KEIT Part Two: the Client Deleted 80% of the Content, and What It Took to Come Back

This is the half of the project that does not go in a pitch deck. Between January and May 2025 this site went from 7,600 organic visits a month to a peak reading of about 22,000, recorded on a single day in May. Then the client deleted roughly 80% of the content, rewrote most of what survived, and by October 2025 the monthly figure was 4,700: minus 79% measured against that peak reading, and it sat there for six weeks.

No links were lost. No manual action. No algorithm update aimed at the vertical. Authority Score, referring domains and backlinks did not move by a single point for twelve months. The infrastructure was identical. Google had simply changed its mind about what the site was.

The recovery, from 4,700 to 14,500 a month by March 2026, up 208%, did not come from putting the content back. It came from measuring where the site’s semantic centre of gravity had drifted to, and moving it back. The single number I care about most is a different one: Italian traffic share went from 8% at the original peak to 16% now, on a site with less traffic in absolute terms than it had in April 2025.

Client: KEIT Day Hospital, aesthetic surgery, Albania, serving the Italian market · Crash window: May to October 2025 · Recovery window: December 2025 to March 2026 · Measured with: Semrush, Google Search Console, and embedding analysis of the full site in RankWit AI · Consultant: Jeansy Sese, Head of SEO

Executive summary

In May 2025 the site hit a peak reading of around 22,000 organic visits. Two weeks later the client made two editorial decisions independently and in good faith: delete about 80% of published content as obsolete or off-brand, and rewrite most of the rest.

The monthly average for May was 11,300, and it kept falling for six consecutive months to a floor of 4,700. Throughout, every off-page signal was flat: Authority Score 30, 358 referring domains, 3,100 backlinks, unchanged. That combination, collapsing on-page performance with a frozen off-page profile, rules out most of the usual explanations and leaves one: the site’s topical classification had moved.

Embedding analysis of every indexable URL, compared against the barycentre of the target clusters, found that the site no longer read as a surgical clinic. It read as a health-awareness magazine. Social campaigns, general prevention articles and civic content made up roughly 55% of the pages. Core surgery was 30%.

Three levers over four months, with no link building of any kind: selective de-indexing of out-of-cluster content, reinforcement of the surgical core, and reweighting of the internal link graph. The surgery and medical tourism clusters went from 40% to 75% of the site. The barycentre’s cosine similarity to the aesthetic surgery cluster went from about 0.64 to about 0.81. Traffic tripled.

What this case study does not claim: the site has not returned to its 22,000 peak, and may not. The recovery happened in a window that also contained Google algorithm updates, the expansion of AI Overviews and medical tourism seasonality, so I cannot attribute 100% of it to the intervention. The limits section says exactly what I can and cannot defend.

What happened after the April 2025 peak

Two decisions, taken on the marketing side with reasonable editorial rationales:

• Around 80% of published content deleted. Pages judged obsolete, repetitive or no longer aligned with the brand’s positioning were removed.
• Most surviving articles rewritten, often from scratch, with new angles and new communication objectives.

The problem is not that either decision was unreasonable. It is that a site is not classified by its owner’s editorial intentions. It is classified by its aggregate semantic footprint.

The interesting data is not in the traffic column. It is in the last three.

The diagnostic signal: off-page flat, on-page collapsing

Through the entire decline the domain’s off-page signals were static: Authority Score 30 for twelve consecutive months, 358 referring domains with nothing lost and nothing acquired, 3,100 backlinks flat. That pattern eliminates a list of causes:

• Not a link problem. Nothing was lost.
• Not a domain trust problem. Authority Score never moved.
• Not a severe algorithmic penalty. No de-indexing.
• Not a manual action. No message in Search Console.

One category remains compatible with the data: the site changed how it appears to Google, and that changed its topical classification. It did not lose authority. It lost relevance to its own target cluster.

Why a content cleanup can destroy rankings

There is a widespread story in digital marketing: cut the weak content, the strong content remains, Google rewards you. Like most simplifications it is partly true and operationally dangerous.

The accurate version: cutting content changes the semantic average of the site, which is where the engine places the domain in topic space. Cut 10% and the centre of gravity barely moves. Cut 80% and it can land in a different cluster.

Two concrete effects of mass deletion:

1. Aggregate reranking. A site is not evaluated page by page in isolation, it is evaluated as an ecosystem. When 70 to 80% of URLs start returning 410 or redirecting inside a concentrated window, the topical distribution of the domain is recalculated. Surviving pages are re-evaluated against the new composition, not the old one.
2. Loss of semantic triangulation. A rhinoplasty page does not stand alone. Much of its authority comes from the adjacent pages that support it, link to it and define its context: anaesthesia, recovery, costs, the clinic, the surgeon, testimonials. Remove 80% of the site and a page that was supported by six neighbours is supported by one. Its relative weight in the internal graph collapses.

Rewriting adds a third effect: it moves each page relative to its own historical embedding. The engine has a memory of how you treated a topic. Rewrite every page from scratch with a different angle, tone and vocabulary and it faces content that is formally at the same URLs but semantically different, which opens a period of instability where previous rankings get re-queried.

The three together explain the pattern exactly: off-page unchanged, on-page reconstituted, rankings sliding for six months.

The diagnosis: embedding analysis of a whole site

The operational question is: where did the site’s centre of gravity move to? Answering it means vectorising the site. Six steps.

• 1. Content extraction. Crawl every indexable URL, excluding archives, tags and pagination. For each one, extract the main text: title, H1, body. Discard navigation chrome, footer and widgets.
• 2. Embedding per URL. Convert each text into a high-dimensional vector. Here this ran through RankWit AI, which handles extraction, embedding, distance calculation and cluster comparison in one environment. Every URL becomes a point, and the domain becomes a cloud of points you can analyse.
• 3. Site barycentre. The vector average of all points: a single vector representing the site’s average position in semantic space.
• 4. Target barycentres. The same calculation for the reference clusters, here aesthetic surgery, medical tourism Albania, rhinoplasty, breast surgery, liposuction and blepharoplasty. Each cluster is characterised from 15 to 20 representative pages taken from the leading competitors in the sector.
• 5. Site-to-cluster distance. Cosine similarity between the site barycentre and each target barycentre. Higher similarity, better alignment.
• 6. Internal distribution. For every URL, cosine similarity to every cluster barycentre, then label each page with its nearest cluster. This produces a percentage distribution: how much of the site belongs to which cluster?

Step 6 is the one that found the problem.

The finding: it was not a clinic, it was a magazine

Read another way: less than half the site was aligned with the clusters the client wanted to compete in commercially. The rest was content with real editorial merit, breast cancer awareness campaigns, civic initiatives, general prevention articles, that was pulling the site’s centre of gravity away from aesthetic surgery.

To Google, KEIT was a hybrid: part clinic, part civic magazine, part lifestyle. An entity that is hard to classify is an entity shown less often on core queries, because the algorithm prefers to serve results where the query-to-document match is unambiguous.

And the 80% deletion had not fixed this. In places it made it worse, because the cleanup was not selective: it removed technical surgical articles alongside social content, leaving a residual distribution that was still unbalanced.

The correction: recentre the site on the surgery cluster

The logic was not to restore the deleted content. It was to deliberately change the site’s composition so that its barycentre moved inside the aesthetic surgery cluster. Three levers.

Lever 1 — Selective pruning of out-of-cluster content

Not every social page was removed. Each was judged on two criteria: the traffic it actually brought, since a non-commercial page that still attracts relevant visitors has a role, and its distance from the target barycentre, with pages under 0.55 cosine similarity to the surgery cluster becoming candidates.

About 45 URLs ended up de-indexed: dated social campaigns, press releases about finished civic initiatives, general prevention articles unconnected to surgery. Deliberately noindex,follow with removal from menus and internal links, not 410, which triggers aggressive re-crawling. The goal was to subtract them from the average without losing their historical link equity.

Lever 2 — Reinforce the surgical core

In parallel, the surgery cluster was expanded and densified:

• Procedure pages, rhinoplasty, breast surgery, abdominoplasty, liposuction, blepharoplasty, rewritten with dense technical vocabulary: procedural terms, anaesthesia protocols, recovery times, patient selection criteria.
• Semantic triangulation: for each surgical entity, three coherent but distinct attributes, for example technique, recovery and outcome for rhinoplasty, with the ring closed by justified internal links.
• A consolidated domain vocabulary: every surgical page now uses the same 40 to 60 technical terms consistently, with targeted co-occurrence density.

Lever 3 — Redistribute weight in the menu and architecture

The main menu was rebuilt around the surgical categories. Social and civic news sections moved to the footer or to submenus. The point was not to hide them, it was to reduce their weight in the internal link graph, and therefore their contribution to the centre of gravity the engine infers from navigation signals.

The distribution after four months

Surgery plus medical tourism went from 40% to 75% of the site. The site barycentre’s cosine similarity to the aesthetic surgery cluster went from about 0.64 to about 0.81.

The measured recovery, October 2025 to March 2026

Three things in that curve are worth more than the headline number.

• The floor is a plateau, not a point. Six weeks at the minimum without going lower. That says the floor was residual authority, the off-page profile that never moved, rather than an indefinite downward trend.
• The inflection point is sharp and datable, late November to early December 2025, and it coincides with the rollout of the first post-analysis corrections: out-of-cluster pages de-indexed, first menu changes.
• The growth happened through multiple Google algorithm updates, visible as markers on the Semrush graph. This was not a recovery that waited for the storm to pass. It went through it.

The number that matters more than the traffic

Less absolute traffic than at the peak, and twice the share of the Italian market for its cluster. That is the point of the whole case: the traffic was not rebuilt by putting content back. What was rebuilt is the precision with which Google maps this clinic onto Italian surgery queries. The site now captures a much larger slice of the semantic market it sits in, from a smaller body of content. Authority density per page went up.

And again, for emphasis: Authority Score 30, referring domains 358, backlinks 3.1K, unchanged throughout. The recovery is 100% content-driven. No link building campaign, no press release with links, no guest posts. Composition only.

A side effect worth naming: AI Overviews

In the Semrush keyword position tracking, an AI Overviews band appears for the first time in December 2025 and grows from there. Before that date this site had never been cited in a Google AI Overview for its own cluster. The timing overlaps the recentring and the parallel GEO work, and it supports an operational rule worth stating plainly: a site that is semantically coherent with its cluster gets rewarded by classic results and by AI search at the same time. They are two manifestations of the same underlying signal.

Why the recovery works: positive aggregate reranking

The mechanism that produced the crash produced the recovery, with the sign reversed. Once the site’s composition changed substantially, 45 or more out-of-cluster URLs de-indexed, the surgical core densified, the menu restructured, the engine re-crawled and re-evaluated the aggregate. This time the evaluation returned: a site about aesthetic surgery, with a barycentre coherent with the cluster, stable off-page signals and a triangulated internal structure.

The result was progressive positive reranking: not one shock but a series of micro-improvements spread over three to four months, visible as a slow ascending curve. Because the domain had never lost off-page authority, relevance was the only thing that had to be restored, and traffic came back with a query distribution far more concentrated on the target clusters. That concentration is the 16% traffic share.

A crash, a floor and a climb with no off-page intervention anywhere in the sequence is the characteristic signature of a topical perception problem, not an authority problem. Anyone in a similar situation can recognise the pattern from the shape alone.

When to run embedding analysis, and how

You do not need a crash to justify it. This case shows it is a preventive diagnostic, as useful as a technical audit or keyword research.

Triggers:

• Before any content pruning, at any scale. Knowing in advance whether a cut will move the barycentre is worth months of avoided recovery.
• After a significant editorial change: rebrand, pivot, change of editorial line.
• When traffic falls with no identified off-page cause. If authority, links and penalties are all clean, the problem is aggregate on-page, and embeddings find it.
• At onboarding, to see where a site sits in semantic space before proposing any strategy at all.

Toolkit: RankWit AI handles embeddings, barycentres, cluster analysis and site-versus-competitor comparison in a single flow. The hand-built alternative is a modern embedding model, such as text-embedding-3-large or an open-source equivalent, plus Screaming Frog for the crawl and numpy or scikit-learn for the maths, plus a reference set of 15 to 20 pages per target cluster taken from the sector’s main competitors.

Cost: for a site of 200 to 400 URLs the full analysis is four to six hours of work. The information return is disproportionate to that cost.

Limits and honesty about the data

• Single sample. One client, one crash, one recovery. This is an operational proof of concept on a real case, not generalisable statistical evidence.
• Causality is not isolated. The recovery window, December 2025 to March 2026, also contained Google algorithm updates, the expansion of AI Overviews and medical tourism seasonality. I cannot demonstrate that 100% of the recovery is attributable to the embedding work. What I can demonstrate is that a static off-page signal would not have produced a 208% recovery on its own, without the measurable changes in composition.
• The peak and the floor are not the same kind of number. The 22,000 is the highest single reading of May 2025; the 4,700 and the 14,500 are monthly figures. The minus 79% is therefore a drop measured against a best day, not against a sustained monthly level, and the sustained May level was 11,300. Against that, the floor is a fall of about 58% and the current 14,500 is above it. Both framings are in this page on purpose, because only one of them was ever going to be quoted.
• The time effect is incomplete. At 14,500 the site is still below its 22,000 peak reading. Another two or three months might take it past that peak, or the plateau might settle below it. Twelve-month data will say.
• Model dependency. RankWit AI uses an internal model tuned for SEO analysis. With a different model the absolute cosine distances would differ. The relative distributions across clusters, which is what actually drives the decisions, are robust across models.
• No blame assigned to the client. These were legitimate editorial decisions about their own content. This case is not about errors in any normative sense. It is about an information gap on how semantic ranking works, one that most sites and most consultants do not cover.

What to apply to your own site

• Never cut content in bulk without measuring the barycentre first. A blind cleanup can cost six months or more of recovery.
• Do not rewrite everything in parallel. A concentrated rewrite is a second aggregate shock on top of the deletion. Work in staging, test, roll out gradually.
• Traffic is not the sensitive metric. Traffic share on the cluster is. It answers the real question: is the site capturing its slice of the semantic market it competes in?
• Off-page stable plus on-page falling means a composition problem. Do not go looking for links. Go looking for your centre of gravity.
• Prefer noindex,follow to deletion when you want to subtract a page from the average while keeping its historical equity.
• Judge the internal link graph, not just the page. A page’s weight comes substantially from the neighbours that support it.

Timeline

Conclusion

When the crash started we did not know how badly the site was misaligned. We knew the client had deleted a lot and rewritten a lot, and the working assumption was normal post-cleanup instability. At three months of instability the pattern was worrying. At six months, with off-page provably static, it was clear the problem was not temporary: it was structural, in the composition of the site.

Embedding analysis gave a picture of the problem that no traditional SEO tool could produce. It is not an exotic method. It is close to how modern engines actually see a site, and the fact that most SEOs do not use it yet is a competitive advantage rather than a technical limitation.

The sentence I would keep from this project: modern SEO is management of the site’s average vector, not management of individual pages. Every editorial action, publishing, deleting, rewriting, moving, linking, moves the point the site occupies in semantic space. Push it towards your cluster and traffic grows. Push it sideways and traffic falls. Off-page modulates the intensity of the signal. It does not decide its direction.

Traffic falling with your links intact? That is a composition problem, and it is measurable before you guess at it. Advanced SEO consulting.

Related: the Security-SEO case study, where a cloaker malware tried to erase 22,000 visits a month at a medical tourism clinic · how I isolate an intervention when several things move at once.

Italian version: Caso studio SEO KEIT Parte 2.