SEO teams spend plenty of time creating new pages while existing pages quietly lose rankings, clicks, and revenue.
Updating those pages can recover performance. But rewriting too much wipes out the SEO equity they have already built.
This is the 14-step process ground transportation marketplace hoppa uses to diagnose content decay and make targeted updates — and how they turned it into a system with Claude Code.
Content decays. Not dramatically — no penalty, no algorithm update to blame. Just a slow drift.
A ranking creeps down. Impressions hold steady while clicks quietly fall. An AI Overview eats the top of the results. Google notices before your analytics dashboard does.
The example comes from a marketplace with pages covering airports, resorts and routes across multiple countries and languages. Before touching its Antalya Airport transfers page, the 56-day diagnostic showed:
Buried, not invisible. Nothing was broken. It was just stale.
New pages start at zero. Updating a page that already ranks is a different job entirely — you are working with a live asset:
Teams "refresh" a decaying page by rewriting it top to bottom and lose every ranking it had. That is a self-inflicted demotion.
Every update starts with a 56-day window in Search Console. Not 90 days, not a year.
Two reasons. It is wide enough to be reliable and narrow enough to stay within one season — on a seasonal site, a wider window averages June against January and produces the wrong conclusion.
Second, the same window is used when the update goes live. SEOTesting offers 2, 4, 6 or 8-week test periods; eight weeks equals 56 days.
This tagging discipline is really the whole game.
Most self-inflicted damage comes from mistaking a Keep section for a Fix.
Here it stops being generic. Antalya's query data turned up:
Personas come from a two-source process rather than a single dataset.
The foundation is a sitewide taxonomy: a Search Console export covering the last 16 months, with every query clustered into a master persona set that holds across the whole portfolio.
For any specific update, that sitewide set is paired with SEOTesting's Query Fan-Out tool, fed a seed term for the destination. It generates synthetic queries people would plausibly ask within an LLM interface rather than type into a search box, clustered the same way.
The two datasets — 16 months of real GSC behaviour and synthetic fan-out queries for that destination — combine into the final set.
For Antalya, that produced four personas:
What changes at an airport destination every 18 to 24 months gets checked against current sources rather than assumed from the original brief:
Then the angle. The original angle on a commercial page tends to start generic — something like "we make airport transfers easy."
That is not an argument. It is a brochure, and it said nothing to the four different personas the data just surfaced.
Antalya's new angle:
Different travellers need different transfers, and the page should surface the right answer based on who is actually looking, instead of presenting every option to everyone and hoping they self-sort.
Not a brief for the whole page: a delta brief covering only what changes.
Each section gets one of the four labels explicitly, with a reason attached. It runs roughly 1,500 words for a page this size and reads more like an engineering change request than a writer's brief — which is the right tone for the work.
Only the sections tagged "fix" and "add" get written.
That meant refreshed copy and keyword coverage across the fix list, plus one add: a persona chooser, "Expert Airport Transfer Finder."
A visitor picks the option closest to their situation, and the module surfaces the vehicle recommendation, price range, and detail that matters to that persona.
Notably, the component does not add new information. All of it already existed somewhere on the page, spread across vehicle explainers, the group-size guide and the FAQ. It gives each visitor a direct path to the part that is already theirs, instead of asking them to scan the whole page.
Including the "keep" sections.
Staying does not mean it is still accurate, so every numeric claim — distances, prices, transit times, terminal assignments — gets reverified against current sources.
Most "AI-refreshed" pages skip this part, updating the surface text while leaving the underlying facts untouched.
Every image gets checked for three things: still accurate, still on-brand, still meeting current performance spec (WebP/AVIF, properly sized, lazy-loaded). Anything failing gets replaced.
The rule for everything else is preserved unless there is a specific reason not to:
When the brief calls for something visual rather than prose, vibe-code it:
The old workflow for a custom component was Figma mockup, design review, dev sprint, QA — two weeks best case. This is closer to an hour for moderate complexity.
That is the only reason a persona chooser is feasible per page across a portfolio this size rather than a rare special-case build.
Every update runs through SEOTesting, connected to both Search Console and GA4.
The same locked window and control-versus-test structure are used on both sides simultaneously, so a content update is judged by revenue and sales, not just rankings.
Measured against the 56-day baseline locked before a single word changed:
| Metric | Control | Test | Change |
|---|---|---|---|
| Clicks/day | 39.55 | 49.00 | +23.88% |
| Impressions/day | 2,652 | 2,744 | +3.44% |
| Avg. position | 14.89 | 10.87 | improved |
| CTR | 1.49% | 1.79% | +0.30pp |
| Queries/day | 366 | 389 | +6.28% |
Every single metric moved the right way.
Most updates that move one number quietly cost you on another. This one didn't.
And it was not down to any single piece of the delta. It is what the diagnostic, the angle, the rewritten sections, and the new component produced together.
That is the whole point of locking a baseline before touching anything: it is the only way to know an update did something, rather than getting lucky with a publish date.
That whole process, done properly by hand, is about two focused days per page.
hoppa runs thousands of pages across nine languages. Two days a page is fine math until you multiply it across a portfolio that size. Even at one update per page per year, the arithmetic does not survive contact with reality.
The real cost is the opportunity cost: every month a page sits at position 14.89 instead of 10.87 is a month of impressions that never got the chance to convert.
So the 14 steps were mapped into Claude skills, with Claude Code running the process instead of a person running it with AI help on the side.
The judgment didn't move to the machine. The enforcement of that judgment did.
Every skill that touches content runs inside a project preloaded with the brand book, terms and conditions, pricing policy, and a library of brand-specific reference material — so brand constraints are always in context instead of being re-explained each run.
| Step | Skill | Function |
|---|---|---|
| 1 | hoppa-intelligence | Pulls the 56-day GSC window and locks the baseline automatically |
| 2 | Audit Skill | Runs keep/fix/remove/add tagging against actual query performance |
| 3–4 | Competitor-Gap Skill | Defended queries, gap-close queries and new intents from Ahrefs, plus a SERP read |
| 5–7 | Editorial Intelligence | Persona revalidation, query fan-out gap detection, local-knowledge refresh |
| 8 | Delta-Brief Skill | Generates the brief with keep/fix/remove/add explicit |
| 9 | hoppa-editorial | Writes only "fix" and "add" sections, calibrated against gold-set tone benchmarks |
| 10 | hoppa-scientific-refiner | Fact-checks new and kept content |
| 11 | Image-Auditor | Runs the accuracy, on-brand and spec checks automatically |
| 12 | seo-preservation | Locks the slug, protects CTR-earning meta titles, extends schema |
| 13 | Component Generator | Turns the brief's spec into the working component |
The hard gates sit at steps 5–7. The update cannot proceed without:
Skip those and you get generic AI output, which is exactly why so much AI-assisted content reads the same regardless of who published it.
A failed fact-check on a "keep" section automatically bumps it to "fix." No human has to ask "is this still true?" first.
If the new content's voice drifts from the kept content's, step 9 refires until they match.
Measurement isn't a separate skill so much as the discipline that wraps around all of it. Tests are still set up manually.
Getting the content right was only half the job. The other half was getting it live — which until recently meant a person copying finished sections into the CMS, checking formatting, and hitting publish.
Now Claude connects directly to the CMS (Strapi) through an MCP server, and deployment runs as its own step:
If any check fails, the deploy halts and flags it rather than silently shipping something broken. When everything passes, it produces a single ready-to-publish report for a human to approve.
Content production and CMS deployment are now a single continuous pipeline rather than two jobs with a person bridging them.
Once the pipeline worked for one page, the question became how many it could run at once without compromising quality.
Two content updates were tested in parallel. Technically it works — nothing errors out, nothing breaks. But quality drops on both.
The two runs share the same underlying agents, and pushing two batches through simultaneously visibly softens the output on each:
So they do not run it that way. One update runs at a time, start to finish, working through a batch sequentially.
It is slower on paper. It is also the difference between output you trust on the first read and output that needs a second pass to catch what got rushed. At this stage of the tooling, that trade is not close.
Not yet updated:
Fully updated:
Remove full rewrites from the default menu. Refreshing a ranking page is editing, not re-authoring. Skipping the tagging step erases that distinction.
Lock the baseline first. Without stored pre-update data you cannot prove the effect — and that proof funds the next round.
Fact-check the "keep" sections too. This is where most AI refreshes cut corners, and why stale numbers survive.
Design hard gates into the automation. Blocking progress without personas, local knowledge and an angle is the mechanism that stops AI content from all reading the same.
Resist parallel execution. Doubling throughput at the cost of quality on both runs is rarely the better trade.
Do not mistake a lift for causation — see Seven Reasons SEO Tests Fail.
Manage the decay backlog like technical debt — the framework in Technical Debt in SEO transfers directly.
Promote recurring corrections into rules. On updating system instructions when the same edit repeats, see Feedback Loops for AI Content Workflows.
Shipping new pages is visible work. Repairing old ones is not. The revenue is usually in the second one.
It is wide enough to be reliable while staying within a single season, and it matches the 8-week test period offered by SEO testing tools so baseline and test use identical windows.
Keep (still ranks and accurate, leave alone), Fix (right idea, stale execution, rewrite in place), Remove (wrong, redundant or harmful), and Add (the data indicates a gap).
Against the locked 56-day baseline, the Antalya page saw clicks/day +23.88%, impressions/day +3.44%, average position improving from 14.89 to 10.87, CTR +0.30pp and queries/day +6.28% — every metric moving the right way.
The 14 steps were mapped to Claude skills covering GSC baseline locking, section tagging, competitor gap analysis, persona revalidation, delta brief generation, writing, fact-checking, image auditing, SEO preservation and component generation. Judgment stayed with humans; enforcement moved to the system.
Technically yes, but the runs share the same underlying agents and quality drops on both — diagnostics get shallower, briefs looser and writing needs more editing. They run one update at a time, sequentially.
Yes. Staying does not mean still accurate, so every numeric claim — distances, prices, transit times, terminal assignments — gets reverified. This is the step most AI refreshes skip.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
An E-E-A-T checker was never automatable because there is no API for it. Here is how an AI coding agent read Google's quality guidance, crawled a page sample, and produced a scored audit — plus where it stops.
A GEO audit is not checking whether your brand appears in an AI answer. Here is a nine-point GEO checklist split into four on-site items, four off-site, and one measurement layer.
LinkedIn, YouTube, Reddit, and TikTok are building systems to identify low-value AI content. Here are the limits of watermarking and what content still gets distributed.