, which Google can treat as a soft 404. It's on our fix list.
Where the site stands
The honest answer is: early. In the three months to November 5, Google Search Console recorded 49 clicks and about 5,700 impressions for polydraft.ai, with an average position of 52.8. Ahrefs showed 0 organic keywords for the site in the US on November 5. Links alone haven't changed that: Ahrefs put our Domain Rating at 0.6 with 4 referring domains in late August, and at 7 with 40 referring domains by October 5, still with no organic keywords. A new domain with few links takes time, and most of our posts are weeks old. We'll update these numbers here as they change.
What PolyDraft does today, and what’s next
The build is what PolyDraft does for every customer: research (keyword data depends on PolyDraft's research provider being available), a page plan with one keyword per page, original copy, the technical layer, a quality check and deployment to your own Cloudflare account.
The release workflow above is now part of the product. New PolyDraft sites come with a content calendar in their admin: topic clusters and a dated publishing queue drawn from keyword and question data (or, when keyword data isn't available, from questions in search results), plus AI-predicted sub-queries that answer engines run for those topics. On top of it, the SEO and GEO autopilot reads each site's Search Console data every week, writes new articles, refreshes the ones that slip and reports what moved. polydraft.ai was built before the calendar existed, so we still run ours with our own workflow.
If you want to see how a build works, read how it works or look at the sites in our showcase.
Frequently asked questions
Was any of this website written by people?
The original build was produced by PolyDraft's agent from a written brief. Since then, people and AI agents have edited pages, and Vincent Chen approves blog releases before they go live.
What data did the keyword plan use?
We don't have the original build's keyword file in a form we can publish. Today we plan from Ahrefs data in our keyword library, with one primary keyword per page; some page owners and their data dates are in the table above.
How is the site maintained now?
Through the release workflow in this log: a keyword from our library, a brief and draft by an AI agent, an automated QA script, a separate fact-check, approval, a live check, and links from related posts.
Has the site ranked yet?
Barely. Google showed polydraft.ai about 5,700 times in the three months to November 5, for 49 clicks, at an average position of 52.8. Ahrefs showed 0 organic keywords in the US on the same date.
Continue with the next task
- Judge any AI-built site with an independent checklist: How to evaluate an AI website builder
- See the playbook for a site with no traffic yet: SEO launch plan for a new website
- Keep published pages current: SEO content refresh workflow
- How we use AI agents in this work: How to use AI agents in SEO
About this article. Drafted by an AI agent following PolyDraft's blog workflow, checked by an automated QA script, fact-checked by a separate AI agent, and approved by Vincent Chen. Release details come from our release records, post dates from our CMS, search figures from Google Search Console and Ahrefs with the dates shown, and screenshots were captured on November 6, 2026. Updated November 6, 2026: rewritten as a dated log, with the release workflow, mistakes, current search numbers and keyword owners added, and claims corrected. Earlier update: October 8, 2026, source boundaries clarified.
Sources
- polydraft.ai: robots.txt, llms.txt, Pricing, How it works
- Google Search Central: robots.txt introduction, HTTP status codes and soft 404 errors
- Schema.org: Schema.org vocabulary
- IndexNow: IndexNow protocol
Revision history: November 6, 2026: corrected the publication date to August 12, 2026, when it first went live, and redrew the posts-per-month chart to match.


