DEV Community: Polydraft

Vibe Coding Invents Your Pricing. Your Marketing Site Needs a Source of Truth

Vincent — Tue, 10 Nov 2026 03:28:33 +0000

If you scroll X this week, the vibe-coding argument is still going. One camp says it gets you a demo, not a product. Another jokes that you are too scared to touch the output because everything might break. A quieter complaint is more useful if you are trying to ship a marketing site: a lot of current vibe-coding tools have no source of truth.

That last gap is what blows up after launch.

You paste a one-line brief into a chat. The model returns a polished homepage, a pricing table, three testimonials, and a line about thousands of happy customers. It looks shippable. Then you notice the monthly plan is a number you never provided. The quoted company does not exist. The feature list includes a dashboard you have not built.

The site runs. The facts are fiction.

The failure is architectural, not a bad prompt

A coding agent is excellent at producing markup and components. It is not automatically excellent at remembering what is true about your business.

In a typical vibe-coding loop it goes like this:

You describe the company in a prompt.
The model fills every gap so the page looks complete.
You tidy typography, spacing, and the hero.
You start a new chat, or the context window rolls over.
The next page — or the next edit — invents a slightly different version of the same company.

There is no durable record of the legal name, the offers, the prices, the claims you will and will not make, who can be quoted, and which URLs already exist. The chat is the database. Chats are a terrible database.

This is why "it runs" is not "it ships" for a marketing site. Shipping means a visitor, a journalist, or an AI answer engine can quote a sentence and you will still stand behind it next month.

Lock the facts before you generate the pages

Before you ask any AI website builder or coding agent to produce a homepage, write down a source of truth a human can review. Keep it boring and specific. If a field is unknown, mark it unknown. A blank section is safer than a fluent lie.

The minimum brief I now refuse to skip

Legal and brand. Legal entity, public brand name, domain, locations, and the one sentence you use to describe what you sell.
Offers. Each product or service in one paragraph, who it is for, and what it is not.
Proof you can defend. Only testimonials, case studies, logos, and metrics you can point to a source for. If you cannot, the section stays empty.
Commercial facts. Prices, plan names, what is included, what is not, and the refund or trial policy — or an explicit "contact us" if those are not public yet.
Constraints. Claims the model must never invent: customer counts, awards, "trusted by" logos, performance numbers, invented case studies.
URLs you already own and want to keep.

Fabricated specifics do not just fail to convert. They get quoted, and then corrected in public.

One knowledge base, many pages

Once those facts live outside the chat, page generation becomes a projection of a record, not a fresh improvisation.

That is the difference between prompting a homepage and operating a site:

Change a price in one place, and every plan table can be updated from the same record.
Add a second offer without rewriting the "about" story two different ways.
Grow a campaign page into a multi-page site without inventing a new company on the services page.

If you are still deciding whether to launch one page or a small site, that choice is about traffic and offers, not about whether the facts are stable. We published a landing page vs website decision table for that call. The source of truth should be the same in either case. Launch on the URL you plan to keep, then add pages as sales and search questions show up.

A four-question test you can run on any tool

Ask the tool, or your own agent workflow, these questions before you trust it with a live domain:

Where are the business facts stored after the first prompt?
If I start a new chat, does it still know the approved price and the claims I forbade?
What does it do when it does not know a number — leave a gap, ask, or invent one?
When I correct a fact, do existing pages update from that correction?

If the only memory is the current thread, you do not have a website system. You have a demo generator.

Those questions sit next to the ownership and export checks in our guide to evaluating an AI website builder. The demo will always look fine. The source of truth is what you are left holding after the context window is gone.

What we do in PolyDraft, and what we do not claim

This post is written by the PolyDraft team, so here is the honest version.

When you describe the business or paste an existing public URL, PolyDraft turns your facts, products, goals, and media into a verified knowledge base — the source of truth for every page and later update. Copy is supposed to stay grounded in those reviewed facts. You inspect the complete preview, fix anything that is wrong, and approve what goes live. We do not invent customer results or traffic lifts, and a fact marked unknown should stay blank rather than become a statistic.

That is the product idea: an AI website builder that researches and builds from a record you can review, not from a disposable chat. Whether you start with one page or a full site, the facts should not have to be re-explained — or re-invented — each time you ask for a change.

If you are vibe-coding the site yourself, steal the same pattern. Keep a SOURCE_OF_TRUTH.md (or a small CMS collection) that every prompt is required to read, and treat any fact not in that file as out of bounds.

A short workflow you can use this week

Write the brief above. Have a second person who knows the business strike anything they cannot defend.
Decide landing page vs website, and reserve the URLs you want to keep.
Generate the first preview only after the brief is approved.
Reject any invented metric, logo, or quote on sight. Do not leave it to "fix later."
Make one factual change — a price, a service name — and confirm every affected page updates from the same record.
Only then publish.

Vibe coding can still be the fastest way to get a layout. It is a bad place to store the truth about a company.

SEO/GEO Autopilot Is Not Autopilot: How a Growth Engineer Runs B2B Web Growth With Human Approval

Vincent — Mon, 09 Nov 2026 17:33:12 +0000

TL;DR. “SEO/GEO Autopilot” should automate evidence gathering and draft proposals—not unattended publishing. A Growth Engineer runs B2B web growth as a governed loop: read Search Console / Bing signals → find a justified opportunity → prepare article and internal-link changes → get human approval → verify the live release, then learn. PolyDraft already ships research-led builds, technical SEO, GEO artifacts, and a CMS you operate; a weekly GSC + Bing recommendation cycle is in internal testing (Public Beta is not open). Nobody can guarantee AI citations. Measure search performance, AI appearances, and business outcomes separately.

Disclosure. This post is written by the PolyDraft team—the people who build the product. We describe what ships today, what we are testing internally, and where a Growth Engineer / AI FDE engagement fits. We do not invent customer results, traffic lifts, or citation guarantees.

The problem: “AI SEO” is treated like set-and-forget

Most B2B teams still confuse launch day with growth.

They ship a site (or an AI-generated one), tick “SEO” in a checklist, maybe paste a few keywords into meta fields—and wait. When AI tools enter the picture, the temptation grows: schedule more posts, auto-refresh dates, auto-link every matching phrase, and call it autopilot.

That pattern fails for three reasons:

Helpful content is not a freshness quota. An accurate page may need no edit. Google’s people-first content guidance rewards useful substance, not cosmetic rewrites.
Agents amplify blast radius. An agent with CMS write access can publish inventable claims, wrong redirects, or scaled pages that help nobody—exactly the pattern Google’s spam policies describe as scaled content abuse.
Search, AI appearances, and revenue are different questions. A citation is not a visit. An impression is not a qualified inquiry. Mixing them into one “visibility score” hides the truth.

Principle. Automate the recurring, checkable work. Keep humans accountable for what the website claims.

That is the operating idea behind what we call SEO/GEO Autopilot—and why the word autopilot is deliberately misleading if you expect unattended publishing.

Meet the Growth Engineer

We needed a role name that matches how B2B web channels actually grow in 2026—not a silo of “SEO person,” “content person,” and “web person” handing tickets across tools.

A Growth Engineer sits at the bridge of:

Domain	What they own
Product / facts	Verified business claims, pricing, capabilities, entity names
SEO + GEO	Query-to-page fit, technical access, answer-first structure, crawler policy
Ops / agents	Permissions by blast radius, approval gates, release verification

They do not treat agents as a replacement for judgment. They treat agents as labor for gathering evidence, clustering opportunities, drafting refreshes, and running QA—while a named person approves the exact version that goes live.

When a team wants that role done-for-you on our platform, we offer AI FDE (AI forward-deployed engineering): a PolyDraft engineer runs PolyDraft for you—building a new site from your current one, plus technical SEO, content, and AI-visibility work. The agent still does not publish without your approval. We frame that engagement as a Growth Engineer practice: same loop, same accountability, shipped with you in the driver’s seat.

What “SEO/GEO Autopilot” actually means

On our SEO automation page we draw a hard line:

SEO automation = software for repeatable analysis, checks, and preparation.
Agentic SEO = a bounded workflow with accountable decisions—not permission for an agent to publish anything it generates.
GEO (generative engine optimization) = make your site easy for AI assistants and AI search features to find, read, attribute, and quote—complementing traditional SEO, not replacing it.

PolyDraft’s product direction is a weekly cycle that turns Google Search Console and Bing Webmaster performance into content and internal-link proposals, with human approval before publication. That cycle is in internal testing. Public Beta is not open, and no release date is announced. Starting a free build does not grant access to the weekly workflow.

What is available today as foundation:

Research-led page planning (live volume, difficulty, intent)
Technical SEO defaults (titles, meta, schema, canonicals, sitemap, robots.txt, RSS)
Blog CMS, lead inbox, internal-link manager
GEO artifacts: answer-first sections, llms.txt, AI-crawler-friendly robots.txt

From our own build log (verified product figures, not customer traffic claims): 100+ websites built, 20+ industries, brief to live preview in under 2 hours, content in up to 100 languages. Build free; pay to publish. Sites deploy to your own Cloudflare account and domain.

The best-practice operating loop (5 steps)

This is the Growth Engineer loop we recommend for B2B web channels—whether you run it yourself or with AI FDE. It matches the internally tested weekly cycle on our SEO automation page.

Step	Action	What “done” looks like
1. Read evidence	Pull GSC / Bing performance for a defined period	Finding tied to source, page, and observation window
2. Find opportunity	Incomplete answer, outdated claim, or missing contextual links	A justified improvement to an existing URL preferred over a near-duplicate
3. Prepare changes	Draft article edits + internal-link proposals	Change explained; verified facts and working URLs preserved
4. Human approval	Named person checks facts, priorities, scope, links	No approval → no publication
5. Verify, then learn	Check live page/links; compare later performance carefully	Before/after is an observation, not proof of causation

Weekly review is not a promise of weekly publication or weekly growth. Accurate articles may need no change.

People-first selection questions before you edit:

Signal	Ask before editing	Possible decision
Relevant queries, incomplete answer	Does this URL own the task? Is the missing info verified?	Add the missing explanation to the existing article
Outdated product claim	What does current product docs actually support?	Correct the claim and its source—regardless of traffic
Useful page, few contextual links	Which articles genuinely lead to this next task?	Propose descriptive links from relevant passages
Clicks fall, page still answers its task	Demand, indexing, query mix, device, comparable periods?	Investigate or monitor—unchanged can be correct

Permission matrix: autonomy by blast radius

From our guide on how to use AI agents in SEO, autonomy has five ceilings: Observe → Recommend → Draft → Execute after approval → Execute within policy.

Condensed starting ceilings for a Growth Engineer:

Action	Blast radius	Starting ceiling
Read Search Console / analytics (read-only)	Low	Observe on a schedule
Crawl + report issues	Low	Observe, rate-limited
Cluster keywords / propose owner URLs	Medium	Draft
Write briefs and first drafts	Medium	Draft
Change internal links, titles, descriptions	Medium	Execute after approval
Publish or update indexable pages	High	Execute after approval of the exact version
Change robots.txt, redirects, canonicals	High	Execute after approval (engineer)

Connect data read-only first: Restricted Search Console user, Read-only Bing user, dated keyword/crawl exports. Raise ceilings one level at a time based on a clean track record.

First-party process evidence (our own blog QA)

When we applied this workflow to our own article series, separate checks caught real defects before readers saw them (logged October 12–13, 2026):

Automated QA on first drafts: 3 errors fixed before review (including an unqualified “guarantee” and two overlong meta descriptions)
Adversarial review of the publishing script: 18 confirmed findings (10 distinct problems), including a path that could have re-published a draft a person had deleted
Fact-check against sources: wrong date labels, wrong Google page citations, inflated counts—all corrected before publishing

Judge the agent by corrections and defects before you judge it by traffic.

GEO without the hype

Generative engine optimization services are checkable work, not magic:

Crawler access decisions (per user-agent, not a vague “allow AI”)
Answer-first openings on key pages
Consistent company / product facts everywhere
Structured data that matches visible page content
Dated prompt tests + Search Console / referral analytics—kept separate

Google documents that its AI features are part of Search, with no separate technical requirements beyond ordinary eligibility. Optional files like llms.txt are indexes, not ranking levers. Nobody can guarantee citations. Assistants vary by system, session, and wording.

Measure separately:

Measure	What it tells you	What it does not
Release checks	Right page/metadata live and readable	Whether anyone will cite you
Dated prompt sets	What named assistants answered on a given day	A trend unless repeated under same conditions
Search Console (incl. generative AI report where available)	Impressions / clicks in Google Search & AI features	Citations inside every other assistant
Referral analytics	Visits that pass a referrer from AI tools	Visits that show up as direct / unknown

Do not add unlike metrics into one unexplained visibility score.

What PolyDraft ships today vs. direction

Ships today (features):

AI website builder that researches, writes, builds, and quality-checks a complete preview
Keyword research with volume / difficulty / intent behind page plans
Technical SEO by default + GEO-ready structure (answer-first, schema parity, llms.txt, AI crawler policy)
Owner-operated CMS: blog, inquiry inbox, internal-link manager, localization up to 100 languages
Deploy to your cloud; Agency includes source export
Honest-content policy: no fabricated stats, reviews, or client names

Direction / internal testing:

Weekly GSC + Bing → content & internal-link recommendation cycle with human approval
Recurring supervised recommendations of what to publish or refresh next

Explicit non-promises:

No unattended SEO/GEO publishing
No citation or ranking guarantees
Free build ≠ access to the weekly Autopilot workflow

When AI FDE / Growth Engineer service fits

Option	Who does the work	Good fit when
DIY with the loop above	Your team	You have time and a manageable site
Software + your approval	Tools + you	You want monitoring and drafts
Traditional agency	Specialists	You want strategy handed off—and you keep account ownership
PolyDraft Growth Engineer (AI FDE)	A PolyDraft engineer running PolyDraft for you	You want a new site and its SEO/GEO ops done-for-you on the platform, with approval gates

Ask any provider—including us—for a baseline, change log with approvals, facts sheet, and measurement plan with raw prompt answers. Walk away from guaranteed “AI rankings,” mass-produced pages, or accounts held only in the provider’s name.

Mini checklist (use without buying anything)

Copy this into your next weekly triage:

[ ] Pull GSC (and Bing if available) for a fixed period; attach source + page + dates to each finding
[ ] Prefer improving an existing URL over creating a synonym page
[ ] Require a source for every statistic; leave unknowns as “unknown”
[ ] Draft changes with explicit draft status; approve the exact version that publishes
[ ] After publish: check live status, content, canonical, and key links
[ ] Review robots.txt per AI crawler purpose (search / training / user-triggered)
[ ] Keep search metrics, AI prompt observations, and pipeline metrics in separate columns
[ ] Log human corrections and QA defects—not only traffic

Free utilities if you want a quick technical pass: llms.txt generator and robots.txt generator.

FAQ

Is SEO/GEO Autopilot the same as unattended publishing?

No. Autopilot here means automating evidence gathering and draft preparation. Publication stays behind human approval. PolyDraft’s weekly cycle is in internal testing and is not an always-on publish bot.

Can AI completely automate B2B SEO?

Repeatable tasks yes; factual accuracy, business priorities, and what the site should claim still need accountability. Agents prepare; Growth Engineers decide.

Will GEO get ChatGPT (or any assistant) to cite us?

No provider controls that. You can improve crawler access, clarity, consistency, and measurement. Citations may come slowly—or not at all—even when the work is sound.

What does PolyDraft do today for SEO and GEO?

Research-led builds, technical SEO defaults, GEO artifacts, CMS operations, and owner-requested changes with review before publish. See SEO automation and features. The weekly recommendation cycle is not publicly available yet.

Who is a Growth Engineer in this model?

The person (or AI FDE engagement) who runs product facts + SEO/GEO + agent ops as one governed growth channel—blast-radius permissions, approval gates, and separate outcome measures.

Sources

Primary product pages (PolyDraft):

External guidance:

Build free, publish when ready, operate deliberately. If you want a Growth Engineer engagement on the platform, start a free build or ask the PolyDraft team about AI FDE availability—signup does not automatically unlock the weekly Autopilot workflow.