How to scale SEO content updates with Claude Code

SEO groups spend loads of time creating new pages whereas present pages quietly lose rankings, clicks, and income. Updating these pages can get well efficiency, however rewriting an excessive amount of can wipe out the website positioning fairness they’ve already constructed.

Right here’s the 14-step course of we use to diagnose content decay, make focused updates, and measure the outcomes — plus how we turned it right into a scalable system with Claude Code.

Why present pages deserve extra consideration

One of many manufacturers I handle is a floor transportation market with pages overlaying airports, resorts, and well-liked routes throughout a number of international locations and languages. The methodology on this piece is the one we use there, which is why it’s the instance all through.

Each a type of pages decays over time. Not dramatically: no penalty, no algorithm replace in charge. Only a gradual drift. A rating creeps down, impressions maintain regular whereas clicks quietly fall, and an AI Overview eats the highest of the search outcomes. Google notices earlier than your analytics dashboard does.

Take our Antalya Airport transfers web page. Earlier than we touched it, the 56-day diagnostic confirmed:

• 148,537 impressions.
• A mean place of 14.89.
• A 1.49% CTR.
• 2,215 clicks to indicate for all that visibility.

Buried, not invisible. Nothing was damaged. It was simply stale.

New pages begin at zero, so you are able to do no matter you need with a clean slate. That’s the enjoyable, simple a part of website positioning to speak about. 

Updating a web page that already ranks is a unique job totally, not to mention a business, revenue-driving web page. You’re working with a stay asset:

• Inside hyperlinks already pointing at it.
• Schema already in place.
• A historic baseline you may break when you’re careless.

I’ve watched groups “refresh” a decaying web page by rewriting it prime to backside and dropping each rating it had. That’s not a refresh however a self-inflicted demotion.

Dig deeper: 4 types of content decay and how to fix each one

The handbook course of

Step 1: Learn the 56-day GSC window

Each replace begins with a 56-day window in Search Console, not 90 days or a full yr. It’s huge sufficient to be dependable and slim sufficient to remain inside one season. 

On this occasion, the positioning is seasonal sufficient {that a} wider window simply means averaging June towards January and drawing the mistaken conclusion from the mix. That locked window can be the baseline towards which the eventual check will get measured.

One more reason we use 56 days is that the identical time window is used when the content material replace is pushed stay, and we arrange the website positioning check to trace its influence. SEOTesting, the software we use to arrange website positioning checks, affords 4 check intervals: 2, 4, 6, or 8 weeks. Eight weeks equals 56 days, which is why we use that very same time window because the baseline.

Inside the information, three issues matter most:

• Prime queries: These need to be preserved, no matter else modifications.
• Placing-distance queries (Positions 5-20, weak CTR): These are a budget wins.
• Zero-click queries (excessive impressions, virtually no clicks): These present the web page being served for an intent it isn’t answering.

Step 2: Tag each part

This tagging self-discipline is actually the entire sport:

• Maintain: Nonetheless ranks, nonetheless correct. Don’t contact it.
• Repair: Proper thought, stale execution. Rewrite in place.
• Take away: Fallacious, redundant, or actively hurting.
• Add: The info says one thing’s lacking.

Steps 3-5: Learn opponents, refresh key phrases, rebuild personas

That is the place it stops being generic. For Antalya, the question information turned up:

• A striking-distance alternative: “antalya airport switch” sitting at place 7, with actual quantity behind it.
• A non-public-hire hole: “personal switch antalya” pulling 1,091 impressions and 14 clicks, a 1.28% CTR, as a result of the web page had nothing for vacationers wanting a personal switch relatively than a shared shuttle.
• A comparability hole: “greatest antalya airport transfers” sitting at place 10.5, with no comparability content material on the web page in any respect.

How the personas get constructed

The personas come from a two-source course of relatively than a single dataset. The inspiration is a sitewide taxonomy: a Google Search Console export overlaying the final 16 months, with each question throughout the positioning clustered right into a grasp set of personas that holds throughout the entire portfolio, not only one web page.

For any particular replace, that sitewide set is paired with SEOTesting’s Question Fan-Out software and fed a seed time period for the vacation spot. It generates artificial queries that individuals would plausibly ask inside an LLM interface relatively than sort right into a search field, and people are clustered into personas in the identical method.

The 2 datasets — one constructed from 16 months of actual GSC habits, one from artificial fan-out queries for that particular vacation spot — then get mixed into the ultimate, destination-specific set. For Antalya, that mixture landed on 4 personas:

• The usual shuttle shopper.
• The private-hire shopper.
• The group traveler.
• The day-tripper.

Steps 6-7: Refresh native information, resolve the angle

What modifications at an airport vacation spot each 18 to 24 months:

• Terminal assignments.
• Taxi rank places.
• Rip-off patterns.
• Tipping conventions.
• Peak congestion.
• Evaluation themes on Trustpilot.

All of it will get checked towards present sources, not assumed from the unique temporary.

Then the angle. The unique angle on a business web page like this tends to start out generic: one thing like “we make airport transfers simple.” That’s not an argument. It’s a brochure, and it stated nothing to the 4 completely different personas whose information simply surfaced.

Antalya’s special approach: Totally different vacationers want completely different transfers, and the web page ought to floor the best reply based mostly on who’s truly trying, as a substitute of presenting each choice to everybody and hoping they self-sort.

Step 8: Write the delta temporary

Not a short for the entire web page: a delta temporary overlaying solely what modifications. Every part will get one of many 4 labels from Step 2, explicitly, with a purpose hooked up. It runs roughly 1,500 phrases for a web page this dimension and reads extra like an engineering change request than a author’s temporary, which is the best tone for the work.

Step 9: Write the delta

Solely the sections tagged “repair” and “add” get written. That meant refreshed copy and key phrase protection throughout the “repair” listing, plus one new “add” — a persona chooser, “Professional Airport Switch Finder.” 

A customer picks the choice closest to their scenario, and the module surfaces the automobile suggestion, worth vary, and element that issues to that persona.

The part itself doesn’t add new info: all of it already existed someplace on the web page, unfold throughout the automobile explainers, the group-size information, and the FAQ.

It simply provides every customer a direct path to the half that’s already theirs, as a substitute of asking them to scan the entire web page to search out it. It shipped as one a part of the delta, not by itself, alongside the repair work above. Price retaining in thoughts for the leads to a couple of steps from now.

Step 10: Truth-check all the things

Together with the “maintain” sections. Staying doesn’t imply it’s nonetheless correct, so each numeric declare (distances, costs, transit occasions, terminal assignments) will get reverified towards present sources, together with outdated content material.

Most “AI-refreshed” pages skip this half and replace the floor textual content whereas leaving the underlying details untouched.

Steps 11-12: Audit photos, protect the website positioning fairness

Each picture will get checked for 3 issues:

• Nonetheless correct.
• Nonetheless on-brand.
• Nonetheless assembly present efficiency spec (WebP/AVIF, correctly sized, lazy-loaded).

Something that fails will get changed.

In the meantime, the rule for all the things else is preserved except there’s a particular purpose to not:

• The URL slug by no means modifications.
• The meta title stays if it’s incomes CTR.
• Schema will get prolonged relatively than changed.
• Inside hyperlinks are checked in each instructions: into the web page and out of it.

Step 13: Construct the UI parts

When the temporary requires one thing visible relatively than prose, we vibe-code it:

• Describe the habits in pure language.
• Let Claude or Gemini generate the part.
• Iterate stay till it really works.
• Server-render it so LLMs can learn the content material with out executing JavaScript.

The outdated workflow for a customized part was Figma mockup, design evaluation, dev dash, QA: name it two weeks, best-case situation. That is nearer to an hour for one thing of reasonable complexity, which is the one purpose a persona chooser is possible per web page throughout a portfolio this dimension relatively than a uncommon, special-case construct.

Step 14: Measure the change

Each replace runs by means of SEOTesting, which we’ve linked to each Google Search Console and GA4. The GSC facet provides us clicks, impressions, place, and CTR. The GA4 facet provides us no matter truly issues commercially: the acquisition occasion, on this case, or every other GA4 occasion price monitoring.

The identical locked window and control-versus-test construction are used on each side concurrently, so a content material replace is judged by income and gross sales, not simply rankings.

Right here’s what the entire replace produced on Antalya on the search facet, measured towards the 56-day baseline locked earlier than a single phrase modified:

Each single metric moved the best method: clicks up, impressions up, CTR up, place improved, and question protection up. Most updates that transfer one quantity quietly value you on one other. This one didn’t, and it wasn’t all the way down to any single piece of the delta. 

It’s what the diagnostic, the angle, the rewritten sections, and the brand new part produced collectively. That’s actually the entire level of locking a baseline earlier than touching something: it’s the one method to know an replace did one thing, relatively than getting fortunate with a publish date.

Dig deeper: Refreshing content: How to update old content to drive new traffic

Turning it right into a system by means of Claude Code

That complete course of, achieved correctly by hand, is about two targeted days per web page. hoppa runs 1000’s of pages throughout 9 languages, and two days a web page is okay math till you multiply it throughout a portfolio that dimension. Even at a conservative one replace per web page per yr, the arithmetic doesn’t survive contact with actuality. 

The actual value isn’t the time, both. It’s the chance value: each month a web page like Antalya sits at place 14.89 as a substitute of 10.87 is a month of impressions that by no means acquired the prospect to transform.

So we mapped the 14 steps above into Claude abilities and had Claude Code run the method as a substitute of an individual working it with AI assistance on the facet. The judgment didn’t transfer to the machine. The enforcement of that judgment did.

The inspiration is a devoted Claude undertaking, preloaded with:

• Our model e book.
• Phrases and circumstances.
• Pricing coverage.
• A library of brand-specific reference materials.

Each talent that touches content material runs inside that undertaking, so model constraints are at all times in context as a substitute of being re-explained on each run.

Step 1 → hoppa-intelligence

Pulls the 56-day GSC window routinely and locks the baseline earlier than anything occurs: the identical 4 metrics, prime queries, striking-distance queries, and zero-click queries an individual would pull by hand.

Step 2 → the Audit Ability

Runs the maintain/repair/take away/add tagging routinely, classifying each part towards precise question efficiency.

Steps 3-4 → the Competitor-Hole Ability

Pulls defended queries, gap-close queries, and new intents from Ahrefs, plus a SERP learn on who’s outranking us and what floor they’ve taken.

Steps 5-7 → Editorial Intelligence

Persona revalidation and question fan-out hole detection, plus the local-knowledge refresh. That is the place the arduous gates sit: the replace can’t proceed with out:

• A validated persona set.
• Present native information.
• An outlined angle.

Skip these and also you get generic AI output, which is precisely why a lot AI-assisted content material reads the identical no matter who revealed it.

Step 8 → the Delta-Temporary Ability

Generates the temporary in the identical delta form as Step 8 above (maintain/repair/take away/add specific) and gained’t produce one with no outlined angle hooked up.

Step 9 → hoppa-editorial

Writes solely the “repair” and “add” sections, inside the identical undertaking, calibrated towards gold-set tone benchmarks. If the brand new content material’s voice drifts from the stored content material’s, it refires till they match.

Step 10 → hoppa-scientific-refiner

Truth-checks new and stored content material. A failed verify on a “maintain” part routinely bumps it to “repair.” No human has to ask, “Is that this nonetheless true?” first.

Step 11 → the Picture-Auditor

Runs the identical three checks routinely (correct, on-brand, spec-compliant) and flags no matter wants changing.

Step 12 → seo-preservation

Locks the URL slug, protects a meta title that’s incomes CTR, and extends schema relatively than changing it, the identical preserve-by-default rule as Step 12 above.

Step 13 → the Element Generator

Turns the temporary’s spec for something new (a persona chooser, a comparability desk) into the precise working part, working the identical describe-generate-iterate loop, simply contained in the talent relatively than an individual driving it by hand.

Measurement (Step 14) isn’t a separate talent a lot because the self-discipline that wraps round all of it. We nonetheless arrange the checks manually. The baseline is locked in Step 1 earlier than anything runs, and each batch is learn towards that very same baseline as soon as it’s stay.

Dig deeper: 6 content audit workflows to build in Claude

Closing the loop: Deployment

Getting the content material proper was solely ever half the job. The opposite half was getting it stay, and till just lately, that also meant an individual copying completed sections into our CMS, checking that the formatting held, and hitting publish.

We’ve since closed that hole. Claude connects on to our CMS, Strapi, by means of an MCP server we run for Strapi, and deployment itself now runs as its personal step:

• Staging deploy
• A full section-by-section diff towards the stay web page
• Inside and anchor hyperlink verification
• Schema validation

If any verify fails, the deploy halts and flags it relatively than silently transport one thing damaged. When all the things passes, it produces a single ready-to-publish report for a human to approve earlier than it goes stay.

Content material manufacturing and CMS deployment are actually a single steady pipeline, relatively than two separate jobs with an individual bridging them by hand.

What working this at scale truly taught us

As soon as the pipeline labored for one web page, the apparent subsequent query was what number of it may run without delay with out compromising high quality to the naked minimal.

We examined two content material updates working in parallel. It really works, technically: nothing errors out, and nothing breaks. However the high quality drops on each.

The 2 runs share the identical underlying brokers, and pushing two batches by means of the identical brokers on the identical time visibly softens the output on every: the diagnostics get shallower, the delta briefs get looser, and the writing wants extra enhancing on evaluation.

So we don’t run it that method, and we are able to’t with out giving one thing up. One replace runs at a time, begin to end, and the system works by means of a batch sequentially.

Nonetheless, many URLs are on the listing that week. It’s slower on paper. It’s additionally the distinction between output we belief on the primary learn and output that wants a second go to catch what acquired rushed. At this stage of the tooling, that commerce isn’t shut.

Dig deeper: 7 feedback loops for self-improving AI content workflows

What the distinction appears like on the web page

Right here’s the identical underlying self-discipline utilized to 2 different actual pages, one nonetheless ready within the queue and one by means of the total cycle:

Not but up to date:

• Generic copy
• No native content material
• No FAQ
• Not one of the newer parts

Absolutely up to date:

• Wealthy native element
• A banner tied to an actual present occasion
• Evaluation proof
• Structured sections that route completely different guests to what applies to them

Nothing about that hole required a rewrite from scratch. It required the delta.

The outcomes at portfolio scale

None of this issues if it solely works on one web page. We don’t name an replace a win as a result of it feels higher. Each one in every of these runs as a correct check, with a 56-day baseline locked earlier than a single change ships and measured towards the identical window after. That’s the distinction between a measured replace and a fortunate publish.

Seven in 10 up to date pages noticed natural clicks improve, one in 4 noticed them lower (going out of season was the case for a few of these pages), and a pair got here again flat. 

Throughout all 59 checks, normalized to comparable 56-day home windows, this netted 2,284 further natural clicks total, all touchdown on a business switch web page, not a weblog article. 

Natural purchases completed up 24.9% towards baseline, and natural income was up 20.1%, each learn straight off the GA4 facet. 

That’s the argument for treating content material updates as one of many highest-ROI line gadgets on an website positioning roadmap, relatively than as upkeep work that occurs solely when there’s nothing extra thrilling left to do.

What truly transfers to your crew

Not one of the specifics above is the purpose. Our prioritization weights, our persona schema, our tone benchmarks, and our gate thresholds gained’t imply something in your web site, and so they shouldn’t. Rebuild all of it in your personal area.

What transfers is the form of the system:

• A human units the listing and priorities and approves the output. Automation enforces judgment. It doesn’t change it.
• Onerous gates sit within the center, and so they’re allowed to dam progress. Skip persona validation or a local-knowledge refresh, and also you get generic AI output, which is precisely why so many AI content material pipelines sound an identical to one another.
• Preservation guidelines shield no matter’s already incomes. Breaking a rating isn’t an replace. It’s a demotion with higher branding.
• Each change is measured towards its personal locked baseline, not a imprecise sense of how a web page is “doing.”

Proper now, income picks which pages we have a look at first, and the diagnostic tells us what’s mistaken with them. That’s the right order for a crew defending high-value pages, but it surely’s reactive: decay has normally already value one thing by the point income flags it. The alerts exist sooner than that:

• CTR softening.
• Place drift.
• Protection gaps opening up earlier than they present up in a P&L.

The following model of this constantly screens the portfolio and surfaces pages about to bleed, not simply these already bleeding.

Content material updates on business pages will be a few of the highest-ROI website positioning work accessible, but most groups nonetheless deal with them as upkeep. Doing them correctly, one web page at a time, doesn’t scale.

Ours didn’t both, till we stopped asking the mannequin to write down and began asking it to implement judgment we’d already made, at no matter quantity the queue truly wanted.

In the event you’re looking at a queue of decaying pages and questioning the place to start out, begin with the diagnostic, not the rewrite. Every thing else follows from that.

None of this might have occurred with out the one who turned it from a handbook playbook right into a working Claude talent chain: Yvette Ramirez, our content material strategist, who constructed the implementation and stored iterating on the gates and tone benchmarks till the automated model stopped needing a second go.