A landing page has one job: keep the promise the ad just made. “Dynamic landing pages” describes three different ways to do that, and I would only let one run unattended after setup.
Take an emergency plumber account with 40 ad groups and one website. Someone searches for burst-pipe help, clicks an ad promising 24/7 service, and lands on a generic homepage. The business may offer exactly what they need. The page makes them work to find it.
There are three ways to close that gap:
- Text replacement swaps selected words on an existing page.
- Templated pages reuse one layout for different keywords, services or cities.
- AI-written pages draft distinct pages for ad-group themes using search terms, competitor pages and customer language.
The first can run automatically within approved boundaries. The other two create pages that need review before publication. That distinction matters more than whether a vendor puts “AI” on the button.
Text replacement changes words, not the page
The simplest build never creates a page at all. In a landing-page tool, I can define a parameter such as keyword, write a default headline like Fast Emergency Plumber in Denver, and append ?keyword={keyword} to the Google Ads final URL.
After the click, the tool reads that parameter and replaces the marked text. A burst-pipe keyword produces a burst-pipe headline; a water-heater keyword produces a water-heater headline. The layout, proof and underlying offer stay the same. Unbounce documents this parameter-based setup, and this practitioner walkthrough shows the same URL pattern.
That is useful when the visitor’s wording changes but the service does not. It is less useful when the visitor needs different evidence, a different offer or a different explanation.
Write a default that works on its own. Visitors arriving without the parameter, including through copied links or misconfigured ads, see the fallback. I check both the parameterized URL and the plain one.
I also keep replacement narrow: the headline, first paragraph and, where it reads naturally, the CTA. Repeating the keyword everywhere is not message match. It is a mail merge with ambitions.
The automation boundary is the approved wording. A text swap does not write a whole new offer, but an unchecked keyword can still create a misleading or nonsensical headline. I approve the possible inputs and their surrounding copy before letting the swaps run unattended.
Templates scale a layout, not the evidence
Now take the same plumber account across 12 suburbs. A template can generate 12 URLs from one layout, each with its own headline, relevant reviews and call-tracking number.
The attraction is obvious: I do not have to rebuild the page for every location. But the template only creates the structure. It does not make a review local, prove a service area or turn a copied paragraph into useful information.
A new URL is not the same thing as a new reason to trust the business. Before publishing, I check whether each page accurately represents the service and gives that visitor something relevant.
There is also an important distinction between a paid-traffic page and an organic-search strategy. Google’s doorway-abuse policy covers patterns such as regional pages that funnel visitors to one destination and substantially similar pages positioned between search results and a useful destination. That does not mean every templated location page is automatically a violation. It means swapping city names is not, by itself, an SEO strategy.
For near-identical PPC variants, my default is to keep them out of the organic index with noindex. Calling a page “PPC-only” is not an indexing control. Pages intended to rank need their own useful content, not just their own address.
Templates save production work. They do not remove the approval step.
AI-written pages change the argument, not just the headline
The third build goes further. Instead of swapping words or cloning a layout, an assistant drafts a distinct page for an ad-group theme using live inputs:
- Converting search terms.
- Current ad copy.
- Competitor ads and landing pages.
- Customer language from places such as Reddit.
For the plumber account, a burst-pipe page might lead with shutoff help and the business’s actual arrival window. A water-heater page might need to explain replacement options and the pricing information the business can genuinely provide.
Those are different arguments, not different keyword slots. That is the point of message match: repeating the ad’s promise rather than merely its exact wording.
I build around one ad-group theme, not one page per keyword. Closely related searches can need the same explanation. Generating a separate page for each one adds review work without necessarily adding value.
This is also where automatic publishing becomes the wrong default. An assistant can borrow a competitor’s persuasive structure and accidentally borrow its guarantee too. The sentence may be excellent. The promise may not be yours.
I also check where the ads actually send traffic. AI Max Final URL expansion can select a different page on the domain, which matters when the plan is to test one specific destination. The groas explanation of dynamic landing pages for Google Ads covers the three builds in campaign terms.
The practical rule stays simple: automate the draft, approve the offer, then publish.
How I would build the plumber page through groas
The workflow starts with connections, not a heroic prompt. I connect the ad accounts and website to groas once, then add groas to ChatGPT or Claude over MCP.
The assistant can work with live search terms and ad copy alongside the data groas collects: competitor ads on Google and Meta, their landing pages and funnels, customer discussions on Reddit, AI-search visibility, and SEO data from Semrush and Ahrefs.
My AI is the brain. groas is the hands and the data.
Say the plumber account spends $20k a month and the burst-pipe ad group still points to a generic homepage. I would give the assistant this instruction:
Draft a landing page for the emergency plumber ad group using our top-converting search terms and what competitor pages promise. Don’t publish yet.
The assistant can compare the queries with rival pages and draft a more specific response. But competitor research tells me what the market promises, not what this business can deliver. I still need to verify the arrival window, pricing and suburbs served against the actual offer.
My sequence is deliberately unglamorous:
- Draft: gather the inputs and write the page.
- Approve: check the claims, proof, wording and destination.
- Publish: give the explicit instruction to publish and update the ad group’s final URL.
- Log: record what changed and why.
Only after approval would I say:
Publish it, point the ad group’s final URL to it, and log why.
The value is not that AI makes approval unnecessary. It is that I can review a prepared page instead of assembling every input and writing every line myself.
Separate permission to draft from permission to publish
For this workflow, I allow automatic reading, comparison and drafting. I reserve publishing and traffic changes for explicit approval.
The assistant can write three headline options at 3am. It does not need permission for each adjective. It does need permission to put an unreviewed promise in front of paying visitors.
groas’s autopilot stays off until switched on. When enabled, it runs under the rules set for it; changes are logged with their reasons and can be undone in one step. For new landing pages, my rule is approval before publication. I would not confuse enabling automation with approving every future page it might draft.
The same draft-first pattern appears in WordPress publishing tools that support restricted publishing and reversible changes. The important distinction is between preparing work and committing it.
A log helps explain a mistake. Undo helps reverse a change. Neither makes an invented price acceptable while the page is live.
If a tool offers automatic page publishing with no draft step, I skip it. The boring approval screen is doing useful work.
Decide what a win means before sending traffic
A more relevant page is a hypothesis, not a result. For the plumber account, I care about booked jobs per click and cost per booked job, not whether the new page looks busier or has a lower bounce rate.
Before publishing, I record:
- The baseline: CPA or cost per booked job for that ad group over the last 30 days.
- The measurement: the same conversion tracking for both versions.
- The comparison: how traffic will be split between the old and new pages.
- The stopping rule: the sample requirement and planned test duration.
Then comes the awkward part: checking whether the account has enough traffic to answer the question.
At 95% confidence and 80% power, a 2% baseline conversion rate needs about 39,000 visitors per variant to detect a 20% relative lift; a 10% baseline needs about 6,500. Those are not weekend sample sizes for a burst-pipe ad group receiving 400 clicks a month. They are not two-to-four-week sample sizes either.
The required count has to fit the available traffic. If it does not, I cannot promise a quick, statistically clear verdict on a modest improvement. I need to acknowledge that limit before launching, not explain it away after a pleasing handful of bookings.
Running whole weeks helps avoid comparing mismatched slices of the week, but elapsed time does not replace sample size. Nor should I stop the moment the dashboard looks favorable: repeated peeking followed by an early stop can manufacture false winners.
I write the test plan into the publishing instruction and change log. No adequate sample, no confident verdict.
Three limits automation does not remove
The workflow saves repetitive work. It does not repeal the reasons landing pages fail.
More pages do not automatically mean more value
Fifty suburb pages with identical proof blocks and swapped city names are still thin pages. If they are made primarily to manipulate rankings without adding value, they fall into the territory Google describes as scaled content abuse, regardless of how they were created.
Judge the value per page, not the speed of production. A human copying the same weak page by hand does not make the strategy better.

Fluent claims still need checking
An assistant drafting from rival pages and customer discussions can write an arrival time, warranty and starting price in the same confident voice. Confidence is not verification.
A draft that promises 45 minutes when the business offers same-day service has changed the offer. That is not a small copy edit. Check every operational promise before publication, especially the ones that make the page persuasive.
Message match does not guarantee lift
A page can repeat the ad accurately and still fail to improve bookings. It may not resolve the visitor’s objection, or the observed difference may simply be noise.
The measurement plan exists to separate a plausible mechanism from a demonstrated result. If nobody can show the baseline and the count, I do not accept the lift claim.
The checklist before a page goes live
I use this list before giving the publishing instruction:
- Fallback checked: the headline, first paragraph and CTA work without replacement text; both the parameterized and plain URLs have been tested.
- Claims approved: arrival time, price, warranty and service area match the business, not a competitor’s offer.
- Page reviewed: templates and AI-written drafts have been checked before publication.
- Destination controlled: the intended ad group points to the intended page, with URL-expansion settings considered.
- Indexing decided: near-identical PPC variants are noindexed; pages intended for organic search add real value.
- Change recorded: publication and final URL changes have a reason in the log and a clear undo path.
- Test specified: baseline, conversion tracking, traffic split, sample requirement and planned duration are written down.

Apply the model to a different business
Now take a SaaS ad group for project management for agencies that sends visitors to a generic feature page.
I start with the same question as the plumber example: what promise did the ad make, and what does the current page fail to explain?
If the page already contains the right proof and demo offer, text replacement may be enough to make the opening relevant. If the structure works but the agency audience needs different proof, a cloned layout is only a starting point. It does not supply that evidence.
If the argument itself needs changing, I want the third build. I would ask Claude or ChatGPT with groas to draft from converting search terms, stored rival funnels and Reddit complaints about missed deadlines. Then I would hold the page in draft and check whether it answers those complaints without inventing capabilities or changing the real demo offer.
Only after that review would I publish and point the ad group at it. The 25 Google Ads prompts for Claude and ChatGPT provide wording to start from.
The choice follows from how much must change: words, structure or argument. The approval requirement follows from the same distinction.
That is where groas pulls ahead of a standalone builder. The assistant has the live account data and competitor pages behind the draft, while groas handles publication on instruction, logs the change and provides one-step undo. The work does not stop at generating a page-shaped object.
For teams running it themselves, groas Unlimited is $19 a month, flat, covering every account and feature. For businesses that would rather hand over the work, Done For You starts at $5,000 a month with groas’s senior team. Either way, the useful sequence stays the same: draft, approve, publish, measure.

