Your landing page has had 400 visitors and 11 signups. You have rewritten the headline twice this month, both times on a hunch, and the number has not moved in a direction you can distinguish from noise. So you open the AI Advisor hoping it will hand you a better sentence.
It will not. The advisor has nine sections and not one of them writes copy. What it does have is the only thing that makes a rewrite worth doing: it knows which of your pages converts worst, and it has read what visitors typed into your chat widget and your contact form. That is the page to rewrite and the vocabulary to rewrite it in.
The AI Advisor does not generate copy. Use its Analytics card to pick which page to rewrite (it is fed your top 15 pages with views, visitors, and conversions each), then mine the Chat and Messages cards for the exact phrases visitors used, and write the line yourself in their words. Ship the variant at a new path, not the same one, or the advisor cannot tell the versions apart.
The Nine Sections, and Which Three Matter Here
OperatorStack scores nine sections independently: analytics, signups, referrals, messages, forms, chat, performance, errors, and installation. Each gets its own prompt, its own slice of your data, and its own red/yellow/green priority.
Three of them are inputs to a copy decision. The other six are not, and reading them while you are trying to fix a headline is how an afternoon disappears.
- Analytics answers "which page". It is the only section that sees per-page conversion.
- Chat answers "in what words". It is fed the topics visitors discuss plus the raw text of recent messages from unresolved conversations.
- Messages answers "what they asked when chat did not cover it". It is fed recent subjects and the body of unanswered contact messages.
Do not paste the advisor's ACTION line onto your page. That field is capped at 180 characters because it renders into a fixed-width dashboard component that truncates mid-sentence if the model runs long. It is written as an instruction to you, not as customer-facing copy.
Step 1: Let the Analytics Card Pick the Page
Open the Analytics card and read the page list, not the summary. The section prompt is built from your top 15 pages, formatted one per line as page: N views, N visitors, N conversions. That third number is what matters. A page with 900 views and 2 conversions is a copy problem. A page with 40 views and 2 conversions is a traffic problem, and no headline fixes it.
Check whether the card is even allowed to speak yet. The overview endpoint returns has_sufficient_history: false until your project has 24 hours of events behind it, and the advisor will not score a section before that gate clears. If you installed the script this morning, there is nothing to read.
The advisor also receives a separate New pages (first seen in last 7 days) block and a 14-day daily series for your top 5 pages. Both matter in step 3, when you want it to notice that a rewrite exists.
If you only have one page, the Analytics card cannot rank anything and will mostly comment on the traffic trend. Split your pricing or your FAQ onto its own path before you expect per-page copy advice; a single-path site gives the section one row to reason about.
Step 2: Mine the Chat and Messages Cards for Vocabulary
This is the step that actually changes the words on the page, and almost nobody does it.
The chat section prompt includes a Topics visitors discuss list and, separately, the recent message text from conversations that are still active or unresolved. The messages section prompt includes recent subject lines, themes, and the body of every message that is still unanswered. Between them, you are looking at a transcript of strangers describing your product in their own vocabulary, unprompted, without a survey framing the answer.
Read the raw text, not the advisor's summary of it. You are looking for three things:
A noun you do not use. If four people asked about "the free tier" and your page says "Starter plan", your page is losing the search. Use their noun.
A question your page should have already answered. Every repeated pre-signup question is a line of copy that is missing or buried. Three people asking whether it works with Framer means your integration list is not visible above the fold.
The outcome they named. Founders write features. Visitors write outcomes. If the chat log says "I just want to know where my signups came from", that sentence is closer to a working headline than anything on your page right now.
The Messages card explicitly flags unanswered messages and asks the model to summarize what each is asking. That list is the highest-signal copy input in the product, and it is sitting in a card most founders treat as a support queue.
Step 3: Ship the Variant at a New Path
Here is the mechanic that decides whether the advisor can help you a second time.
The Analytics section is fed pages, keyed on path. If you overwrite your headline at /, the before and after collapse into one row and nothing in the advisor's input distinguishes them. Ship the rewrite at its own path instead, and it appears in the New pages (first seen in last 7 days) block with its own views, visitors, and conversions.
Then tag the conversion side so the numbers are yours rather than inferred:
await OperatorStack.ready;
// Fire on the page that carries the rewrite
OperatorStack.trackEvent("hero_variant_view", {
variant: "outcome_headline",
path: window.location.pathname,
});
And at the moment the visitor converts:
await OperatorStack.ready;
OperatorStack.trackEvent("waitlist_joined", {
variant: "outcome_headline",
path: window.location.pathname,
});
Both calls queue internally until the SDK finishes initializing, so they are safe to run during hydration in Nuxt or React. Events batch and flush every 5 seconds, with a sendBeacon flush on unload, so a visitor who bounces still gets counted.
Split your traffic with UTM parameters on the links you share, which the embed already captures with no extra code. Two copy variants at two paths, each with a tagged share link, is a complete test.
Step 4: Wait Out the Cooldown
You cannot iterate on copy hourly, and the product will not let you pretend otherwise.
A section re-analyzes automatically at most once every 24 hours (ADVISOR_COOLDOWN_HOURS), and the manual refresh button runs on a separate 12-hour floor per project (ADVISOR_FORCE_COOLDOWN_HOURS). The overview response carries refresh_available and refresh_available_in, the seconds left until the button works again.
That cap is doing you a favor. A headline change needs enough conversions to separate signal from noise, and on a pre-launch site that is days, not hours. Refreshing the card at lunch tells you nothing you did not know at breakfast.
What This Will Not Do
Three honest limits, so you do not wait for advice that is not coming.
- No section grades a sentence. The advisor never reads your page. It reads paths, counts, and message text. A beautiful headline on a page with no conversions and an awful one on a page with plenty will be scored identically, on the numbers alone.
- Low conversion counts produce vague advice. The system prompt tells the model to judge relative to the project's own scale and to say so plainly when nothing stands out. With 3 conversions across 15 pages, "nothing stands out" is the correct answer and the one you will get.
- It cannot separate copy from offer. A page that stops converting after a pricing change looks exactly like a page that stops converting after a headline change. Change one thing per path.
Frequently Asked Questions
Can the AI Advisor write my landing page copy?
No. The advisor has nine sections and none of them generate copy. Its ACTION field is capped at 180 characters because it renders into a fixed-width dashboard card, not a hero block. It tells you which page is losing people and what visitors are asking about. The sentence is still yours to write.
How do I get AI to write landing page copy from real data?
Do not start from a generator. Pull the phrasing out of your Chat and Messages cards first, since the advisor is fed the topics visitors discuss and the subjects of recent unanswered messages. Those are words real people typed about your product. Feed those phrases to whatever writing tool you like, or just use them directly.
Why does the advisor say nothing about my new headline?
Most likely because you shipped it at the same URL. The Analytics section is fed your top 15 pages as path, views, visitors, and conversions, so it can only compare copy that lives at different paths. It also receives a separate list of pages first seen in the last 7 days, which is where a variant on a new path shows up.
How soon can I re-check after a copy change?
Not the same afternoon. A section re-analyzes automatically at most once every 24 hours, and the manual refresh is capped at 12 hours per project. The overview response returns refresh_available_in with the seconds remaining. Give a copy change several days of traffic before you read anything into it anyway.
How much traffic do I need before the copy advice is worth anything?
The advisor will not score at all until your project has 24 hours of history, and returns has_sufficient_history: false before that. Past the gate, the limit is conversions rather than views. A page with 60 views and 1 signup cannot tell you whether the headline or the offer is the problem.